Spotlight on our academic community

A Real-time, Data-driven Field Decision Framework for Large-scale Software Deployments

As large e-commerce systems need to maximize their revenue, while ensuring customer quality and minimizing IT costs, they are constantly facing major field decisions like « Would it be cost-effective for the company to deploy additional hardware resources for our premium users?” This project will build a real-time, data-driven field decision framework exploiting customer behaviour and quality of service models, release engineering and guided optimization search. It will benefit both Canadian software industry and society, by improving the quality of service experienced by Canadians.

Machine learning for the analysis of the Large Hadron Collider Data at CERN

The Large Hadron Collider (LHC) is one of the most ambitious experiment ever conducted. It collides protons together near the speed of light to reproduce the conditions of the Universe right after the Big Bang. It possesses all the features of Big Data: 1e16 collisions are produced each year, each producing 1000 particles and each of these particle leaving a complex signature in the 100 million electronic channels of the ATLAS detector. This project will initiate a collaboration between data scientists and physicists to develop the application of machine learning to the analysis of the LHC data.

Valorisation des données et Optimisation Robuste pour guider la Transition Énergétique vers des réseauX intelligents à forte composante renouvelable (VORTEX)

Une modélisation multiéchelles consistant en une famille de modèles hiérarchisés et opérant à des échelles de temps croissantes (journée / semaine à mois / horizon de trente ans), et des outils mathématiques adaptés (jeux à champ moyen répétés, apprentissage machine, optimisation convexe et robuste), sont proposés comme base pour une gestion raisonnée de la transition vers des réseaux électriques intelligents à forte composante renouvelable. Notre projet proposera en particulier des outils pour aider à la maîtrise de la demande énergétique dans un contexte régional.

Real-time Optimal Order Placement Strategies and Limit Order Trading Activity

Our primary objective is to identify how institutional investors can reduce their risk and trading costs by optimizing when and how to execute their trades. Limit order trading activity is an important state variable for this optimization problem in today’s financial markets where most liquidity is provided by limit orders. We thus plan to first analyze how risk and trading costs are affected by limit order trading activity using a novel, large-scale, ultra-high-frequency trading data set. We will then use our findings to guide us in modeling these effects and devising real-time optimal order placement strategies.

Exploiting ML/OR Synergies for Assortment Optimization and Recommender Systems

We propose to exploit synergies between assortment optimization and recommender systems on the application level, and the interplay between machine learning and mathematical programming on the methodological level. Rank-based choice models, estimated in a purely data-driven manner will introduce diversity into recommender systems, and supervised learning methods will improve the scalability and efficiency of assortment optimization in retail.

Data-driven Transplantation Science

End-stage kidney disease is a severe condition with a rising incidence, currently affecting over 40,000 Canadians.

The decision to accept or refuse an organ for transplantation is an important one, as the donor’s characteristics are strongly associated with the long-term survival of the transplanted kidney. In partnership with their health care provider, the transplant candidates need to answer two questions: (1) How long is the kidney from this specific donor expected to last for me? (2) If I refuse this specific donor, how much longer am I expected to wait before getting a better kidney?

We propose to use deep learning to predict the success of a possible matching. The results will contribute to build a clinical decision support tool answering the two questions above and helping transplant physicians and candidates to make the best decision. In addition, the quality of the matching can be the input of optimization algorithms designed to improve social welfare of organ allocations.

Développement d’une méthodologie pour l’’utilisation des données massives issues de compteurs intelligents pour modéliser un parc de bâtiments

Les données disponibles grâce à la généralisation des compteurs communicants représentent une grande opportunité pour améliorer les modèles de parc de bâtiments et les modèles plus généraux de flux énergétiques, mais les connaissances fondamentales à ce sujet sont encore limitées. Le présent projet vise à y remédier en développant une méthodologie permettant d’’utiliser les données massives des compteurs électriques communicants pour caractériser et calibrer, notamment par modélisation inverse, des archétypes de bâtiments qui pourront être intégrés dans le modèle TIMES.

Transformative adversarial networks for medical imaging applications

Following the concept of Generative adversarial networks (GANs), we propose to explore transformative adversarial training techniques where our goal is to transform medical imaging data to a target reference space as a way of normalizing them for image intensity, patient anatomy as well as the many other parameters associated with the variability inherent to medical images. This approach will be investigated both for data normalization and data augmentation strategy, and will be tested in several multi-center clinical data for lesion segmentation and/or classification (diagnosis).

Road user tracking and trajectory clustering for intelligent transportation systems

While traffic cameras are a mainstay of traffic management centers, video data is still most commonly watched by traffic operators for traffic monitoring and incident management. There are still few applications of computer vision in ITS, apart from integrated sensors for specific data extraction such as road users (RUs) counts. One of the most useful data to extract from video is the trajectory of all RUs, including cars, trucks, bicycles and pedestrians. Since traffic videos include many RUs, finding their individual trajectory is challenging. Our first objective is therefore to track all individual RUs. The second objective is to interpret the very large number of trajectories that can be obtained. This can be done by clustering trajectories, which provides the main motions in the traffic scene corresponding to RU activities and behaviors, al

The Electricity Demand Response Potential of the Montreal Metropolitan Community: Assessment of Potential Impacts and Options

This project will develop a clear understanding of the potential benefits and trade-offs of key stakeholders for deploying significant electric power demand response (DR) in the Montreal Metropolitan Community (MMC) area. It is motivated primarily by the desire of Hydro-Québec to increase its export potential, while at the same time by the need to assess DR deployment scenarios and their impacts on the people and businesses of the MMC. Data science is at the heart of this work which will need to discover knowledge on electricity consumption in order to learn how to leverage and control its flexibility.

Towards personalized medicine in the management of epilepsy: a machine learning approach in the interpretation of large-scale genomic data

To date, more than 150 epilepsy genes have been identified explaining around 35% of the cases. However, conventional genomics methods have failed to explain the full spectrum of epilepsy heritability, as well as antiepileptic drug resistance. In particular, conventional studies lack the ability to capture the full complexity of the human genome, such as interactions between genomic variations (epistasis). In this project, we will investigate how we can use machine learning algorithms in the analyses of genomic data in order to detect multivariate patterns, by taking advantage of our large dataset of individual epilepsy genomes. In this multi-disciplinary project, neurologists, geneticists, bio-informaticians and computational scientists will join forces in order to use machine learning algorithms to detect genomic variants signatures in patients with pharmaco-resistant epilepsy. Having the ability to predict pharmaco-resistance will ultimately reduce the burden of the disease.

A machine learning approach to decipher protein-protein interactions in human plasma

Proteins circulating in the human bloodstream make very useful and accessible clinical biomarkers for disease diagnostics, prognostics and theranostics. Typically, to perform their functions, proteins will interact with other molecules, including other proteins. These protein-protein interactions provide valuable insights into a protein’s role and function in humans; it can also lead to the discovery of novel biomarkers for diseases in which the protein of interest is involved. However, the identification of such interactions in human plasma is highly challenging. The lack of proper biochemical controls, which are inherently noisy, makes the confidence assessment of these interactions very difficult. We therefore propose to develop a novel machine learning approach that will extract the relevant signal from noisy controls to confidently decipher the interactome of clinically-relevant proteins circulating in the human bloodstream with the ultimate goal of identifying novel biomarkers.

Simulation and regression approaches in hydropower optimization

We develop optimization algorithms based on dynamic programming with simulations and regression, essentially Q-learning algorithms. Our main application area is hydropower optimization, a stochastic control problem where optimal releases of water are sought at each point in time.

Matching individuals to review tasks based on topical expertise level

The task of selecting an expert to review a paper addresses the general problem of finding a match between a human and an assignment based on the quality of expertise alignment between the two. State of the art approaches generally rely on modeling reviewers as a distribution of topic expertise, or as a set of keywords. Yet, two expert can have the same relative topic distribution and have wide differences in their depth of understanding. A similar argument can be made for papers. The objective of this proposal is to enhance the assignment approach to include the notions of (1) reviewer mastery of a topic, and (2) paper topic sophistication. Means to assess each aspect are proposed, along with approaches to assignments based on this additional information.

Asymmetric Information Tests with Dynamic Machine Learning and Panel Data

To our knowledge, the econometric estimation of dynamic panel data models with machine learning is not very developed and tests for the presence of asymmetric information in this environment are lacking. Most often, researchers assume the presence of asymmetric information and propose models (sometimes dynamic) to reduce its effects but do not test for residual asymmetric information in final models. Potential non-optimal pricing of financial products may still be present. Moreover, it is often assumed that asymmetric information is exogenous and related to unobservable agent characteristics (adverse selection) without considering agents’ dynamic behavior over time (moral hazard). Our goal is to use machine learning models to develop new tests of asymmetric information in large panel data sets where the dynamic behavior of agents is observed. Applications in credit risk, high frequency trading, bank securitization, and insurance will be provided.

Improving the prediction of the emotional and cognitive experience of users (UX) in interaction with technology using deep learning.

The objective of this research project is to leverage new advances in artificial intelligence, and more specifically deep learning approaches, to improve the prediction of emotional and cognitive experience of users (UX) in interaction with technology. What users experience emotionally and cognitively when interacting with an interface is a key determinant of the success or failure of digital products and services. Traditionally, user experience has been assessed with post hoc explicit measures, (i.e. such as questionnaires. However, these measures are unable to capture the states of users while they interact with technology. Researchers are turning to neuroscience implicit measures to capture the user’s states through psychophysiological inference. Deep learning has recently enabled other fields such as image recognition to make significant progress and we expect that it will do the same for psychophysiological inference, allowing the automatic modeling of complex feature sets.

L’utilisation des données financières à haute fréquence pour l’estimation de modèles financiers complexes

Les modèles de marché permettant de reproduire la complexité des interactions entre l’actif sous-jacent et les options requièrent une complexité qui rend leur estimation très difficile. Ce projet de recherche propose d’utiliser les données financières d’options à haute fréquence afin de mieux mesurer et gérer les différents risques du marché.

Nouvelles approches pour la modélisation et la résolution de problèmes de livraisons à domicile

Ce projet porte sur le développement de nouvelles approches permettant de mieux aborder les problèmes de livraisons à domicile qui, suite à l’avènement généralisé du commerce électronique, ont connu un essor très important au cours de la dernière décennie. Une partie des travaux portera sur la modélisation même de ces problèmes, notamment en ce qui concerne les objectifs poursuivis par les expéditeurs. Le reste du projet visera sur le développement d’’heuristiques et de méta-heuristiques à la fine pointe des connaissances pour la résolution efficace de problèmes de grande taille.

Optimization and machine learning for fleet management of autonomous electric shuttles

Recently, a Canada-France team of 11 researchers led by Bernard Gendron (DIRO-CIRRELT, UdeM) has submitted an NSERC-ANR strategic project « Trustworthy, Safe and Smart EcoMobility-on-Demand », supported by private and public partners on both sides of the Atlantic: in Canada, GIRO and the City of Montreal; in France, Navya and the City of Valenciennes. The objective of this project is to develop optimization models and methods for planning and managing a fleet of autonomous electric shuttle vehicles. As a significant and valuable additional contribution to this large-scale project, we plan to study the impact of combining optimization and machine learning to improve the performance of the proposed models and methods.

Deep Learning Methods in Biomedical Research: from Genomics to Multi-Omics Approaches

Deep learning approaches represent a promising avenue to make important advances in biomedical science. Here, we propose to develop, implement and use deep learning techniques to combine genomic data with multiple types of biomedical information (eg. other omics datasets, clinical information) to obtain a more complete and actionable picture of the risk profile of a patient. In this project, we will be addressing the important problem of missing data and incomplete datasets, evaluating the potential of these approaches for prediction of relevant medical phenotypes in population and clinical samples, and developing integration strategies for large heterogeneous datasets. The efficient and integrated use of multiomic data could lead to the improvement of disease risk and treatment outcome predictions in the context of precision medicine.

Modeling and predicting the effect of genetic variants on brain structure and function

Neurodevelopmental disorders (NDs) represent a significant health burden. The genetic contribution to NDs is approximately 80%. Whole genome testing in pediatrics is a routine procedure and mutations contributing significantly to neurodevelopmental disorders are identified in over 400 patients every year at the Sainte Justine Hospital. However, the impact of these mutations on cognition and brain structure and function is mostly unknown. However, mounting evidence suggests that genes that share similar characteristics produce similar effects on cognitive and neural systems.

Our goal: Develop models to understand the effects of mutations, genome-wide, on cognition, brain structure and connectivity.

Models will be developed using large cohorts of individuals for whom, genetic, cognitive and neuroimaging data was collected.

Deliverable: Algorithms allowing clinicians to understand the contribution of mutations to the neurodevelopmental symptoms observed in their patients.

From data-science to brain-science: AI-powered investgation of the neuronal determinants of cognitive capacities in health, aging and dementia

Artificial intelligence is revolutionizing science, technology and almost all aspects of our society. Learning algorithms that have shown astonishing performances in computer vision and speech recognition are also expected to lead to qualitative leaps in biological and biomedical sciences. In this multi-disciplinary research program, we propose to investigate the possibility of boosting information yield in basic and clinical neuroscience research by applying data-driven approaches, including shallow and deep learning, to electroencephalography (EEG) and magnetoencephalography (MEG) data in (a) healthy adults, and aging populations (b) with or (c) without dementia. The proposal brings together several scientists with expertise in a wide range of domains, ranging from data science, mathematics and engineering to neuroimaging, systems, cognitive and clinical neuroscience.

From data-science to brain-science: AI-powered investgation of the neuronal determinants of cognitive capacities in health, aging and dementia

Artificial intelligence is revolutionizing science, technology and almost all aspects of our society. Learning algorithms that have shown astonishing performances in computer vision and speech recognition are also expected to lead to qualitative leaps in biological and biomedical sciences. In this multi-disciplinary research program, we propose to investigate the possibility of boosting information yield in basic and clinical neuroscience research by applying data-driven approaches, including shallow and deep learning, to electroencephalography (EEG) and magnetoencephalography (MEG) data in (a) healthy adults, and aging populations (b) with or (c) without dementia. The proposal brings together several scientists with expertise in a wide range of domains, ranging from data science, mathematics and engineering to neuroimaging, systems, cognitive and clinical neuroscience.

Monte Carlo and Quasi-Monte Carlo Methods for Optimization and Machine Learning

The use of Monte Carlo methods (aka, stochastic simulation) has grown tremendously in the last few decades. They a now a central ingredient in many areas, including computational statistics, machine learning, and operations research. Our aim in this project is to study Monte Carlo methods and improve their efficiency, with a focus on applications to statistical modeling with big data, machine learning, and optimization. We are particularly interested in developing methods for which the error converges at a faster rate than straightforward Monte Carlo. We plan to free software that implements these methods.

Data analytics methods for travel time estimation in transportation engineering

Travel time is considered as one of the most important performance measures in urban mobility. It is used by both network operators and drivers as an indicator of quality
of service or as a metric influencing travel decisions. This proposal tackles the issue of travel time prediction from several angles: i) data pre-processing (map-matching), ii) short-term travel time prediction and iii) long-term travel time prediction. These tasks will require the development of new approaches in statistical and machine learning to adequately model GPS trajectory data and to quantify the prediction error.

Machine learning technology applied to the discovery of new vibrational spectroscopy biomarkers for the prognostication of intermediate-risk prostate cancer patients

Prostate cancer is the most frequent cancer among Canadian men, with approximately 25,000 diagnoses per year. Men with high risk and low risk disease almost always experience predictable disease evolution allowing optimal treatment selection. However, none of the existing clinical tests, imaging techniques or histopathology methods can be used to predict the fate of men with intermediate-risk disease. This is the source of a very important unmet clinical need, because while some of these patients remain free of disease for several years, in others cancer recurs rapidly after treatment. Using biopsy samples in tissue microarrays from 104 intermediate-risk prostate cancer patients with known outcome, we will use a newly developed Raman microspectroscopy technique along with machine learning technology to develop inexpensive prognostic tests to determine the risk of recurrence allowing clinicians to consider more aggressive treatments for patients with high

Developing a machine learning framework to dissect gene expression control in subcellular space

Our multidisciplinary team will develop and use an array of machine learning approaches to study a fundamental but poorly understood process in molecular biology, the subcellular localization of messenger RNAs, whereby the transcripts of different human genes are transported to various regions of the cell prior to translation. The project will entail the development of new learning approaches (learning from both RNA sequence and structure data, phylogenetically related training examples, batch active learning) combined with new biotechnologies (large-scale assays of both natural and synthetic RNA sequences) to yield mechanistic insights into the « localization code » and help understand its role in health and disease.

Deep learning for precision medicine by joint analysis of gene expression profiles measured through RNA-Seq and microarrays

This project aims at developing domain adaptation techniques to enable the joint analysis of gene expression profiles datasets acquired using different technologies, such as RNA-Seq and microarrays. Doing so will leverage the large number of gene expression profiles publicly available, avoiding the typical problems and limitations caused by working with small datasets. More specifically, methods developed will be continuously applied to datasets available for Acute Myeloid Leukemia in which the team has extensive expertise.

Machine Learning for (Discrete) Optimization

The interaction between Machine Learning and Mathematical Optimization is currently one of the most popular topics at the intersection of Computer Science and Applied Mathematics. While the role of Continuous Optimization within Machine Learning is well known, and, on the applied side, it is rather easy to name areas in which data-driven Optimization boosted by / paired with Machine Learning algorithms can have a game-changing impact, the relationship and the interaction between Machine Learning and Discrete Optimization is largely unexplored. This project concerns one aspect of it, namely the use of modern Machine Learning techniques within / for Discrete Optimization.

Gibbs-repulsion and determinantal processes for statistical learning

Non-parametric Bayesian models are very popular for density estimation and clustering. However, they have a tendency to use too many mixture components due to their use of independent parameter priors. Repulsion processes priors such as determinantal processes, solve this issue by putting higher mass on parameter configurations for which the mixture components are well separated. We propose the use of Gibbs-like repulsion processes which are locally determinantal, or adaptive determinantal processes as priors for modeling density estimation, clustering, and temporal and/or spatial data.

Using machine learning to uncover how marketer-generated post content is associated with user-generated content and revenue

This projects proposes how machine learning can be used to improve a company’s communication with its customers in order to increase sales. To that end, we will identify how broadcaster-generated content is associated with user-generated content and revenue measures. In addition, we intend to automate the identification of post content, and to propose personalized recurrent neural networks to identify the writing styles of brands and companies and automate the creation of online content.

Real-time detection and prediction of epileptic seizures using deep learning on sparse wavelet representations

Epilepsy is a chronic neurological condition in which about 20% of patients do not benefit from any form of treatment. In order to diminish the impact of recurring seizures on their lives, we propose to exploit the potential of artificial intelligence techniques for predicting the occruence of seizures and detecting their early onset, such as to issue warnings to patients. The aim of this project is thus to develop an efficient algorithm based on deep neural networks for performing real-time detection and prediction of seizures. This work will pave the way for the development of intelligent implantable sensors coupled with alert systems and on-site treatment delivery.

Knowledge-based inference for question answering and information retrieval

Question answering (QA) is a typical NLP/AI problem with wide applications. A typical approach first retrieves relevant text passages and then determines the answer from them. These steps are usually performed separately, undermining the quality of the answers. In this project, we aim at developing new methods for QA in which the two steps can benefit from each other. On one hand, inference based on a knowledge graph will be used to enhance the passage retrieval step; on the other hand, the retrieved passages will be incorpor

Statistical modelling with distributed systems

Statistical inference requires a large toolbox of models and algorithms that can accommodate different structures in the data. Modern datasets are often stored on distributed systems where the data are scattered across a number of nodes with limited bandwidth between them. As a consequence, many complex statistical models cannot be computed natively on those clusters. In this project, we will advance statistical modeling contributions to data science by creating solutions that are ideally suited for analysis on distributed systems.

Learning independently controllable features with application to robotics

Learning good representations is key for intelligent systems. One intuition is that good features will disentangle distinct factors that explain variability in the data, thereby leading to the potential development of causal reasoning models. We propose to tackle this fundamental problem using deep learning and reinforcement learning. Specifically, a system will be trained to discover simultaneously features that can be controlled independently, as well as the policies that control them. We will validate the proposed methods in simulations, as well as by using a robotic wheelchair platform developed at McGill University.

Analytics and optimization in a digital humanitarian context

When responding to humanitarian crises, the lack of information increases the overall uncertainty. This hampers relief efforts efficiency and can amplify the damages. In this context, technological advances such as satellite imaging and social networks can support data gathering and processing to improve situational awareness. For example, volunteer technical communities leverage ingenious crowdsourcing solutions to make sense of a vast volume of data to virtually support relief efforts in real time. This research project builds on such digital humanitarianism initiatives through the development of innovative tools that allow evidence-based decision making. The aim is to test the proposed methodological framework to show how data analytics can be combined with optimization to process multiple sources of data, and thus provide timely and reliabl

Learning representations of uncertainty for decision making processes

Decision support and optimization tools are playing an increasingly important role in today’s economy. The vast majority of such systems, however, assume the data is either deterministic or follows a certain form of theoretical probability functions. We aim to develop data driven representations of uncertainty, based on modern machine learning architectures such as probabilistic deep neural networks, to capture complex and nonlinear interactions. Such representations are then used in stochastic optimization and decision processes in the fields of cancer treatment, supply chain and finance.

Fundamental Challenges for Big Data Fusion and Strategic Transportation Planning

As more and more transportation data becomes continuously available, transportation engineers and planners are ill-equipped to make use of it in a systematic and integrated way. This project aims to develop new machine learning methods to combine transportation data streams of various nature, spatial and temporal definitions and pertaining to different populations. The resulting model will provide a more complete picture of the travel demand for all modes and help better evaluate transportation plans. This project will rely on several large transportation datasets.

Optimised Hardware-Architecture Synthesis for Deep Learning

Deep learning requires considerable computing power. Computing power can be improved significantly by designing application specific computing engines dedicated to deep learning. The proposed project consists of designing and implementing a High Level Synthesis tool that will generate an RTL design from the code of an algorithm. This tool will optimize the architecture, the number of computing units, the length and representation of the numbers and  the important parameters of the various memories generated.

Equilibrium Propagation Framework: Analog Implementation for Improved Performances (Equipe)

The main aim of this project is to implement the Equilibrium Propagation (EP) algorithm in analog circuits, rather than digital building blocks, to take advantage of their higher computation speed and power efficiency. EP involves minimization of an energy function, which requires a long relaxation phase that is costly (in terms of time) to simulate on digital hardware. But it can be accelerated through analog circuit implementation. Two main implementation phases in this project are: (1) Quick prototyping and proof of concep using an FPAA platform (RASP 3.0), and (2) High performance custom System-on-Chip (SoC) implementation using a standard CMOS process e.g. 65nm to optimize the area, speed, and power consumption.

Combiner l’apprentissage automatique et la recherche opérationnelle pour traiter plus rapidement les grands problèmes d’horaires d’équipages aériens

Nous travaux récents portent sur le développement d’algorithmes d’optimisation exacts qui profitent de l’information a priori sur les solutions attendues pour réduire le nombre de variables et de contraintes à traiter simultanément. L’objectif est de développer un système d’apprentissage machine pour obtenir l’’information permettant d’accélérer le plus possible ces algorithmes d’’optimisation, pour traiter de plus grands problèmes d’’horaires d’’équipages aériens. Ce projet produira en plus des avancements en R. O. des avancements en apprentissage sous contraintes et par renforcement.

Predictive model of colorectal cancer liver metastases response to chemotherapy

Colon cancer is the 2nd leading cause of mortality in Canada. In patients with colorectal liver metastases, response to chemotherapy is the main determinant of patient survival. Our multidisciplinary team will develop models based to predict response to chemotherapy and patient prognosis using the most recent innovations in deep learning architectures. We will train our model on data from an institutional biobank and validate our model on independent provincial imaging and medico-administrative databases.

Matching MHC I-associated peptide spectra to sequencing reads using deep neural networks

Identification of MHC I-associated peptides (MAPs) unique to a patient or tumor is key step in developing efficacious cancer immunotherapy. This project aims at developing a novel approach for exploiting Deep Neural Networks (DNN) for the identification of MAPS based on a combination of next-generation sequencing (RNA-Seq) and tandem mass spectrometry (MS/ MS). The proposed developments will take advantage of a unique dataset of approximately 60,000 (MS/MS – sequence) pairs assembled by our team. The project will also bring together researchers from broad horizons: mass spectrometry, bioinformatics, machine learning and cancer immunology.

Our project dedicates to deal with uncertainty in drone routing for disaster response and relief operations. To tackle the uncertainties arose from disaster scenarios, like uncertain demand locations and quantities for relief supplies, we use data-driven robust optimization (RO) method. This technique protects the decision makers against parameter ambiguity and stochastic uncertainty by using uncertainty sets. Therefore, it is significant to set proper parameters for the uncertainty set: a small set cannot accurately capture possible risks while a larger one may lead to overly conservative solutions. To address this problem, we use machine learning (ML) technique to extract information from historical data and real-time observations, and create the parameters by ML algorithms. After calibrating the uncertainty set, we will determine appropriate models for the problem by considering various theories in RO, such as static RO and multiple stage adjustable RO. These approaches will be measured against other applicable approaches such as stochastic programming.

Ce projet de recherche a pour objectif de développer et mettre en application des techniques d’apprentissage machine afin de grandement améliorer l’identification des électrons par le détecteur ATLAS du LHC, le plus grand accélérateur de particules jamais construit et l’un des projets scientifiques les plus ambitieux de tous les temps.Afin de mener à bien le programme d’ATLAS, il est nécessaire d’identifier et mesurer chacune des particules, lesquelles s’y créer à un taux de 40 milliards par seconde et génèrent un flot astronomique de données. Parmi celles-ci, les électrons revêtent une très grande importance, mais ils sont également excessivement rares, ne représentant qu’une infime fraction. Considérant la taille et complexité des données disponibles, le problème d’identification des particules aussi rares que les électrons constitue un terrain d’application idéal pour les méthodes d’apprentissage machine. Les algorithmes actuels d’identification des électrons sont très simples et ne font pas usage de ces méthodes de telle sorte qu’une percée dans ce domaine serait une première mondiale qui pourrait éventuellement paver la voie à des découvertes majeures en physique des particules.

Multiple sclerosis (MS) is a disease, with a high rate in Canada, that leads to major sensory and motor problems. This disease affects the neuronal signal transmission in both brain and spinal cord, creating lesions, which are observable on images acquired with an MRI scanner. The count and volume of lesions on an MRI scan of a patient is a crucial indicator of the disease status and commonly used by doctors for the diagnosis, prognosis and therapeutic drug trials. However, the detection of lesions is very challenging and time consuming for radiologists, due to the high variability of their size and shape.This project aims at developing a new, automatic and fast method of MS lesion detection on MRI data of spinal cord, based on newly developed machine learning algorithms. The new algorithm’s performance will be tested on a large dataset involving patients coming from different hospitals in the world. Once the algorithm is optimized, it will be freely available as part of an open-source software, already widely used for spinal cord MRI processing and analysis. A fundamental goal of this project is the integration of this algorithm in hospitals to help radiologists in their daily work.

When we are walking through a crowd, or playing a sport, our brain continuously makes decisions about directions to go to, obstacles to avoid and information to pay attention to. Fuelled by the successful combinations of quantitative modeling and neural recordings in nonhuman primates, research into the temporal dynamics of decision-making has brought the study of decision-making to the fore within neuroscience and psychology, and has exemplified the benefits of convergent mathematical and biological approaches to understanding brain function. However, studies have yet to uncover the complex dynamics of large-scale neural networks involved in dynamic decision-making in humans. The present research aims to use advanced data analytics to identify the neural features involved in tracking the state of sensory evidence and confirming the commitment to a choice during a dynamic decision-making task. To this end, we will use cutting-edge electrophysiological brain imaging (magnetoencephalography, MEG), combined with multivariate machine learning algorithms. This project, for the first time, will shed light on the whole-brain large-scale dynamics involved in dynamic decision-making, thus providing empirical evidence that can be generalized across subjects to test and refine computational models and neuroscientific accounts of decision-making. By providing a quantitative link between the behavioral and neural dynamics subserving how decisions are continuously formed in the brain, this project will contribute to expose mechanisms that are likely to figure prominently in human cognition, in health and disease. Moreover, this research may provide neurobiological-inspired contributions to machine learning algorithms that implement computationally efficient gating functions capable of making decisions in a dynamically changing environment. ln addition to advancing our knowledge of the way human brains come to a decision, we also foresee long-term health implications for disorders such as Parkinson’s disease.

When making decisions, medical professionals often rely on past experience and their own judgment. However, it is often the case that an individual decision-makerfaces a situation that is unfamiliar to him or her. An adaptive treatment strategy (ATS) can help such biomedical experts in their decision-making, as they are a statistical representation of a decision algorithm for a given treatment that optimizes patient outcomes. ATSs are estimated with large amounts of data, but an issue that may occur is that such sources of data may be subject to unmeasured confounding, whereby important variables needed to ensure the causal inference are missing. The idea behind this research project is to develop a sensitivity analysis to better understand and to quantify the impact of unmeasured confounding on decision rules in ATSs.

Language understanding and generation is a unique capacity of humans. Automatic summarization is an important task in Natural (human) Language Processing. This task consists in reducing the size of discourse while preserving information content. Abstractive summarization sets itself apart from other types of summarization since it most closely relates to how humans would summarize a book, a movie, an article or a conversation. From a research standpoint, automatic abstractive summarization is interesting since it requires models to both understand and generate human language. In the past year, we have seen research that have improved the ability of Neural Networks to choose the most important parts of discourse while beginning to address key pain points (e.g. repeating sentences, nonsensical formulations) during summary text generation. Recent techniques in Computer Vision image generation tasks have shown that image quality can be further improved using Generative Adversarial Networks (GAN). Our intuition is that the same is true for a Natural Language Processing task. We propose to incorporate newest GAN architectures into some of the most novel abstractive summarization models to validate our hypothesis. The objective is to create a state-of-the-art summarization system that most closely mimics human summarizers. This outcome will also bring us closer to understand GANs analytically.

Combiner l’A.I. et la R.O. pour optimiser les blocs mensuels d’équipages aérien.Nos travaux récents portent sur le développement de deux nouveaux algorithmes Improved Primal Simplex (IPS) et Integral Simplex Using Decomposition (ISUD) qui profitent de l’information a priori sur les solutions attendues pour réduire le nombre de variables et de contraintes à traiter simultanément. Actuellement cette information est donnée par des règles fournies par les planificateurs. L’objectif de recherche sera de développer un système utilisant l’intelligence artificielle (IA) pour estimer la probabilité que la variable liant deux rotations fasse partie de la solution d’un problème de blocs mensuels d’équipages aériens. L’apprentissage se fera sur les données historiques de plusieurs mois, de plusieurs types d’avions et de plusieurs compagnies. L’estimation des probabilités doit se faire à partir des caractéristiques des rotations et non à partir de leurs noms. Une rotation ne revient pas d’une compagnie à l’autre ni d’un mois à l’autre. Il faudra identifier les caractéristiques pertinentes. Il faudra de la recherche sur l’apprentissage pour profiter des contraintes du problème. Il y a des contraintes entre le personnel terminant des rotations et celui en commençant par la suite. La validation de l’apprentissage se fera en alimentant les optimiseurs avec l’information estimée et en observant la qualité des solutions obtenues et les temps de calcul. Il y aura de la recherche à faire dans les optimiseurs pour exploiter au mieux cette nouvelle information.

Le but de ce projet est de créer un prototype capable de générer des métriques importantes pour les systèmes de transport intelligents. La solution proposée utilisera la technologie 3D LiDAR pour mesurer la distance des objets (stationnaire et mobile), les compter, mesurer leur vitesse et identifier leur classe en temps réel. Ces données seront disponibles localement ou via une plate-forme infonuagique. Ces métriques pourront être utilisées pour l’optimisation des feux de signalisation en temps réel, faire des analyses de sécurité et permettre une meilleure gestion et planification du réseau de transport.

L’objectif de ce projet est d’ajouter des fonctionnalités et d’améliorer les performances des modèles d’optimisation actuels de la plateforme Gray. Mise au point durant la thèse, la plateforme Gray atténue la barrière d’entrée pour la réalisation de recherche, même de base, sur la planification du traitement contre le cancer. L’objectif est d’offrir un produit compétitif sur le marché qui tirera parti de l’apprentissage profond pour aider à une planification de traitement de radiothérapie adaptative.

To turn the ever increasing amounts of data into social and economic value, two tasks need to be performed: 1) extracting knowledge from the data, and 2) incorporating this knowledge in the operations that drive the society. The machine learning community addresses the first task by extracting the knowledge from the data and capturing it into ‘learned models’. The second task is studied by the operations research community under the label of ‘optimization’. However, these techniques have been developed almost independently. This makes it less straightforward to integrate them and turn the knowledge obtained from a learned model into actionable decisions. In this project, we exploit the fundamental similarities between the two tasks to develop an integrated system that performs both tasks together. We apply our system to problems in business and finance and demonstrate how this approach can help players in these sectors to use their data for improving their operations.

This research is focused on using mathematical tools to accelerate the training time of Deep Neural Networks (DNN)s. DNNs are a powerful tool in Artificial Intelligence, behind applications in machine translation, image recognition, speech recognition and other areas. However training the DNNs requires huge computational resources, which is costly both financially, and in the human effort required to implement them. This research will use advanced mathematical tools to improve the time required to train DNNs.

This project deals with the planning of intermodal rail transportation, integrating methodologies from operations research and machine learning in a new and innovative way. Intermodal container freight transportation is the backbone of international trade and supports a large part of Canadian and North-American imports and exports. Canada has one of the largest rail networks in the world and Canadian railway companies are both network and terminal operators. They face many large-scale optimization problems that are complex because of their sheer size and the uncertainty that affects planning and operations on a continuous basis. The project focuses on a tactical network load and block planning problem that involves decisions related to blocking and railcar fleet management. Assuming that the train schedule is given, the problem entails three consolidation processes: assignment of containers to railcars, of railcars to blocks and of blocks to trains. The project will be dedicated to designing a service network design model and associated solution method that allows to solve realistic, large scale, instances.

When designing a machine learning algorithm, it is crucial for the designer to understand the input data to which this algorithm will be applied. It is well known that real-world data for any task has a lot of structure, exploiting which allows for faster learning and more accurate prediction. However, understanding this structure is a highly nontrivial task, given the high dimension of the data. In this project we propose to develop a mathematical framework for learning the structure hidden in the data, via the lens of probability theory. Assuming the data is generated by some stochastic process, we would like to infer its distributional properties. Then a natural question is, which distributions are harder to learn, and which ones are easier. The aim of this project is to answer this question from statistical and computational perspectives, at least for a variety of commonly used classes of distributions, such as mixture models and graphical models.

Gestion en temps réel du cargo aérien. Le projet à moyen terme est le développement d’un système de gestion du cargo dans les compagnies aériennes en commençant avec Air Canada. Ce système traitera la planification stratégique, tactique et l’opération en temps-réel. Le niveau stratégique évalue des scénarios à long terme sur l’organisation du réseau, les marchés à développer, les alliances à conclure. Le niveau tactique optimise le choix des itinéraires entre chaque paire de villes pour une semaine type d’une saison. Durant l’opération, les vendeurs pourront obtenir en ligne le meilleur itinéraire pour acheminer une nouvelle commande et le prix de revient. À chaque niveau de décision, il faut estimer la demande pour l’horizon considéré et optimiser l’acheminement de cette demande dans le réseau de transport comprenant des avions tout cargo, l’espace disponible dans les soutes des vols passagers, des sous-contrats avec d’autres transporteurs aériens et routiers. La recherche portera sur le développement de nouvelles méthodes d’estimation de la demande et d’optimisation de l’acheminement dans un grand réseau.

The main objective in using optimization approaches for demand and supply management in home healthcare is to match supply and demand by influencing patients/caregivers choices for service time slots/working shifts. Our aim with this project is to develop a decision support tool to deal with approaches for demand and supply management in home healthcare. Specifically, we will implement stochastic models to forecast future demands and to predict caregivers’ absenteeism. Then, we will design and develop choice models to consider patient and caregiver choice behavior. These models will predict the probability of choosing a particular alternative from an offered set (e.g. visit time slots, working shifts) given historical choice data about an individual or a segment of similar individuals. These models will be embedded into an optimization approach that will compute a time slotting/scheduling plan or a pricing strategy to optimally balance the allocation of cost-effective schedules to caregivers and the improvement of service quality.

Every day, we experience thousands of situations, but we only remember few of them. Episodic memory is the only memory system that allows people to consciously re-experience past experiences and it is the most sensitive to age and neurodegenerative diseases. It is thus critical to better understand how to enhance learning and memory in both healthy and clinical populations. Emotions are known to robustly strengthen the formation of long-term memories. Characterizing the influence of positive emotions (joy, happiness) on memory could be pivotal in improving memory therapies, yet the underlying brain mechanisms are still surprisingly misunderstood. In this project, we will use a fully data-driven approach to identify the key neuronal processes strengthened by positive emotions that distinguish events we will durably remember from events we will forget. We will combine for the first time high spatial and temporal resolution brain imaging techniques and state-of-the-art machine-learning algorithms. This will be achieved by assessing the ability of multidimensional (across space, time and frequency) arrays of brain data to predict future memory accuracy.

Ride-sharing services such as Uber, Lyft or Didi Chuxing match a group of drivers providing rides with customers through an online ride-sharing platform. This business model faces a number of fundamental challenges. Indeed, the drivers considered as independent contractors choose the area they wish to serve, if they accept or reject rides, and when they start and stop working. With no direct control over the drivers, the ride-sharing platform can only use incentives and select the information it provides to drivers and customers in order to improve the quality of service and balance supply and demand. This project aims to develop a model that anticipates how the drivers respond to provided information, which is a combination of request statistics, prices at various locations and times, and estimation of the state of the road network. Moreover, it intends to generate location and time-dependent pricing schemes and optimal information filters in order to optimize the efficiency of the system. For example, the filters control the amount of information to release to drivers about the requests in order to balance the supply and demand and avoid the drivers from deserting some areas.

Hydro-Québec is geographically well positioned to make significant profits in neighboring electricity markets. Facing political mandates to retire coal and nuclear power plants, the markets of Ontario, New York and New England are under increasing stress to provide stable, baseload electricity production. We will utilize the vast historical data from these markets to model their future evolution. Using the insights obtained from this analysis, we will be able to determine optimal strategies for Hydro-Québec to maximize its profits through targeted investment decisions in market interconnections. The results will be generalizable to other provincial utilities in Canada and their participation in the relevant electricity markets.

Acute myeloid leukemia (AML) is the most common form of leukemia in adults. Despite advances in supportive care to treat therapy-related complications, the majority of AML patients will not exceed the two-year survival mark because of relapse. This dismal outcome reflects the sub-optimal treatment orientation of poorly understood subtypes of AML. To improve the treatment and outcome of patients, Dr. Guy Sauvageau and colleagues initiated in 2009 the Leucegene project which has become an internationally acknowledged leader in genetic and biological characterization of AML. Exploiting the most innovative technologies, this program already allowed the sequencing of 452 primary human AML specimens. While several types of genomic alterations have been explored in AML, some of them, such as modifications to the chromosome structure, remain elusive despite their known importance in cancer. We are convinced that this is due to unsuitable analysis and we propose here an innovative machine learning approach to efficiently identify these modifications. Tests will be carried out on our sequenced AML specimens. Ultimately, the method will be released to help the scientific community to exploit its cancer data. From a biomedical point of view, it will allow for better definition of AML subgroups, as well as an increase in the chances of identifying new markers for this disease. With the goal to accelerate the transfer of new knowledge from the laboratory to the bedside, this project will help ensure the correct classification and treatment of AML.

To understand brain mechanism of cognitive functions is the ultimate goal of neuroscience studies, which also provides fundamental guidance for developing new techniques in artificial intelligence. With accumulated evidence in animals and humans, functional dynamics is suggested to be the non-stationary nature of cognitive process. In this project, we aim to apply deep learning models to characterize the spatial and temporal dynamics of BOLD signals at rest and during cognitive tasks. To account for the temporal dependence of MRI signals, a convolutional recurrent neural network is first used to characterize the spatial and temporal dynamics of resting-state data, and then to map the dynamic somatotopic maps during movement of tongue, hand and foot. The model is further adjusted for classification of functional dynamics among multiple task conditions. The derived characteristic functional dynamics, including sequential temporal response functions and corresponding activation patterns, reveals the dynamic process of human cognitive function and provides essential guidance for brain simulation. Furthermore, our proposed method could also be used in clinic applications, for instance searching for temporal and spatial biomarkers for Alzheimer’s disease and evaluating the treatment effects of precision medicine.

L’optimisation combinatoire occupe une place prépondérante dans notre société actuelle. Que ce soit la logistique, le transport ou la gestion financière, tous ses domaines se retrouvent confrontés à des problèmes pour lesquels on recherche la meilleure solution. Cependant, un grand nombre de problèmes très complexes reste encore hors de portée des méthodes d’optimisation actuelles. C’est pourquoi, l’amélioration de ces techniques est un sujet crucial. Parmi ces dernières, les diagrammes de décisions semblent avoir un avenir prometteur. Un diagramme de décision est une structure qui permet de représenter de manière compacte un problème tout en préservant ses caractéristiques. Cependant, leur efficacité est extrêmement dépendante de l’ordre des variables utilisé pour leur construction. L’objectif de ce projet est d’utiliser les méthodes récentes d’apprentissage automatique pour ordonner les variables lors de la construction d’un diagramme de décision. Les contributions de ce projet permettront la résolution de problèmes combinatoires plus complexes, et plus larges que ce que peuvent faire les méthodes de l’état de l’art. Nous nous consacrerons principalement aux problèmes réels liés au transport et à la logistique. Ce projet sera effectué en partenariat avec l’entreprise Element-AI.

Recent machine learning approaches have led to impressive demonstrations of machines solving a great variety of difficult tasks, which previously were thought to be restricted to humans. Applied to areas such as health care, environmental challenges, optimization of transport and logistics, or industrial processes, these advances will lead to improved living conditions and the creation of value. While being loosely inspired by biological neural systems, artificial neural networks starkly differ from their biological counterparts in almost every respect. In particular, brains can learn from very few examples, infer causal relationships, and seamlessly transfer skills to new tasks, whereas current machine learning models require enormous amounts of data to just master a single task. To overcome some of these limitations, we introduce new, brain-inspired models for learning and memory, which will allow for meaningful information to be extracted from data more efficiently. The resulting systems will lead to improved, more powerful machine learning systems, which can be applied in numerous contexts, including medical applications, automation, robotics, or forecasting.

A quarter million people every year are affected by spinal cord injury (SCI), which causes paraplegia. When the lesion is incomplete some recovery can occur. Spinal cord stimulation can be applied to help people with SCI to regain control of the paralyzed legs. In the last year Prof. Martinez and I demonstrated a new neuroprosthetic concept whereby cortical stimulation is applied to improve walking. This novel strategy empowers the brain’s own residual networks and increases voluntary control of leg movement with long lasting beneficial effects for recovery. « Fire together, wire together » is the established rule for neural repair. Here we propose to combine for the first time brain and spinal stimulation into an unique combined neuroprosthesis. This approach is compelling, but complicated by the overwhelming amount of stimulation parameters that needs to be characterized. We propose to solve this problem with machine learning. The first ever intelligent neuroprosthesis will monitor changes in muscular activity to explore and learn an optimal set of stimulation parameters. Our results can be rapidly translated to clinical tests.

Epilepsy is a chronic neurological condition that affects as many as 1 in every 100 Canadians. While first line of treatment consists of long-term drug therapy more than a third of patients suffer from seizures that are resistant to antiepileptic drugs. Due to their unpredictable nature, uncontrolled seizures represent a major personal handicap and source of worriment for patients. In addition, persistent seizures constitute a considerable public health burden due to high use of health care resources, high number of disability days or unemployment, and low annual income. Some of the difficulties and challenges faced by drug-refractory patients can be overpassed by implementing algorithms able to anticipate seizures. With accurate seizure forecasting, one could ameliorate refractory epilepsy management improving social integration, productivity and quality of life. Our main objective is the development of a real-time seizure prediction system, based on deep learning, intended to warn patients or caretakers about an incoming seizure and recommend advisory measures.

Drug combinations can simultaneously target redundant biological pathways and thus offer unique advantages for disease treatment. By growing human cancer cells each with a specific gene deletion in the presence of a drug, we identified gene deletions which make cells more sensitive or more resistant to the growth-inhibition effects of >230 different drugs. In this proposal, I outline a plan to develop software tools to exploit this resource in order to precisely characterize drug mechanisms and to predict useful drug combinations. Tumor growth relies on overactive biomass and energy production, and I found that close to 80% of the drugs we screened altered the sensitivity of cells to the deletion of metabolic genes. I will use a genome-scale mathematical model of cell metabolism to attribute these effects to the lowered activity of other metabolic genes, those whose activities we predict are directly affected by the drug. After modelling the effects of each drug on cell metabolism, we will simulate the effects of drug combinations to identify pairs which effectively block the generation of small molecules important to tumor growth.available data on the effectiveness of drug combinations, I will train a machine learning algorithm to use similar gene deletion data as well as drug molecular similarities to predict useful drug combinations for the treatment of cancer and potentially other diseases. I will also attempt to predict the direct molecular targets of each drug by modelling molecular signalling in cells, leveraging known signalling pathways, molecular interaction networks and pairs of gene deletions sensitive or resistant to similar sets of drugs.

The rise of antibiotic resistance has put antimicrobials, once believed to be miracles of modern medicine, into jeopardy1. The current death toll of AMR is ~800,000 per year (i.e., ~100 per hour) and is expected to rise to ~ 16 million in 2050. In Canada alone, the financial burden of antibiotic resistance is ~ $200 million annually. A key knowledge in our battle against antibiotic resistance is to predict the growth rate of bacteria at different concentrations of antibiotics. Recently, the response of bacterial strains to antibiotics were measured for all possible mutations in important enzymes that confer resistance to beta-lactam antibiotics (e.g., penicillins, ampicillins, etc.). In this project, I will employ the power of machine learning to develop predictive models of resistance at different antibiotic dosages from the available large-scale datasets. This approach would have immediate impacts on the design of antibiotic dosages that prevent or delay the onset of resistance. In this integration of machine learning with biochemistry and molecular medicine, we will seek the potentials of data science to aid decision-making in medicine.

Un défi majeur dans le secteur du bâtiment est l’’absence de systèmes continus de suivi de la performance et d’évaluation des écarts de performance entre la situation observée et celle désirée. L’opération non-optimale des systèmes de Chauffage, Ventilation et Conditionnement d’Air (CVCA) entraîne des pertes énergétiques et de confort des occupants. Le secteur du bâtiment représente environ un tiers de la consommation énergétique au Québec et au Canada. Ainsi, des mesures d’’efficacité dans ce secteur produisent des impacts positifs majeurs. L’idée de ce projet est d’utiliser des méthodes d’apprentissage profond pour exploiter les larges jeux de données générés par les systèmes CVCA. Le but final est d’automatiser la détection et le diagnostic des anomalies, et d’optimiser l’opération des équipements. Les bénéfices incluent une gestion améliorée de l’énergie et une meilleure prise en compte du confort des occupants. Au terme de ce projet, des outils de détection / diagnostic / contrôle basés sur l’apprentissage profond seront développés et testés à des fins d’implantation dans des bâtiments réels.

The increasing use of algorithmic decision-making in domains that affect people’s lives, has raised concerns about possible biases and discrimination that such systems might introduce. Recent concerns on algorithmic discrimination have motivated the development of fairness-aware mechanisms in the machine learning (ML) community and the operations research (OR) community, independently. While in fairness-aware ML, the focus is usually on ensuring that the inference and predictions produced by a learned model are fair, the OR community has developed methods to ensure fairness in solutions of an optimization problem. In this project, I plan to build on the complementary strengths of fairness methods in ML and OR to address these shortcomings in a fair data-driven decision-making system. I will apply this work to real-world problems in the areas of personalized education, employment hiring (business), social well-being (health), and network design (transportation). The advantage of my proposed system compared to the existing works is that it: 1) incorporates domain knowledge with data-driven probabilistic models, 2) detects and describes complex discriminative patterns, 3) returns a fair decision/policy, and 4) breaks negative/positive feedback loops.

The increasing use of algorithmic decision-making in domains that affect people’s lives, has raised concerns about possible biases and discrimination that such systems might introduce. Recent concerns on algorithmic discrimination have motivated the development of fairness-aware mechanisms in the machine learning (ML) community and the operations research (OR) community, independently. While in fairness-aware ML, the focus is usually on ensuring that the inference and predictions produced by a learned model are fair, the OR community has developed methods to ensure fairness in solutions of an optimization problem. In this project, I plan to build on the complementary strengths of fairness methods in ML and OR to address these shortcomings in a fair data-driven decision-making system. I will apply this work to real-world problems in the areas of personalized education, employment hiring (business), social well-being (health), and network design (transportation). The advantage of my proposed system compared to the existing works is that it: 1) incorporates domain knowledge with data-driven probabilistic models, 2) detects and describes complex discriminative patterns, 3) returns a fair decision/policy, and 4) breaks negative/positive feedback loops.

The cumulative solar and wind power capacity integrated mainly into low and medium voltage grids in Canada represents 9% of total available power capacity in 2015, and is expected to more than double by 2040. This reality will create not only opportunities for sustainable energy production, but also challenges for the system operator due to the uncertain power fluctuations from supporting multiple prosumers (customers who can alternatively behave as energy consumers or producers). This project addresses the question of how to involve prosumers in the energy management process for provision of ancillary services in the grid (e.g. voltage control) while mitigating unsuitable emerging effects. The idea is to consider a three-layer optimization problem related to different voltage levels (high, medium and low). Grid incentives will be optimized at the high voltage level, while lower levels will optimize the dispatch among grid prosumers to maximize their involvement. An Agent-Based Modelling framework will provide a backbone for this multi-level optimization, enable bi-directional information flows, and make it possible to handle the challenges of high data volume and complexity.

The cumulative solar and wind power capacity integrated mainly into low and medium voltage grids in Canada represents 9% of total available power capacity in 2015, and is expected to more than double by 2040. This reality will create not only opportunities for sustainable energy production, but also challenges for the system operator due to the uncertain power fluctuations from supporting multiple prosumers (customers who can alternatively behave as energy consumers or producers). This project addresses the question of how to involve prosumers in the energy management process for provision of ancillary services in the grid (e.g. voltage control) while mitigating unsuitable emerging effects. The idea is to consider a three-layer optimization problem related to different voltage levels (high, medium and low). Grid incentives will be optimized at the high voltage level, while lower levels will optimize the dispatch among grid prosumers to maximize their involvement. An Agent-Based Modelling framework will provide a backbone for this multi-level optimization, enable bi-directional information flows, and make it possible to handle the challenges of high data volume and complexity.

Learning driving behavior from smartphone location and motion sensors Monitoring and tracking vehicles and driving behavior are of great interest to better assess safety and understand the relationship with potential factors related to the infrastructure, vehicles and users. This has been implemented in recent years by car insurance to better assess their customers’ risk of crash and offer usage-based premium. Driver monitoring and analysis or driver behavior profiling is the process of automatically collecting driving data (e.g., location, speed, acceleration) and predicting the crash risk. These systems are mainly based on Global Positioning System (GPS), which suffers from accuracy issues, e.g. in urban canyons, and is insufficient to detect normal and risky driving events like steering and braking to assess the driving behaviours. To address this problem, researchers have proposed the integration of GPS, Inertial Navigation System (INS) and motion sensors, and map-matching (MM) in a single hybrid system. INS is fused with GPS and used during signal outages to provide continuous positioning (dead reckoning). Map matching is the process of estimating a user’s position on a road segment, which provides more contextual information like road geometry and conditions, historical risk of the segment and other drivers’ behaviour. The objective of this work is to improve the understanding of driver behaviour and crash risk by integrating location and motion data, driving events and road attributes using different machine learning algorithms.The objective of this work is to improve the understanding of driver behaviour and crash risk by integrating location and motion data, driving events and road attributes. The specific objectives are the following: 1) to detect risky driving events, namely hard acceleration/braking, compliance to signalization (e.g. speed limits), sharp steering, tailgating, improper passing and weaving from location and motion data using machine learning (ML); 2) to apply map-matching algorithms to extract road-related attributes; 3) to cluster driver behaviour based on the time series of location and motion data, detected driving events and road-related attributes.

We study the context of a transportation network manager who wants to take decisions on infrastructures, assets and resources to deploy in order to achieve its objectives. The network manager has to take into account that there are several classes of users, most of which pursue their own objectives within the rules stated by the manager, while others have objectives that are antagonist to those of the manager. Our goal is to develop methodology to help the transportation network manager. The application that motivates this research project is based on the transportation network design problem faced by a vehicle inspection agency who wants to inspect a maximum number of vehicles on a given territory under a limited budget. In such application, it is important to take into account the fact that some users will react to the installation of new vehicle inspection stations by diverting from their usual path to avoid inspection. Other applications of interest include the design of transportation networks that are resilient to major accidents and terrorist attacks. In this context, the network manager must anticipate potential threats posed by hostile users.

The railway industry represents one of the most important means of freight transportation. In Canada only more than 900,000 tons of goods are moved every day. where one of the major companies of the sector is Canadian National Railways (CN). An important component in their overall structure is the locomotive fleet management. The high cost of each locomotive and the large number of them required to satisfy train schedules makes the locomotive planning highly valuable. This in turn, represents an environmental and macroeconomic effect of great importance. Although several variants of locomotive planning problems have been studied before there is still a huge gap between the state-of-practice and the state-of-the-art. Thus, we will first study an optimization model that is tailored for CN’s requirements. Moreover, we will investigate on the development of specialized solution methods that incorporate machine learning with operations research techniques to obtain optimal solutions within reasonable time. This will provide a tool for the partner company to better evaluate scenarios in the locomotive planning and give value to the data while representing an important scientific contribution for the optimization community.

Protein homeostasis describes the cells capability to keep its proteins in their correct shape and function through a complex regulatory system that integrates protein synthesis, folding and degradation. How cells maintain protein homeostasis is a fundamental phenomenon, an understanding of which has direct implications for prevention and treatment of severe human diseases such as Alzheimer’s and Parkinson’s. The protein quality control is regulated through specific enzymes called molecular chaperones that assist the (re)folding of proteins thus managing a complex and varied proteome efficiently. Although the specificity of interactions of these chaperones with their client proteins is known to be the key to the efficient allocation of protein quality control capacity, a significant yet unanswered question lies in rationalizing the principles of this specificity. This project aims to systematically define the principles of sequence specificity across eukaryotic chaperone network through a combination of molecular modelling and machine learning methods. To this end, the peptide sequences that confer chaperone specificity will be identified systematically using a robust docking procedure accelerated by a Random Forest model. To account for the conditional interdependencies of the energetic contributions of the peptide residues binding to the chaperone receptor and to capture them, probabilistic graphical models will be developed and deep learning methods will be applied to the large dataset obtained from docking simulations. This project, through the unique and rich dataset we will construct and the sophisticated analyses we will apply, will not only unravel the sequence specificity in protein homeostasis interactions during health and disease, but also provide the necessary guidelines for how it can be re-engineered for rational therapeutic intervention.

Recurrent neural nets are neuroscience-inspired AI algorithms that are revolutionizing the machine learning of complex sequences. They help power a variety of widely used applications such as Google Translate and Apple’’s Siri. But they are also big, complicated models, and learning them is a delicate process, up to now requiring much fine-tuning to avoid the parameter adjustments getting out of control. The human brain also faces this stability problem when it learns sequences, but it has a robust, working solution that we are only beginning to understand. Bringing together experts in neuroscience, applied math, and artificial intelligence, we will adapt sophisticated methods for measuring stability from the mathematics of dynamical systems. We will develop learning algorithms that use this information to efficiently guide the learning, and will employ them in a neuroscience study that compares artificial and brain solutions to learning complex task sequences. Our goal is to unify and extend our understanding of how natural and artificial recurrent neural nets learn complex sequences.

Facility location arises as an important field in combinatorial optimization with applications to logistics and data mining. In facility location problems (FLPs), one seeks to find the location of some supply points and to assign customers to those supply points so as to optimize a certain measure of performance. In data mining, several FLPs can be used with the purpose of modelling and solving clustering problems. The p-center problem (PCP) is an example of such type of problem, in which one seeks to find the location of p points (namely the centers) so as to minimize the maximum dissimilarity between any customer and its closest center. This problem is extremely difficult in practice. In a recent article co-authored by the candidate, the most classical variant of the PCP (namely the vertex PCP) is solved by an iterative algorithm for problems containing up to a million data points within reasonable time limits. This is more than 200x larger than previous algorithms. In this project we aim at extending some of the ideas used in that article to solve other classes of facility location problems for large datasets.

Road traffic crashes are a serious concern. Typically, dangerous locations in the road network are identified based on historical crash data. However, using crashes is not ideal, as crash data bases contain error and omissions and crashes are not perfect predictors of safety. Our earlier work demonstrates how mobile sensor data, such as GPS travel data collected from regular drivers, can be used to substitute crash data in the safety management process within Quebec City. However, advanced statistical models must be developed to convert the collected sensor data into predicted crash counts at sites throughout the network. This project proposes three advancements to crash models developed in previous work. First, methods for imputing missing data will be proposed and explored. The effect of these methods on the final predicted crash counts will also be quantified. Second, techniques for expanding analysis to an entire road network will be developed. Third, the developed models will be tested on additional datasets in Montreal and Toronto. The ability to predict levels of safety with mobile sensor data is a substantial contribution to the field of transportation.

Data-mining sleep brain signals using machine-learning: effect of caffeine (EEG).

Développement d’approches en apprentissage par machine pour prédire la distribution intracellulaires des acides ribonucléiques (ARNs).

Optimisation de ressources hydroélectriques pour l’intégration des énergies renouvelables.

Adaptive cortical neuroprosthesis for neuromuscular control.

Modèle d’évaluation intégrée « BaHaMa » : développement d’une version multirégionale.

Données automatisées d’alimentation pour prédire les maladies des veaux laitiers.

Évaluation des mouvements du tronc chez des adolescents avec et sans scoliose idiopathique.

Studying fine-scale population structure using neural networks.

Classification de tumeurs du foie à l’aide d’un réseau neuronal convolutif.

Outil d’analyse méthodologique et contextuelle d’articles scientifiques en santé mentale.

Reclassification des systèmes de journalisation : une approche par apprentissage machine.

Optimisation de la calibration de mesures neurophysiologiques.

Recherche de modèle prédictif de décisions en justice.

Évaluation de la perte d’information engendrée par l’inspection visuelle ou automatisée des données EEG.

Modèle prédictif de réponse de métastases hépatiques de cancers colorectaux à la chimiothérapie avec techniques d’intelligence artificielle.

Learning normalized inputs for iterative estimation on medical image segmentation.

Éthique et science cognitive de l’attention manipulée par l’IA.

Discoverning the RNA structural determinants of the RNA-binding proteins.

Topographie cérébrale du rythme thêta dans la régulation émotionnelle: Étude pilote chez une population atteinte de schizophrénie.

Modélisation de la périodicité d’un signal brut provenant de données EEG.

Application de méthodes d’apprentissage machine pour l’amélioration des traitements contre les cancers pédiatriques.

Interactions ARN-protéines dans la maturation des microARN.

Cortical control of motor recovery: a dynamical systems perspective.

Lymphogranulomatose vénérienne : facteurs de risque et présentation clinique.

Machine Learning to Nudge Health Behaviours.

Mobilité et performance des communications des villes intelligentes.

Validation de l’automatisation du score automatisé de défaillance multiviscérale Pediatric Logistic Organ Dysfunction-2 (aPELOD-2).

Extensions de l’algorithme du gradient stochastique pour l’estimation de modèles mixed logit.

Accessing chromatin interactions by high-resolution analyses of correlated regulatory element variation.

Approches computationnelles pour investiguer les dérèglements épigénétiques héritables.

The occurrence of cyber risk incidents.

Rétablissement de dérèglements épigénétiques héritables dans les Cellules Embryonnaires par Édition de l’Épigénome.

Optimisation multivariée d’une neuroprothèse corticale pour le contrôle moteur.

Caractérisation de la famille des pharmacogènes CYP4F.

Projet en IoT et Villes.

Développement d’un outil d’analyse syntaxique pour structurer des données liées aux prescriptions électroniques.

Optimisation de l’utilisation de la flotte de locomotives du CN.

Détection de mutation de novo dans les gènes du Cytochrome p450.

Self-Supervised Learning of Visual Representations from Videos.

Réseaux d’interactions écologiques et changements climatiques : inférence et modélisation par des techniques d’apprentissage automatique

Domain Agnostic Adversarial Attacks for Security and Privacy.

Modélisation des interactions entre les modes de transport par l’intégration de différentes sources de données

Artificial Neural Networks and Bispectrum for Epileptic Seizure Prediction.

Evaluation of a Directed Acyclic Graph for Cysticercosis using Multiple Methods.

Spatial-Temporal Traffic Pattern Analysis and Urban Computation Applications based on Tensor Decomposition and Multi-scale Neural Networks.

Deep ecology: Bringing together theoretical ecology and deep learning.

Learning to Select Cutting Planes in Integer Programming.

Representing the World Through Predictions in Intelligent Machines.

Computer Vision for Safe Interactions between Humans and Intelligent Robots.

Stage chez Le Devoir, spécialité science des données.

Stage chez Le Devoir, spécialité science des données.

Stage chez Le Devoir, spécialité communication.

Accélération de méthodes d’optimisation pour les problèmes volumineux par évaluation inexact

Nous nous intéressons aux problèmes d’optimisation continue, non convexe et sans contraintes dans lesquels l’évaluation des valeurs de l’objectif et de son gradient sont obtenues à l’issue d’un processus coûteux. Nous supposons que l’on peut obtenir à moindre coûts des approximations de l’objectif et de son gradient à un niveau de précision souhaité. Nous regarderons l’impact de ces hypothèses sur la convergence et la complexité de méthodes d’optimisation classiques ainsi que les économies pouvant être effectuées sur le temps de calcul et la consommation énergétique. Cette étude est motivée, entre autre, par les problèmes d’inversion sismique dont la taille peut avoisiner les centaines de millions de variables et dont la fonction et le gradient peuvent être approximés par la résolution d’un problème aux moindres carrés linéaires. L’économie de temps de calcul et d’énergie est un enjeu majeur de l’ère de l’intelligence artificielle et de l’exploration des données volumineuses et cette approche est nouvelle est prometteuse en terme de retombées économiques et environnementales.

Oberman Mathematical approaches to adversarial robustness and confidence in DNN

Deep convolutional neural networks are highly effective at image classification tasks, achieving higher accuracy than conventional machine learning methods but lacking the performance guarantees associated with these methods. Without additional performance guarantees, for example error bounds, they cannot be safely used in applications where errors can be costly. There is a consensus amongst researchers that greater interpretability and robustness are needed. Robustness can be to differences in the data set where the models can be deployed, or even robustness to adversarial samples: perturbations of the data designed deliberately by an adversary to lead to a misclassification.

In this project, we will study reliability in two contexts: (i) developing improved confidence in the prediction of the neural network, using modified losses to improve confidence measures (ii) modified losses which result in better robustness to adversarial examples. The overall goal of the project is to lead to more reliable deep learning models.

Modèles prédictifs de l’allocation des dons de la Croix-Rouge canadienne en réponse aux feux de forêt

Au Canada, les inondations et les feux de forêt sont les catastrophes naturelles qui provoquent le plus de dégâts. Les efforts de la Croix-Rouge canadienne (CRC) visant à atténuer les impacts des feux de forêt dépendent fortement de la capacité des organisations à planifier, à l’avance, les opérations logistiques de secours. Le premier objectif du projet est l’élaboration de modèles permettant de prédire l’allocation des dons en argent aux bénéficiaires, selon les caractéristiques socio-démographiques de la région et du bénéficiaire ainsi que selon les caractéristiques des feux (sévérité et type). Le second objectif sera de comprendre les facteurs qui ont un impact significatif sur les besoins de la CRC lors d’une réponse à un feu de forêt, afin de faciliter la planification des opérations logistiques et les appels de financement.

h detach Modifying the LSTM Gradient Towards Better Optimization

Recurrent neural networks are known for their notorious exploding and vanishing gradient problem (EVGP). This problem becomes more evident in tasks where the information needed to correctly solve them exist over long time scales, because EVGP prevents important gradient components from being back-propagated adequately over a large number of steps. We introduce a simple stochastic algorithm (\textit{h}-detach) that is specific to LSTM optimization and targeted towards addressing this problem. Specifically, we show that when the LSTM weights are large, the gradient components through the linear path (cell state) in the LSTM computational graph get suppressed. Based on the hypothesis that these components carry information about long term dependencies (which we show empirically), their suppression can prevent LSTMs from capturing them. Our algorithm\footnote{Our code is available at https://github.com/bhargav104/h-detach.} prevents gradients flowing through this path from getting suppressed, thus allowing the LSTM to capture such dependencies better. We show significant improvements over vanilla LSTM gradient based training in terms of convergence speed, robustness to seed and learning rate, and generalization using our modification of LSTM gradient on various benchmark datasets.

Extraction de connaissances en français basée sur une traduction des textes en anglais combinée à l’utilisation d’outils développés pour l’anglais

Plusieurs institutions gouvernementales rendent disponible sur leurs sites web un très grand volume de documents qui ne sont écrits que dans la langue officielle du pays. Or, de plus en plus ces institutions désirent transformer ces documents en une base de connaissances, déployée en un ensemble de données ouvertes intégrées au Web sémantique. C’est le cas notamment du ministère de la Culture et des Communications du Québec, qui met à la disposition du public un répertoire du patrimoine culturel du Québec, très riche en informations textuelles, mais qu’il est malheureusement difficile d’intégrer aux données des autres acteurs culturels du Québec, ou de lier à toutes les connaissances patrimoniales qui sont déjà présentes dans le réseau de données ouvertes Linked Open Data (LOD).

Plusieurs travaux ont déjà été proposés pour soutenir l’effort d’extraction de connaissances à partir de textes : des annotateurs sémantiques, qui identifient dans un document les entités qui y sont citées (personnes, organisations, etc.) et les lient à leur représentation dans une base de connaissances du LOD; des extracteurs de relations, capables d’extraire du texte des relations entre deux entités (par exemple, « X est l’auteur du roman Y »); des extracteurs d’événements et d’informations temporelles. Dans la très grande majorité des cas, ces outils ont été développés pour l’anglais, ou offrent de piètres performances lorsqu’appliqués au français.

Nous proposons donc d’explorer une approche qui consiste à produire, à partir d’un corpus de documents en français, une version équivalente traduite sur laquelle seront appliqués les outils déjà existants pour l’anglais (le service Syntaxnet de Google, par exemple). Cela implique qu’il faudra tenir compte des erreurs et inexactitudes qui résulteront de l’étape de traduction. Pour y arriver, des techniques de paraphrase et de simplification de texte seront explorées, l’hypothèse ici étant que des phrases simples sont plus faciles à traduire et que cette simplification n’aura pas d’impact majeur sur la résolution de la tâche si la sémantique est préservée lors de cette simplification. On notera aussi que certains aspects de la langue, comme l’anaphore, perturbent la traduction (le module de traduction aura du mal à choisir entre les pronoms « it » et « he » pour traduire le pronom « il »). Il faudra dans ces cas mesurer précisément leur impact et proposer des solutions de contournement.

En bref, le projet proposé permettra de déterminer dans quelle mesure les services de traduction actuellement disponibles préservent suffisamment le sens du texte pour pouvoir exploiter des outils développés pour une autre langue. L’hypothèse que nous désirons valider est que leurs lacunes peuvent être comblées par certains prétraitements du texte original, et que ces prétraitements peuvent être implémentée à faibles coûts (en temps et en ressources).

Use of Deep Learning Approaches in the Activity Prediction and Design of Therapeutic Molecules

The proposed research is to employ Deep Learning and Neural Networks, which are both fields of Machine Learning, to more accurately predict the effectiveness, or “activity”, of potential therapeutic molecules (potential drugs). We are primarily concerned with predicting a given molecule’s ability to inhibit the growth of primary patient cancer cells (cells taken directly from a patient). The Leucegene project at the Institut de Recherche en Immunologie et Cancérologie (IRIC) has tested the activity of a large number of compounds in inhibiting the growth of cancer cells from patients afflicted with acute myeloid leukemia. The proposed research will use this activity data, along with several other data sources, to build an algorithm that can better predict the effectiveness that a molecule will have in inhibiting cancer cell growth. This means that before a molecule is even synthesized in a chemistry lab, a good estimation of its effectiveness as a therapeutic compound can be made, almost instantly. The first approach is to use Neural Networks and “representation learning”, in which features of the molecules that are important to improving activity are identified automatically by the algorithm. This will be done by representing the molecules as graphs and networks. Another approach that will be taken is the use of “multi-task learning” in which the prediction accuracy of an algorithm can be improved if the same algorithm is trained for multiple tasks on multiple datasets. The « multiple tasks » that will be focused on are multiple, but related, drug targets that are essential to cancer cell growth. Moving beyond activity prediction alone, these machine learning architectures will be expanded to design new chemical structures for potential drug molecules, based on information that is learned from drug molecules with known activities. These approaches have the capacity to improve the predictions about whether molecules will make effective drugs, and to design new molecules that have even better effectiveness than known drugs. Research progress in this area will lower the cost, both in money and time, of the drug development process.

Attacking the Reality Gap in Robotic Reinforcement Learning

As Reinforcement Learning (RL) becomes an increasingly popular avenue of research, one area that stands to be revolutionized is robotics. However, one prominent downside of applying RL in robotics scenarios is the amount of experience today’s RL algorithms require to learn. Since these data-intensive policies cannot be learned on real robots due to time constraints, researchers turn to fast, approximate simulators. Trading off accuracy for speed can cause problems at test time, and policies that fail to transfer to the real world fall prey to the reality gap: the differences between training simulation and the real-world robot. Our project focuses on theoretically analyzing this issue, and provides practical algorithms to improve safety and robustness when transferring robotic policies out of simulation. We propose algorithms that use expert-collected robot data to learn a simulator, allowing for better modeling of the testing distribution and minimizing the reality gap upon transfer. In addition, we study the transfer problem using analysis tools from dynamical systems and continual learning research, looking for indicators in neural network dynamics and optimization that signal when the reality gap is likely to pose an issue. Lastly, we use the analysis to synthesize an algorithm which optimizes for the metrics that signal good, “transferable” policies, allowing safer and more robust sim-to-real transfer.

Leveraging Machine Learning and Magnetoencephalography for the Study of Normal and atypical states of Consciousness

Understanding the neural processes and network dynamics underlying conscious perception is a complex yet important challenge that lies at the intersection between cognitive brain imaging, mental health, and data science. Magnetoencephalography (MEG) is a brain imaging technique that has many qualities favorable to investigating conscious perception due to its high temporal resolution and high signal to noise ratio. However MEG analyses across space, time and frequency is challenging due to the extreme high-dimensionality of variables of interest, and susceptibility to overfitting. Furthermore, high-computational complexity limits the ease with which investigators might approach some cross-frequency coupling metrics believed to be important for conscious perception and integration, across the whole brain. To mitigate such challenges, researchers frequently rely on a variety of multivariate feature extraction and compression algorithms. However, these techniques still require substantial tuning, and are limited in their application to the kinds of high-order tensor structures encountered in MEG. New methods for the study of conscious perception with MEG are thus needed.

In this project, we will leverage very recent advances in computer science and machine learning that extend algorithms currently used in neuroimaging research, to extreme high-dimensional spaces. Taken together, the proposed research will apply state-of-the-art techniques in machine-learning and electrophysiological signal processing to overcome current obstacles in the study of the brain processes that mediate conscious perception. This work will constitute an important contribution to neuroimaging methodology, neuropharmacology, and psychiatry. Beyond expanding our understanding of healthy cognition, this research may ultimately provide novel paths to the study of psychiatric disorders that involve altered conscious perception, such as Schizophrenia.

Recherche de nouvelle physique au Grand collisionneur de hadrons (LHC) à l’aide de l’apprentissage profond

Le Grand collisionneur de hadrons (LHC) se situe au cœur de la recherche fondamentale en physique. Avec sa circonférence de 27 km, celui-ci constitue le plus grand et plus puissant accélérateur de particules au monde. Ceci en fait le meilleur outil afin d’étudier le domaine de l’infiniment petit. C’est d’ailleurs au LHC que le boson de Higgs fut découvert, menant à l’obtention du prix Nobel de physique en 2013.

Cependant, le modèle standard, référence qui dicte les lois régissant les particules et leurs interactions, possède plusieurs lacunes que les physiciens et physiciennes n’ont toujours pas réussi à combler. Plusieurs théories furent élaborées, mais aucune d’entre elles ne fut observée au LHC. Face à ce défi, la communauté de physique des particules devra utiliser une nouvelle approche.

Le groupe ATLAS de l’Université de Montréal s’est ainsi tourné vers l’intelligence artificielle. Le projet élaboré par cette collaboration, et l’objectif principal de cette recherche est de développer un algorithme d’apprentissage profond qui permettrait de détecter des anomalies dans les données. L’algorithme développé sera ensuite utilisé sur les données du détecteur ATLAS dans l’espoir de découvrir des signaux de nouvelle physique et d’améliorer notre compréhension de l’univers.

Modélisation de la probabilité d’acceptation d’une route dans un contexte de covoiturage

Le covoiturage touche aux algorithmes fréquemment étudiés de tournées de véhicule, de ramassage et de livraison avec fenêtres de temps et de transport à la demande dynamique. Toutefois, très peu d’études s’attardent au contexte où les conducteurs et les passagers peuvent rejeter une proposition de route. Alors que le rejet d’une route proposée est rare lorsque les conducteurs sont des professionnels, c’est plutôt la norme dans un contexte de covoiturage. La modélisation de la probabilité d’acceptation d’une route se pose alors comme un problème central dans le développement d’une application mobile de covoiturage de qualité.

Le modèle d’apprentissage automatique développé devra estimer, selon les caractéristiques de l’utilisateur (notamment s’il est conducteur ou passager) et les routes alternatives proposées, la probabilité d’acceptation d’une route. De prime abord, cette modélisation pose deux défis :

(1) La façon dont les acceptations et les refus sont collectés pose un problème de type logged bandit. À ce titre, plusieurs propositions peuvent être offertes en même temps et un utilisateur peut en accepter plusieurs. De plus, les offres peuvent être activement refusées, simplement ignorées ou acceptées. Puisque les offres sont affichées séquentiellement, celles qui apparaissent en premier ont plus de chance d’attirer l’attention de l’utilisateur. L’ordre des propositions a donc probablement une influence sur la probabilité d’acceptation.
(2) Le comportement des nouveaux utilisateurs, pour qui très peu d’information est disponible, devra être inféré à partir des clients similaires de longue date. Ceci est en soi un problème difficile.

Hardware Acceleration of Speech Recognition Algorithms

Speech recognition has become prevalent in our lives in recent years. Personal assistants, such as Amazon’s Alexa or Apple’s Siri are such examples. With the rise of deep learning, speech recognition algorithms gained a lot of precision. This is due, mostly, to the use of neural networks. These complex algorithms, used in the context of a classification task, can distinguish between different characters, phonemes or words. However, they require lots of computations, limiting their use in power-constrained devices, such as smartphones. In my research, I will attempt to find hardware-friendly implementations of these networks. Deep learning algorithms are usually written in high-level languages using frameworks such as Torch or Tensorflow. To generate hardware-friendly representations, models will be adapted, using these frameworks. For example, recent findings have shown that basic networks can use weights and activations represented over 1 or 2 bits and retain their accuracy. The reduction of the precision of the network parameters is called quantization. This concept will be one of the many ways used to simplify the networks. Another aspect of this research will be to revisit methods of representing voice features. Traditionally, spoken utterances were converted to Mel Frequency Cepstrum Coefficients (MFCCs) which are essentially values representing signal power over a frequency axis. These coefficients are calculated roughly every 10 ms and are then sent to the model network. A representation of lower precision can greatly reduce the computational costs of the network. The overall goal of the research will be to improve the calculation speed and to diminish the power consumption of speech recognition algorithms.

Dépistage et diagnostic des troubles de la parole et du langage chez les enfants utilisant l’IA 

Le but de ce projet est de créer une plateforme Web où les cliniciens seront guidés dans la différenciation des troubles de la parole et du langage chez les enfants en posant simplement quelques questions de fond et en analysant les échantillons de parole des enfants pour y déceler les erreurs de parole et de langage. La solution proposée utilisera une reconnaissance de la parole qui est suffisamment précise dans les  » troubles de la parole  » qui sera développée au cours de ce projet et ensuite un algorithme sera déployé pour l’analyse détaillée des erreurs. Une partie de ce projet était le projet postdoctoral du boursier. Les objectifs à long terme de ce projet sont de rendre cet outil accessible aux parents de ces enfants également.

Développement d’outils basés sur l’apprentissage automatique pour soutenir la modulation thérapeutique de l’épissage alternatif

Les gènes humains sont constitués de séquences codantes pour des protéines (les exons) interrompues par des séquences non-codantes (les introns). Les séquences codantes sont jointes selon différentes combinaisons lors d’une étape importante de l’épissage alternatif (EA). Lors de maladies, l’EA est souvent dérégulé mais l’impact fonctionnel d’une telle dérégulation et les stratégies pour la corriger restent largement inconnus en raison de sa complexité et du manque d’outil informatique pour analyser de grands ensembles de données d’EA. Nous proposons de développer un ensemble d’outils informatiques utilisant des algorithmes d’apprentissage automatique qui nous permettront de comprendre l’impact d’une dérégulation de l’EA dans la pathogenèse et ainsi d’aider aux développements de nouvelles stratégies thérapeutiques ciblant les mécanismes de l’EA.

Temporal abstraction in multi-agent environment

Temporal abstraction refers to the ability of an intelligent agent to reason, act and plan at multiple time scales. The question of how to obtain and reason with temporally abstract representations has been extensively studied in classical planning and control theory, and more recently it has become an important topic in reinforcement learning, especially through the framework of options. The theoretical development of options is based on the framework of Semi-Markov Decision Processes (SMDPs), in which an agent interacts with its environment by observing states and taking actions. As a result of an action, the agent receives an immediate reward, and transitions to a new state drawn from some distribution, after a certain period of time which is also drawn stochastically. Both the state and the dwell-time distribution are dependent only on the agent’s state and action. However, in many cases of practical importance, an agent may be faced either with more general environments, in which the environment may be partially observable, or there may be multiple agents acting in the environment. For example, in energy markets or in transportation there may be many agents, who would be interacting with each other and making decisions without being able to observe relevant information except at specific time points.We propose to focus on establishing a mathematical framework for temporal abstraction which would work in Decentralized Partially Observable Markov Decision Processes. In a multi-agent system, agents take decision and exchange their information at designated decision epochs. In general, the decision epochs are given by the realizations of a random sequence. Instead of looking at every instant of time, when an action is taken by an agent, we are interested in the Decentralized Semi-Markov Decision Processes (Dec-SMDPs), in which a Partially Observable Markov Decision Process (POMDP) corresponding to an agent is embedded between any two successive decision epochs. In between two such decision epochs, each agent chooses actions so as to maximize the total return over a finite or infinite horizon, i.e., it solves a POMDP problem. The optimal decision epochs are chosen based on a given criterion, e.g., exchanging information at some goal states fixed a priori or when the increase of reward from the last decision epoch is less than a threshold. The overall performance, which is to be maximized through such sequential decision making consists of two rewards. The exchange of information in encouraged by an extrinsic reward along with an intrinsic reward that is maximized in between two consecutive decision epochs.We would like to investigate two aspects of this problem setup. First, we are interested in formally establishing the framework for Partially Observable Semi-Markov Decision Processes and its extension to decentralized (multi-agent) problems. We would like to investigate if under certain simplifying assumptions in the planning problem, the posterior beliefs (i.e., belief on the state of the environment based on past information and current action) exhibit certain monotonicity and symmetry properties so that we can infer what the structure of optimal policies could be. We also want to establish the general Dec-SMDP framework for modeling this problem and characterize its properties in comparison with SMDPs.In the subsequent analysis, we would like to investigate learning algorithms for these families of problems. We will build on standard reinforcement learning algorithms for temporal abstraction, such as option-critic, and provide extensions in our case that are consistent with the theoretical characterization of these problems. We will also examine the performance of both value-function-based and policy-gradient style algorithms in this context. We will compare the results that can be obtained using our framework to results in which each agent ignores the others and only tries to optimize myopically its own reward. We will use both standard simulated small problems from the multi-agent literature, designed to emphasize specific aspects, as well as larger scale domains that correspond to simulate transportation and energy markets, where multiple agents work in a cooperative setting to achieve a common task in a decentralized manner, e.g., self-driving cars and smart-grids. In such applications the agents occasionally communicate among themselves and use a common information to update a belief about the state of the world and a local information to decide about their individual policies and terminations of such policies.

Towards Learning Language Based Navigation in Visually Rich 3-D Environments

A long term goal of artificial intelligence and robotics is a robot able to perform manual tasks by understanding language instructions or questions and using visual and other sensory input to navigate and interact in a complex environment to achieve it’s goals. Advances in machine learning have succeeded in important perceptual sub-tasks of this problem: object recognition, speech recognition, natural language processing among others. However, how to integrate these successes with sequential decision making and multi-modal reasoning across language, vision, and other modalities is an open question that has been difficult to study. Very recently visually rich 3-D simulations and tasks have arisen aimed to allow the development of algorithms for learning language directed navigation of robots. Even in these constrained simulations, straightforward application of existing machine learning and reinforcement learning techniques are unable to effectively tackle this new set of challenges. We aim to develop methods for these problems focusing on visual relational reasoning and ideas from human learning. We also strive to advance the nascent evaluation methodology of these algorithms. Besides making steps towards our ambitions of creating intelligent agents, methods developed to solve these tasks can be directly applied in household automation, robotic assistants, manufacturing, and autonomous driving.

Visual Imitation Learning With Partial Information

For many control and decision-making tasks, it is complex to describe the desired behaviour we hope to elicit from a robot. Many complex tasks that we want robots to be able to perform are dependant on a skill that people acquire at a young age, imitation. The ability of animals to learn from demonstrations has triggered research across many disciplines. This work will push the possibilities of imitation learning by creating methods that will allow robots to learn from diverse video demonstration. Of particular interest are skills that involve interaction with objects in the real world. Imitation learning is a tough problem but is also a very important one. If we make enough progress on imitation learning average people could program robots by providing a few demonstrations of the desired task in the real world.

High Fidelity Data Collection for Precision Agriculture with Drone Swarms

Projections from United Nations show that by 2050 we will need to produce 70% more food. However, agriculture already takes over 38% of the land and it is the largest user of freshwater in the world. What can we do to improve the way food is produced? We propose high precision agriculture! It uses big data to support decisions increasing productivity and reducing the use of land, water, fertilizers, pesticides, herbicides and fungicides. The use of more intelligent methods is also beneficial to biodiversity, changing the way natural resources are managed from an one-size-fits-all approach to a tailor-made solution. Yet, traditional data sources are known to have limited resolution and even low altitude remote sensing (e.g. airplanes or unmanned aerial vehicles – UAVs) can only see from a fixed perspective: above. Additionally, according to PwC there’s a $32.4bn market for UAVs in the agriculture industry. This project proposes to improve productivity and sustainability by increasing the precision of the data collected down to the individual plant level with the use of Artificial Intelligence (Deep Convolutional Neural Networks) powered autonomous micro aerial vehicle swarms capable to fly among crops (e.g. corn, soybean and oats). With the high resolution data collected by a swarm of small and cost effective drones, farmers will be able to take advantage of all machine learning technology already available to optimize food production, maximize yield and minimize impact in the environment.

New Frontiers in Learning for Discrete Optimization

In addressing current and future societal needs, both the public and private sectors are deploying increasingly complex information and decision systems at unprecedented scales. The algorithms underlying such systems must evolve and improve rapidly to keep up with the pace. This project focuses on algorithms for Discrete Optimization, a widely used tool for decision-making and planning in industrial applications. The goal is to devise Machine Learning (ML) methods that streamline the process of algorithm design for discrete optimization, particularly in new, uncharted domains where classical paradigms may not be effective.

Deep Learning to understand the LHC data

What is our universe made of? To answer this question with the current technology, the Large Hadron Collider (LHC) has been under its stable operation since 2009, and has collected data of enormous number of proton-proton collisions, O(1e16/year). The protons are accelerated to nearly the speed of light. Each collision reproduces the high energy state that our universe once had right after the Big Bang. The research of fundamental particles under such high energy condition is very important to better understand the laws of our universe, which might tell us the future of our cosmos. Single collision at the LHC produces O(1000) particles, whose data are collected via millions of readout channels from the detector. Therefore, the data to be analyzed at the LHC amount to O(30PB/year), and become complex i.e. it would be a suitable target to apply and study machine learning (ML) techniques. However, many areas of the analyses have yet to be improved using advanced ML algorithms such as Deep Learning (DL). This project will accelerate the application of ML/DL to several aspects of LHC data analyses, with particular focuses on the particle identification, and the data quality evaluation, in order to support a potential discovery of new particles.

Angiographie Myocardique Ultrasonore Super résolue

Les maladies cardiovasculaires sont responsables de plus de 30% des décès dans le monde dont plus de 7 millions chaque année sont imputable aux maladies coronariennes. Chez les patients présentant des maladies coronariennes connues ou suspectées, l’imagerie est souvent la première étape du diagnostic. Malheureusement, aucune technique non-invasive ne permet aujourd’hui de cartographier l’anatomie et la fonction des vaisseaux intra myocardiques irriguant le cœur. Le développement d’échographes ultrarapides a récemment permis le développement d’une nouvelle méthode d’angiographie super-résolue, basée sur la détection de microbulles injectées, permettant de cartographier des vaisseaux sanguins à l’échelle capillaire (<10 µm). Cette technique ne peut cependant pas être directement être appliquée au cœur puisqu’elle nécessite encore aujourd’hui plusieurs minutes d’acquisitions et est très sensible au mouvement. Notre objectif principal est la mise au point d’un système ultrasonore de cartographie super résolue de la micro vascularisation intra myocardique en 3D par apprentissage machine destinée au diagnostic précoce des maladies coronariennes. L’utilisation de réseaux de neurones récurrents devrait permettre de prédire la structure et les paramètres du réseau vasculaire et ainsi améliorer le pronostic des patients tout en minimisant la complexité des examens.

Theoretical Understanding of Deep Neural Networks

Deep neural networks have led to state-of-the-art results in a wide range of applications including object detection, speech recognition, machine translation and reinforcement learning. However, the optimization techniques for training such models are not well-understood theoretically. Furthermore, it is unclear how the optimization procedure affects the ability of these models to generalize to new data. In this project, we propose to design scalable theoretically-sound optimization algorithms exploiting the underlying structure of deep networks. We also plan to investigate the interplay between optimization and generalization for these models. We hope that this project will result in improved methods for training deep neural networks.

Apprentissage profond pour les propriétés électroniques des matériaux quantiques

Certains matériaux ont des propriétés qui ne s’expliquent qu’avec les lois de la physique quantique: ce sont les matériaux quantiques. Il est souvent difficile de calculer les prédictions théoriques de leur propriétés, comme c’est le cas pour la supraconductivité à haute température critique, ou les phases topologiques de la matière. Cela ralentit la recherche sur ces matériaux, et donc le développement de nouvelles technologies. En collaboration avec l’Institut Quantique (IQ), à Sherbrooke, le présent projet est d’utiliser les méthodes de pointe en apprentissage profond pour améliorer nos outils de prédictions pour les matériaux quantiques. D’une part, nous cherchons à améliorer les avancée récentes basées sur l’apprentissage profond pour les calculs dits ab initio, qui permettent le calcul des propriétés électroniques et chimiques des molécules et cristaux à partir seulement de leur configuration atomique. D’autre part, nous cherchons à intégrer l’apprentissage profond aux méthodes de pointes pour électrons fortement corrélés, c’est-à-dire pour les matériaux où la configuration seule ne suffit pas, car les électrons interagissent fortement. Il s’agit de la toute première collaboration entre IVADO et l’IQ, qui permettra de développer une l’expertise de pointe en intelligence artificielle pour la modélisation de matériaux quantiques.

How do we know what we know: neuropsychology, neuroimaging, and machine learning unraveling the neuro-cognitive substrate of semantic knowledge.

Human intelligence has two key components: the ability to learn and that of storing a representation of what has been learned. A deeper understanding of how semantic representations are instantiated in biological neural networks (BNNs) will have a two-fold beneficial impact on society. First, it will improve clinical practice providing better diagnostic and prognostic tools for patients with impaired semantic processing. Second, it will inform the development of human-like representations in artificial neural networks (ANNs), leading towards general artificial intelligence. Through a multidisciplinary approach that includes experimental psychology, cognitive neuroimaging, and machine learning, we will shed light on how semantic representations (1) vary across individuals – both healthy volunteers and neurodegenerative patients, (2) are encoded in the brain – thanks to functional magnetic resonance imaging and magnetoencephalography, and (3) can generalize across tasks and stimuli modalities – enabling human adaptive behaviors. The extensive multimodal dataset we will acquire and analyze with state-of-the-art analytical tools will thus pave the way to groundbreaking scientific discoveries for both BNNs and ANNs.

Optimization Algorithms for Machine Learning and Deep learning

In this project, we are interested in the development of efficient algorithms for solving convex and non-convex optimization problems. Convex optimization lies at the heart of many classical machine learning tasks. In this project, one of our goals is the development of provably convergent algorithms for solving structured convex optimization problems. Interesting Directions: Define the weaker assumptions which guarantee convergence of optimization algorithms like Adam, Adagrad, SGD with momentum. What is the optimal mini-batch size of these algorithms? What is the optimal selection for learning rate and momentum parameter? What is the optimal sampling? Deep Neural Networks (DNNs) are the state-of-the art machine learning approach in many application areas. However, the optimization methods used for training such models are not well-understood theoretically. In this project we are interested in the design of novel optimization algorithms that exploring the underlying structure of DNNs. Interesting Directions: Can we theoretically provide an explanation of the heuristics (stagewise stepsize, batch normalization, etc) used in the training of DNNs? Is it possible to design methods that generalize well to new data by studying the loss landscape of DNNs? Can we design efficient distributed data-parallel algorithms with aim to accelerate the training of DNNs?

Using Generative Adversarial Networks to Visualize the Impacts of Climate Change

It is difficult to overstate the importance of fighting climate change. A recent report from the Intergovernmental Panel on Climate Change determined that dramatic and rapid changes to the global economy are required in order to avoid mounting climate-related risks for natural and human systems. However, public awareness and concern about climate change often does not match the magnitude of threat to humans and our environment. A primary reason for this is cognitive biases which tend to downplay the importance of effects we don’t see or experience personally. Therefore, making abstract predictions of climate change impacts understandable, relatable, and well-communicated is vital in helping to overcome the barriers to public awareness and action with regards to climate change. To contribute to overcoming these challenges, we propose to use a Generative Adversarial Network (GAN) to simulate imagery of the impact that climate-change induced flooding will have on buildings and houses in North America. Our GAN can then be hosted on the Web and used as a tool to help the public understand – both rationally and viscerally – the consequences of not taking sufficient action against climate change.

User Behavior Modeling for Intelligent Information Systems

Intelligent information systems such as search engines, recommender systems, digital assistants, and social chatbots are ubiquitous today. Machine learning algorithms are the core components of these systems. Therefore, the development of more sophisticated machine learning models for the next generation of intelligent information systems relies on the amount and quality of the training data. As a by-product of operating these systems, we can log a large amount of user interaction data and use it to train and optimize the machine learning models. For example, users’ clicks can be used as implicit relevance feedback to optimize Web search engines. However, optimizing the information system with observed user behavior logs is a non-trivial task as they only provide implicit and noisy signals and depend strongly on context. This project addresses this problem by first building reliable and generalizable user behavior models from the observed user behavior log and then utilizing them to optimize the intelligent information systems. This project will advance the research and development of intelligent information systems by solving the bottleneck of data availability.

Large scale optimization solvers in Julia for data science

L’étude d’algorithmes pour résoudre les problèmes d’optimisation est devenue au fil des années la base de la science des données et par extension ses multiples applications dans des secteurs clés tels que la santé, le transport, l’énergie, la finance … De nos jours, de nouveaux défis impliquent une quantité toujours grandissante de données à traiter ainsi qu’un accroissement de la difficulté des modèles utilisés. Ce projet a pour but de nouvelles avancées dans des outils numériques en Julia pour résoudre des problèmes d’optimisation complexe de très grande taille. Dans cette étude, nous considérons deux exemples que sont les problèmes d’optimisation sous contraintes d’équation aux dérivées partielles et les problèmes d’optimisations avec contraintes dégénérées qui surviennent en particulier dans l’étude de la théorie des jeux.

Apprentissage automatique pour des systèmes MIMO massifs à faible consommation d’énergie

L’année 2020 marquera le début du déploiement à grande échelle de la 5^e génération de réseaux cellulaires. Cependant, la question de l’impact énergétique de cette nouvelle génération de réseaux se pose. Actuellement, plus de 70% des coûts énergétiques des opérateurs proviennent des infrastructures radio. Ce bilan énergétique va s’alourdir avec l’introduction des communications en bandes millimétriques. Grâce à l’utilisation d’un grand nombre d’antennes du côté de la station de base, il est théoriquement possible d’augmenter l’efficacité énergique du système. Cependant,  les chaînes de transmission doivent être dupliquées, consommant énormément d’énergie en pratique. De nombreuses solutions sont proposées dans la littérature, mais celles-ci demandent une puissance calculatoire élevé, sans certitudes d’obtenir les performances optimales. Partant de ce constat, l’objectif de ce projet est d’étudier et de proposer de nouvelles solutions économes en énergie, performantes et implantables pour mettre en œuvre les techniques de réduction d’énergie pour des systèmes multi-antennes. La technologie des réseaux de neurone profonds est prometteuse pour résoudre de tels problèmes complexes. Une de leurs grandes forces réside dans leur capacité à apprendre les spécificités de l’environnement réel de fonctionnement. De plus, leur utilisation en télécommunications est facilitée par la possibilité d’aisément générer de vastes bases d’apprentissage.

Decoding auditory perception in cochlear implants users with machine learning

Predicting outcomes and personalizing care have long been significant challenges in health research. One area where little progress has been made concerns Cochlear Implants (CI), which can restore hearing in the deaf. However, clinical outcomes (speech and emotion perception) vary greatly across implantees in the absence of a clear picture as to why this is the case. Due to progress in machine learning, it is believed that outcomes could be improved by identifying the specific neuro-functional markers of CI use. To address this issue, we aim to identify the neural mechanisms underlying impaired auditory processes in CI users, with an initial focus on emotion perception deficits. For this, machine learning will be used to integrate neuroimaging and acoustical data from empirical experiments into predictive models. An extensive EEG data set of CI and normal hearing participants’ brain responses elicited by emotional sounds will be analyzed. For each group, we will identify: (1) the pattern of brain responses that can discriminate emotions and (2) the specific acoustic features (e.g., tempo, pitch) used for emotion perception.The identified neuro-markers will serve as a proof of concept, toward the broader implementation of machine learning to improve the quality of life of CI users.

Apprentissage automatique pour la modélisation d’événements extrêmes

En contexte de réchauffement climatique, les administrations publiques devront assurer le maintien de la sécurité publique et contenir les impacts socio-économiques et environnementaux qu’engendrent les événements naturels extrêmes. La complexité de ces événements de même que les risques imminents qu’ils apportent à la population nécessitent le développement de modèles de plus en plus complexes afin d’assurer une modélisation adéquate de ces phénomènes. L’utilisation d’approches plus avancées tels que les algorithmes d’apprentissage automatique permettrait de répondre à cette problématique en augmentant, d’une part, la précision des estimations mais également en répondant à la complexité du problème qui devient élevée avec la dimensionnalité des variables à l’étude et la dépendance spatio-temporelle des données. Une modélisation prédictive est d’autant plus cruciale sachant que ces phénomènes augmentent en fréquence, en intensité et en durée en raison du réchauffement global de la planète. L’objectif du projet est d’utiliser des algorithmes d’apprentissage automatique afin d’estimer les probabilités d’occurrence d’événements extrêmes. À terme, des outils de gestion basés sur l’apprentissage automatique seront développés et testés à des fins d’implantation. Les principaux bénéfices incluent une modélisation améliorée des événements extrêmes pour une meilleure gestion des risques (sur les plans économique, social et environnemental) liés à ces événements.

Deep Spatiotemporal Modeling for Urban Traffic Data

Large volumes of spatiotemporal data are increasingly collected and studied in modern transportation systems. Spatiotemporal models for traffic data are critical components of a wide range of intelligent transportation systems (ITS), such as ride sharing, transit service scheduling, signal control, and disruption management. The spatiotemporal data exhibit complex attributes, which introduce numerous challenges needs to be dealt with. Despite the abundance of spatiotemporal modeling techniques developed in different domains, it is still an open issue of making full use of the characteristics of the spatiotemporal datasets. The goal of this postdoc project is to develop new spatiotemporal models for urban traffic data based on deep learning and tensor learning. The specific objectives of this project are to: (1) characterize the spatiotemporal propagation properties of traffic data by deep spatiotemporal neural networks; (2) decouple interaction between external factors and traffic pattern by disentangle representation; (3) capture the strong regularity in collective travel behavior by low-rank tensor factorization and (4) utilize the cross-variable relationship by deep factors models. We will apply our models to large-scale and multivariate spatiotemporal data imputation and prediction. This project will lead to fundamental research advances to spatiotemporal modeling and urban intelligent transportation systems (ITS).

Générateur de données synthétiques pour améliorer l’apprentissage en science des données

Navigational Assistant for the Visually Impaired (NAVI)

Méthodes d’apprentissage automatique dans l’aide à l’élaboration de plans d’interventions en éducation

Automation of signal processing methods for feature construction on physiological signals

Traitement de Données Génomiques par Différentes Architectures de Réseaux de Neurones.

Reverse-engineering of and migration towards scalable NoSQL data architecture

Intégration et traitement temporel de la vision dans le cerveau lors de la reconnaissance d’attributs faciaux : modélisation de données MEG et comportementales en apprentissage machine.

Multi-Relational Link Prediction Using Graph Neural Networks (SEARL)

Learning algorithms for functional cortical neurostimulation

Les mécanismes neuronaux de la prise de décision

Probing learned network structure in a multi-task setting

Annotation des chromosomes sexuels de l’ornithorynque par apprentissage automatique

Projet de stage sur la modélisation du secteur énergétique du Mexique

Modèles prédictifs de l’allocation des dons de la Croix-Rouge canadienne en réponse aux feux de forêt

Automatic integration and deployment of AI models

Méthodes lagrangiennes pour la résolution de problèmes de conception de réseaux

Un agent d’apprentissage pour une neuroprothèse cortico-spinale Marina

Outil de planification en temps réel pour les mines souterraines

Automated communication analysis for Software-aided emergency management

Utilisation de réseau profond pour la simplification automatique de textes

Using Deep Learning to Identify Electrons at the Large Hadron Collider (LHC)

Investigation de dérégulations épigénétiques héritables par approches computationnelles

Investigation of systematic errors in genome assembly algorithms

Prediction of Fiber Quantity and Quality in Forest Supply Chains Using Artificial intelligence Methods

Studying social influence using a neuroimaging and data science approach

IA sécuritaire dans les essaims de drones : Développer une approche permettant aux petits essaims de drones de suivre les humains en toute sécurité

Simulation multiniveaux pour les applications des villes intelligentes

Création d’une interface Julia pour NOMAD pour l’ajustement automatique des hyper-paramètres d’algorithmes d’optimisation.

Évaluation de scores de risque polygénique selon le sexe et la structure populationnelle

Traitement de la dégénérescence par régularisation en optimisation continue

Analyses multivariées des interactions entre différents processus attentionnels (EEG): une approche orientée sur les données.

Étude de l’amas de galaxies supermassif MACSJ1447.7+0827

Évaluation empirique de méthodes de prévision de la demande

Visualisation et Analyse de données des réseaux des villes intelligentes.

Duckietown AI Driving Olympics.

Epileptic seizure detection by combining smart wear monitoring and artificial intelligence techniques.

Validation d’une batterie de tests visuels comme aide au diagnostic de troubles neurologiques.

Application heuristique des sciences des données à la théorie cinématique des mouvements humains.

Banque de données des issues cliniques à court et long termes des Échafaudages Vasculaires Biorésorbables comparativement aux Stents pharmacoactifs de 2^e génération

Performances des micro-PMUs dans les Villes Intelligentes

This project, which has already received approval from Health Canada, the Québec Ministry of Health and Social Services and the Montréal Heart Institute’s Ethics Committee, seeks to evaluate a colchicine-based treatment, including its impact on mortality rates and pulmonary complications. The scope of the task (recruitment of a cohort of 6,000 subjects) and the extremely tight deadline (as quickly as possible) for this type of project require the implementation of new digital recruitment and follow-up tools.

Funding amounts:
IVADO: $125,000
scale ai: $100,000
TransMedTech Institute: $50,000

Colcorona clinical trial

Polytechnique Montréal news release

In a prerequisite for drug development, this project seeks to identify molecules that may specifically associate with SARS-CoV-2. To do this, researchers from Mila and IRIC will first use neural networks to automatically generate billions of potential molecules. An enhancement algorithm will then be used to select the most promising ones for biological evaluation and possible clinical trials.

Funding amounts:
IVADO: $100,000
scale ai: $125,000
Canada Excellence Research Chair in Data Science for Real-Time Decision-Making: $25,000

Learn more

News release

Like other viruses, the virus responsible for COVID-19 mutates and changes over time. These mutations can lead to changes in its spread, in the demographic impact of the disease, or even in the effectiveness of certain treatments. In order to adapt to this reality, this project aims at a real-time genomic analysis of the virus through molecular modelling, focusing primarily on the variants observed in Quebec. The results will provide information for both public health and healthcare, as well as facilitate the work of researchers developing new treatments.

IVADO funding amount: $100,000

The current COVID-19 pandemic, like others before it, originated in a host animal. However, the ecology, origin and development of these hosts and their viruses remain largely unknown. In order to address this shortcoming, this project aims to model animal populations that act as reservoirs for these pathogens in order to complete knowledge of the disease and to anticipate future resurgences, or outbreaks of new viruses.

IVADO funding amount: $45,000

Methods of diagnosing COVID-19 require chemical reagents whose supplies are limited. One consequence is the potentially extensive spread of the virus by asymptomatic individuals. In order to more easily assess whether or not testing is needed, this project proposes to use Raman spectroscopy and artificial-intelligence algorithms to estimate an individual’s total viral load and, if necessary, then determine whether the coronavirus is present. Ultimately, this could make it possible to significantly reduce the number of tests to be carried out, a benefit for remote regions or regions with limited infrastructure.

Funding amounts:
TransMedTech Institute: $33,100
IVADO: $11,000

Using an algorithm based on machine-learning techniques, this project seeks to develop a protocol for better understanding the structural components involved in the vital functions of SARS-CoV-2 or any other RNA virus. This technique will make it possible to produce a list of therapeutic targets to counter their replication and proliferation, thus offering new perspectives for the development of drugs to be used in current or future clinical studies.

IVADO funding amount: $17,500

News release

Personalized tracking of COVID-19 cases allows for step-by-step monitoring of the spread of the disease and helps public health officials evaluate the effectiveness of the measures implemented. However, when the number of cases becomes too high, individual follow-up becomes impossible, making it very useful to track the virus at the genetic level. To this end, this project proposes a phylogenetic analysis of the virus, including the ability to link locally sampled cases to each other and to link them to cases in other countries. In this way, it is possible to estimate the virus’s movements and transmission speed. It will then be easier to determine its rate of introduction from outside the country and to assess the proportion of local or community transmission within populations.

IVADO funding amount: $15,000

Numerous COVID-19-related databases exist, but no virtual platform currently incorporates a significant proportion of these sources. This makes it difficult to do a global analysis of them, and to make connections between this often-medical information and external factors, especially socio-political ones. In this perspective, this international project aims to develop a multifactorial open-source platform, enabling the integration and continuous addition of new information.

IVADO funding amount: $10,000

Learn more

The goal of this project is to provide the scientific community with a platform to predict potential targets for a vaccine against COVID-19. This interactive platform uses an algorithm’s ability to predict which parts of the virus will be exposed on the surface of infected cells and thus generates a list of potential targets. This algorithm, developed by Tariq Daouda in the laboratories of Sébastien Lemieux and Claude Perreault, has already been used successfully, enabling the current situation to be approached from a different angle. Offered to researchers through a portal, it will make it possible to accelerate the development of vaccines against COVID-19, but also against other emerging viruses.

Find out more

News release

Many telemedicine tasks (such as responding to 811) involve healthcare professionals. This project proposes to set up several standalone telephone assistance solutions to free up these experts who are currently in high demand, whether to answer citizens’ routine questions, do follow-ups, make appointments or help navigate through healthcare facilities.

Funding amount: $500,000

Find out more

Chester is an existing prototype of a radiology assistant that can recognize certain pneumonia-related characteristics. During the current pandemic, this project aims to improve Chester’s disease predictions with the aim of enhancing the management of patient care. How will this be done? By combining artificial intelligence and image recognition, while widely disseminating a public database of clinical metadata for a large number of COVID-19 cases (as well as SARS and other pneumonia cases).

Find out more

Ressources

• Better understanding SARS-CoV-2 by reconstructing its history through genetics.

“This project aims to analyze the viral sequences at different stages of the evolution of the virus and thus identify indicators associated with the geographical regions where patients have tested positive for COVID-19.”

• Open platform for studying the immune system and developing treatments

“Data-efficient deep learning to better model immune response: (…) building an open-source platform leveraging the latest AI technologies to model pathways in the immune system in order to better predict immune response. (…) we work on AI approaches that can contribute to the process of vaccine design(…)”.

Find out more

• Creation of a biobank

“The main objective of this Québec-wide infrastructure is to provide researchers with the samples and data they need for their work. This will facilitate the co-ordination of research and support efforts for the development of new disease biomarkers, with a view to creating vaccines and drugs.”

Find out more

• Omics profiling of COVID-19 progression mechanisms and specific analysis of immune responses in young patients

“This project [will] provide a mechanistic understanding of SARS-CoV-2 virus progression to assess the risk of specific medical profiles and patients, as well as to help identify binding targets for potential antiviral agents and vaccines. (…) An example of an active research question is to understand the apparent resilience of young children to severe infection, which is somewhat atypical for such epidemics.”

Find out more

Creation of a new, free COVID-19 screening device with a self-administered questionnaire assessing healthcare workers’ symptoms prior to a client visit.

Find out more

Creation of a free HVAC (heating, ventilation and air conditioning) optimization service in response to COVID-19, using a “zone by zone” approach supported by cloud computing technologies.

Find out more

Establishment of an economic and industrial impact forecasting model that, with the help of private databases, evaluates each economic sector of a city, region, RCM or other, in order to describe the situation and its risk factors.

Find out more

EdLive makes its distance-learning technology available to schools and businesses. Thanks to this initiative and the collaboration of EdLive partners, several thousand students across Québec are taking courses online.

Find out more

L’assistant Watson pour les citoyens est maintenant disponible gratuitement pour aider les gouvernements et les institutions de santé à répondre aux questions courantes sur la COVID-19.

Apprenez-en plus

Mental health capsules online. Tips and strategies for better coping with this high-risk period for stress.

Find out more

Implementation of a free cybersecurity monitoring service to safeguard the security of companies’ and organizations’ IT equipment during this crisis.

Find out more

Launch of a COVID-19 rapid response call by Thales and its artificial intelligence (AI) research centre cortAIx.

Find out more

Free use of the Valital recruitment platform to more quickly find candidates or volunteers in the medical and research fields.

Find out more

Developing Biologically inspired Deep Neural Network for Continual Lifelong Learning

We humans are able to continually learn throughout our lifetime which is called lifelong learning. This capability is also crucial for computational systems interacting in the real world and processing continuous streams of data. However, the current deep learning systems struggle to continually acquire the incremental information available over time from non-stationary data distributions. They tend to forget the knowledge which is acquired earlier upon learning the new one – such a problem is called catastrophic forgetting. In this project, we will study biological factors of lifelong learning and their implications for the modelling of biologically motivated neural network architectures that can improve life-long learning capability of computational systems by reducing catastrophic forgetting problem.

Off Policy Batch Reinforcement Learning for Healthcare

Artificial Intelligence (AI) has an increasing impact on our everyday life, one being in health care. Today most of the successful applications of AI in healthcare are for diagnosis or prediction, but not for the treatment. But AI agents also have the potential for sequential decision making such as assisting doctors in reassessing treatment options, as well as in surgery. The branch of AI that is a natural fit for handling such sequential decision-making problems is known as Reinforcement Learning (RL).So far most of the successful applications of RL have been in the video game environments. But there are relatively fewer applications of RL in healthcare. One of the reasons is that unlike games, in healthcare the RL agents cannot interact with the environment to explore new possibilities to learn the optimal treatment policy. Trying new treatment options on patients without knowing their consequences is not only unethical but also can be fatal. So, the agent has to learn retrospectively from previously collected batches of data. In RL literature, this is called Off-Policy Learning. Challenges in off-policy evaluation, sparse reward, non-stationary data, and sample inefficiency are some of the roadblocks for using RL safely and successfully in healthcare. During my Ph.D. I aim to tackle some of these challenges in the context of healthcare.

Unsupervised representation learning

Unsupervised representation learning is concerned with using deep learning algorithms to extract ‘useful’ features (latent variables) from data without any external labels or supervision. This addresses one of the issues with supervised learning, which is the cost and lack of scalability in obtaining labeled data. The techniques developed in this field have broad applicability, especially with regard to training smart ‘AI agents’ and domains where obtaining labeled data is difficult.’Mixup’ (Zhang et al) is a recently-proposed class of data augmentation techniques which involve augmenting a training set with extra ‘virtual’ examples by constructing ‘mixes’ between random pairs of examples in the training set and optimizing some objective on those mixed examples. While the original mixup algorithm simply performed these mixes in input space (which comes with a few limitations) for supervised classification, recent work (Verma et al, Yaguchi et al) proposed performing these mixes in the latent space of the classifier instead, achieving superior results to the original work.One intuitive way to think about ‘latent space mixing’ is that we can imagine that the original data is generated by *many* latent variables, the possible configurations of which increase exponentially as the number of latent variables increases. Because of this we only see a *very small* subset of those configurations in our training set. Therefore, mixup can be seen as allowing the network to explore *novel* combinations of the latent variables it has inferred (which may not already be present in the training set), therefore making the network more robust to novel configurations of latent states (i.e. novel examples) at test time. Empirical results from the works cited corroborate this hypothesis.

The first stage of my PhD was exploring mixup in the context of unsupervised representation learning (building on the work of Verma et al, which I also co-authored), in which the goal is to learn useful latent variables from unlabeled data. This was done by leveraging ideas from adversarial learning and devising an algorithm which is able able to mix between encoded states of real inputs and decoding them into realistic-looking inputs indistinguishable from the real data. We showed promising results both qualitatively and quantitatively, and recently published our findings at the NeurIPS 2019 conference.

Some preliminary experiments suggest that one of our proposed variants of ‘unsupervised mixup’ has a connection to ‘disentangled learning’, which explores the inference of latent variables which are conceptually ‘atomic’ but can be arbitrarily composed together to produce more abstract concepts (which is similar to how we as humans structure information in the brain). This lays the groundwork for some more exciting research to pursue during my PhD.

City-Scale Traffic Data Imputation and Forecasting with Tensor Learning

With recent advances in sensing technologies, large-scale and multidimensional urban traffic data are collected on a continuous basis from both traditional fixed traffic sensing systems (e.g., loop detectors and video cameras) and emerging crowdsourcing/floating sensing systems (e.g., GPS trajectory from taxis/buses and Google Waze). These data sets have provided us with unprecedented opportunities for sensing and understanding urban traffic dynamics and developing efficient and reliable smart transportation solutions. For example, forecasting the demand and states (e.g., speed, volume) of urban traffic is essential to a wide range of intelligent transportation system (ITS) applications such as trip planning, travel time estimation, route planning, traffic signal control, to name just a few. However, there are two critical issues that undermine the use of these data sets in real-world applications: (1) the missing data and noisy nature make it difficult to get the true signal, and (2) it is computationally expensive to process large-scale data sets for online applications (e.g., traffic prediction). The goal of this project is to develop new framework to better model local consistencies in spatiotemporal traffic data, such as the {sensor dependencies} and {temporal dependencies} resulting from traffic flow dynamics. The scientific objectives are to: (1) Develop nonconvex low-rank matrix/tensor completion models considering spatiotemporal dependencies/correlations (e.g., graph Laplacian [spatial] and time series [temporal]) and traffic domain knowledge (e.g., fundamental diagram, traffic equilibrium, and network flow conservation). (2) Incorporate Gaussian process kernels and neural network structure

Inverse Reinforcement Learning with Robust Risk Preference

RL/IRL methods provide powerful tools for solving a wide class of sequential decision-making problems under uncertainty. However, the practical use of these techniques as a computational tool has been limited historically owing to multiple factors like the presence of high-dimensional continuous state and action spaces in many real-world decision problems, the stochastic and noisy nature of the real world systems compared to the simulated environments, and the indifference of traditional reward and utility functions to the risk preference of the agent. I am excited about the possibility of directing my future research towards building risk-aware MDP models as they would provide stronger reliability guarantees than their risk-neutral counterparts. one typical modeling premise in RL/IRL is to optimize the expected utility (i.e., an assumption that humans are risk-neutral), which deviates from actual human behaviors under ambiguity. Recent work suggests such an effort can provide stable solutions for high-dimensional state space problems, thus making them more applicable for practical use cases.As an effort in this direction, under the guidance of Prof. Erick Delage, I am working towards developing risk-aware IRL/RL algorithms for portfolio selection problems. Applications that I am interested in include, but are not limited to, i) learning the agent’s risk profile using inverse learning methods and ii) Risk sensitive exploration in RL setting. Our work tries to formulate the inverse learning model from a distributionally robust optimization (DRO) point of view where the agent performs at least as well as the expert in terms of the risk-sensitive objective. We plan to achieve this by building an ambiguity set for the expert’s risk preference and train the agent to learn by taking a worst-case approach, thus shielding the agent from the ambiguity in the underlying risk distribution.

Mesure de la pulsatilité cérébrale et son impact sur la cognition chez la souris vasculairement compromise par imagerie ultrasonore

Les maladies cardiovasculaires peuvent être à l’origine d’un vieillissement cérébral accéléré. Les artères, telles l’aorte ou les carotides, sont riches en fibres élastiques, permettant d’adoucir les fluctuations de la pression sanguine (ou pulsatilité) lors du cycle cardiaque dans les vaisseaux cérébraux en aval. Avec l’âge et la maladie, les artères deviennent plus rigides, entraînant une augmentation de la pulsatilité en aval et menant à des altérations microvasculaires. Cartographier la pulsatilité dans l’ensemble du réseau vasculaire cérébral pourrait donc devenir un biomarqueur permettant de diagnostiquer les maladies neurodégénératives. Jusqu’à récemment, suivre l’évolution du pulse dans le réseau vasculaire cérébral n’était pas possible : la microscopie optique ne permet que la mesure des micro-vaisseaux à la surface du cerveau, tandis que l’IRM haut champ permet d’imager un cerveau entier mais n’a pas une résolution spatiotemporelle et sensibilité suffisantes pour mesurer de petits vaisseaux. Une nouvelle technique ultrasonore pourrait relever ce défi : la Microscopie par Localisation Ultrasonore (MLU). Basée sur la localisation et le suivi de microbulles injectées comme agents de contraste, elle permet de cartographier les vaisseaux avec une résolution de l’ordre de 5 µm dans l’ensemble du cerveau. Cependant, cette méthode nécessite de suivre les microbulles durant 10 minutes pour finalement n’obtenir qu’une unique image de la vascularisation cérébrale. Notre objectif est de parvenir à rendre cette méthode dynamique, en la synchronisant avec l’ECG et la respiration afin d’obtenir non pas une image unique mais un film d’au moins une pulsation cardiaque, afin d’observer les variations de vitesse du flux sanguin au cours du cycle, et d’en déduire la pulsatilité. Cette nouvelle méthode permettra de démontrer pour la première fois la variation de la pulsatilité dans le cerveau entier, d’établir un lien de corrélation entre l’augmentation de la pulsatilité et les pertes de cognition ainsi que les dommages cérébraux et d’établir la mesure de la pulsatilité comme biomarqueur pour suivre l’évolution de maladies cardiovasculaires et/ou neurodégénératives.

Apprentissage profond et classification des états cognitifs pour la modulation en temps réel des interactions humain-machine en milieu automatisé.

L’univers du travail est en train d’être profondément modifié. Des technologies comme la robotique et des applications de l’intelligence artificielle s’intègrent de plus en plus dans les tâches du travail. L’objectif de cette recherche est de prendre en compte l’humain dans un environnement composé de machines. Ces dernières ne sont pas capables de percevoir que l’employé, avec lequel elles collaborent, est fatigué, absent mentalement, ou tout simplement distrait. Un collègue, dans ce cas-ci, s’ajusterait ou le préviendrait afin qu’il reprenne ses esprits. La machine, quant à elle, poursuivrait son activité sans s’ajuster augmentant le risque d’accident ou d’erreur. Pour répondre à ce problème, ce projet porte sur le développement d’un système qui s’adapte à l’état cognitif de son utilisateur comme la fatigue, la charge mentale ou la fatigue.Les interfaces cerveau-machines utilisent des mesures neurophysiologiques de l’être humain pour surveiller, s’adapter, ou se faire contrôler. À l’intérieur, des algorithmes d’apprentissage machine permettent de classifier l’état cognitif à partir des données capturées en temps réel. En utilisant les signaux électriques dégagés par le cerveau et la dilatation de la pupille, il est possible de discriminer entre plusieurs états dans le temps la situation de l’opérateur.

Le prototype développé pourra ainsi donner l’ordre à d’autres machines de ralentir la cadence ou de prévenir lorsque l’employé avec qui elles collaborent semble fatigué ou peu vigilant. Ce prototype sera développé et évalué en laboratoire dans un environnement contrôler. Il s’agit d’une preuve de concept à l’entreprise. Les interfaces cerveau-machines sont aujourd’hui principalement utilisées en médecine pour des prothèses, système d’assistance à la parole ou fauteuil roulants. Les retombées sont majoritairement en sécurité au travail (transport, manufacture) et dans l’optimisation de l’interaction humain-machine (collaboration humain-machine).

Coded Neural Network

Les réseaux de neurones profonds connaissent un engouement généralisé en ce début de décennie. Or, les progrès dans ce domaine s’accompagnent d’une hausse des besoins en capacité de calcul devançant la loi de Moore. Dans ce contexte, on cherche à proposer un ensemble de méthodes permettant d’optimiser les besoins en énergie de réseaux de neurones profonds en prenant en compte les caractéristiques physiques (mémoire, processeur, etc) du système accueillant le réseau pour son usage final.

L’objectif est d’avoir une méthode suffisamment générale pour s’adapter aux tâches et réseaux variés que les concepteurs pourraient vouloir déployer dans leurs applications, et réduisant la charge énergétique selon un compromis capacité du réseau/énergie contrôlable.
Ces travaux permettraient d’une part de gagner en énergie sur les systèmes des utilisateurs (par exemple, téléphone intelligent), ce faisant favorisant les usages déconnectés. D’autre part, on cherche à toucher les utilisations en data-centers, si voraces en énergie.
Ce projet de recherche permettra à terme de tirer parti au mieux des ressources allouées à l’apprentissage automatique dans sa phase d’exploitation, pour s’assurer de son acceptabilité sociale d’une part et de sa viabilité technique et économique d’autre part.

Computational Investigations of Pragmatic Effects in Language

This thesis focuses on natural language processing (NLP), specifically computational pragmatics, using deep learning methods. While most NLP research today focuses on semantics (literal meaning of words and sentences), my research takes a different approach: I focus on pragmatics which deals with intended meaning of sentences, one that is context-dependent. Correctly performing pragmatic reasoning is at the core of many NLP tasks including information extraction, summarization, machine translation, sentiment/stance analysis. My goal is to develop computational models where pragmatics is a first-class citizen both in terms of natural language understanding and generation. I have already made strong progress toward this goal: I developed a neural model for definiteness prediction [COLING 2016] — the task of determining whether a noun phrase should be definite or indefinite — in contrast to prior work relying on heavily-engineered linguistic features. This has applications in summarization, machine translation and grammatical error correction. I also introduced the new task of presupposition triggering detection [ACL 2018 — best paper award] which focuses on detecting contexts where adverbs (e.g. “again”) trigger presuppositions (e.g.,“John came again” presupposes “he came before”). This work is important because it is a first step towards language technology systems capable of understanding and using presuppositions and because it constitutes an interesting testbed for pragmatic reasoning. Moving forward, I propose to examine the role of pragmatics, particularly presuppositions, in language understanding and generation. I will develop computational models and corpora that incorporate this understanding to improve: (1) summarization systems e.g. in a text rewriting step to learn how to appropriately allocate adverbs in generated sentences to make them more coherent and (2) reading comprehension systems where pragmatic effects are crucial for the proper understanding of texts and where systems can answer questions of pragmatic nature whose answers are not found explicitly in the text. By the end, the thesis would present the first study on presuppositional effects in language to enable pragmatically-empowered natural language understanding and generation systems

Enhancing the Drug Discovery Process: Bayesian Inference to evaluate Efficacy Characteristics of potential Drug Through Uncertainty

During the multi-phase drug-discovery process, many compounds are tested in various assays which generates a great deal of data from which Efficacy Metrics (EM) can be estimated. Compounds are selected with the aim of identifying at least one sufficiently potent and efficient to go into preclinical testing. This selection is based on the EM meeting a specific threshold or by comparison to other compounds.

Current analysis methods suggest point estimates of EM and hardly consider the inevitable noise present in experimental observations, thus failing to report the uncertainty on the EM and precluding its use during compound selection. We propose to extend our previously introduced statistical methods (EM inference and pairwise comparison) to the ranking of a panel of compounds and to combinatorial analysis (multiple compounds tested simultaneously). Given an EM threshold, we aim at identifying the compounds with the highest probability of meeting that criteria.

We use a hierarchical Bayesian model to infer EM from dose-response assays (single- and multi-doses), yielding empirical distributions for EM of interest rather than single point estimates. The assay’s uncertainty can thus be propagated to the EM inference and to compound selection. We are thus able to identify all compounds of an experimental dose-response dataset with at least 1% chance of being amongst the best for various given EM, and to characterize the effects of each compounds of a combinatorial assay.

This novel methodology is developed and applied to the identification of novel compounds able to inhibit cellular growth of leukemic cells.

Uncertainty in Operations Research, Causality and Out-of-Distribution Generalization

My research focuses on two main directions: widening the operations research toolbox using recent advances in deep learning and learning causal structures. Both aspects have the potential to be useful in various applications, for example the optimization of railway operations, gene expression studies as well as the understanding of different protein interactions in human cells.Together with Emma Frejinger and its team at the CN chair, we developed a methodology which allows to predict tactical solutions given only partial knowledge of the problem using deep neural networks. We demonstrated the efficiency of the approach on the problem of booking intermodal containers on double-stack trains. Moreover, we are currently working to apply machine learning techniques to standard operations research problems such as the knapsack and the travelling salesman problem in hope of gaining insight about classical algorithms to solve them.

More recently, I have been interested in the nature of causal reasoning and how machines could acquire it. Typical machine learning systems are good at finding statistical dependencies in data, but often lack the causal understanding which is necessary to predict the effect of an intervention (e.g. the effect of a drug on the human body). Together with my co-authors, we developed « Gradient-Based Neural DAG Learning  », a causal discovery algorithm which aims at going beyond simple statistical dependencies. We showed the algorithm was capable of finding known causal relationships between multiple proteins in human cells.

In the future, I will work to make machine learning more adaptive and able to reuse past knowledge in order to learn new patterns faster. This is something humans do all the time, but which is hard for current algorithms. I believe causality is part of the answer, but other frameworks like meta-learning, transfer learning and reinforcement learning are going to be necessary. Apart from bringing us closer to human-level intelligence, making progress in this direction would benefit many applications. For instance, if a machine learning system is used to predict tactical solutions to a railway optimization problem, the distribution of problems it faces might shift due to changes in trade legislation, hence rendering the predicted solutions far from optimal. We should aim to build systems which can adapt to a changing world quickly.

Bioéthique écosystémique et mégadonnées: santé, agriculture et écologie

Les problèmes actuels sont globaux, liant société, économie et environnement à la santé. L’antibiorésistance par exemple provient d’un mésusage d’antibiotiques en santé et en agriculture qui vient réduire l’efficacité de ceux-ci. Pour attaquer ce problème, de larges collaborations entre médecins, agriculteurs et écologistes deviennent nécessaires, mais demeurent limitées par bons nombres de défis techniques (ex. : partage de données) et éthiques (consentement, sécurité) apparaissant dès l’intégration les données et les connaissances pour intervenir de façon concertée. L’objectif de cette thèse est d’étudier ces enjeux affectant la circulation des données entre santé, agriculture et écologie afin de proposer un modèle de gouvernance des données maximisant l’accès et la protection des données pour appuyer la recherche, la surveillance et l’intervention tout en maintenant la confiance des fournisseurs de données.

Ce projet fondera son analyse éthique sur une cartographie des relations entre les intervenants clés pouvant supporter un réseau de partage de données entre la santé, l’agriculture et l’écologie. Quatre études de cas sont amorcées et permettent de décrire le processus de constitution de ce réseau aux niveaux interministériel, intersectoriel, interprofessionnel, interpersonnel (certification éthique obtenue). Le devis ethnographique réalisé en étroite collaboration avec ces 4 milieux d’accueil supportera l’écriture d’un cadre de gouvernance par théorisation ancrée. Il sera ensuite comparé aux initiatives internationales (Danemark, Angleterre, États-Unis). Cette thèse permettra d’appuyer la mise en œuvre de réseaux structurants de partage de données intersectorielles au niveau de la médecine vétérinaire au Québec et jettera les bases d’un cadre de gouvernance pour l’interconnexion des bases de données entre organisations et secteurs.

Connexions entre réseaux récurrents, automates pondérés et réseaux de tenseurs pour l’apprentissage avec données séquentielles

À plusieurs reprises dans l’histoire, des découvertes ont été faites parallèlement par plusieurs scientifiques. On n’a qu’à penser au calcul infinitésimal développé indépendamment par Newton sous l’influence de ses travaux sur les lois universelles du mouvement et Leibniz inspiré par le principe philosophique de l’infiniment petit. À l’intersection entre plusieurs disciplines, ce type de découvertes n’atteignent leur plein potentiel que grâce à la contribution des différentes expertises. Dans cet ordre d’idées, de nombreuses équivalences peuvent être tracées entre les formalismes développés en physique et en intelligence artificielle. En particulier, une méthode pilier de la formulation moderne employée en physique quantique, les réseaux de tenseurs, peut être reliée aux réseaux récurrents, l’une des principales familles de modèles adaptés aux données structurées en apprentissage profond. Ces derniers sont également connectés aux automates pondérés, qui sont des modèles au cœur des méthode formelles et de vérification en informatique théorique. L’exploration des liens entre ces trois méthodes (réseaux de tenseurs, réseaux récurrents et automates pondérés) permet de tirer profit des garanties théoriques offertes par les méthodes formelles, de l’expressivité et des nombreuses applications des réseaux récurrents, tout en faisant le pont avec les débouchés des réseaux de tenseurs dans les domaines des matériaux et de l’informatique quantiques. Le projet vise ainsi à créer des passerelles entre ces différentes disciplines et exploiter les progrès faits dans l’une au profit des autres.

Neuroprosthesis development to recover the gait after spinal cord injury in rats

Spinal Cord Injury (SCI) interrupts the communication between the brain and the spinal locomotor networks, causing leg paralysis. When SCI is incomplete (iSCI), some nerve fibers survive the lesion and patients with iSCI can eventually regain some motor abilities. The goal of this study is to assess in the rat model whether combined brain and spinal stimulation can lead to a superior locomotion recovery after spinal cord injury. Artificial Intelligence (AI) techniques will be employed to track the motor activity, drive the stimulation and optimize the strategy in real time. By refining the spatiotemporal stimulation parameters, the intelligent algorithm will help the rat’s brain to generate leg trajectory that features a better clearance of the ground during swing, stronger leg extension and higher posture during stance. We expect that optimized neuroprosthetic stimulation will result in locomotor patterns that are more similar to intact rats and will facilitate the recovery of voluntary control of locomotion. The results will provide a framework for the future development of efficient neuromodulation interfaces and prosthetic approaches for rehabilitation.

Quantitative susceptibility mapping framework for assessing cortical development in neonates after severe deoxygenation at birth

Hypoxic ischemic encephalopathy (HIE) is a newborn brain pathology that is common but, unfortunately, not well understood. HIE affects 1.5 per 1000 live births in developed countries and is the leading cause of death and devastating sequelae in terms of neonates cognitive, behavioural, and physical disabilities. The most effective clinical treatment, therapeutic hypothermia, improves the survival rate; however, the repercussions of HIE remain unclear for survivors. As of today, the understanding of altered cortical growth mechanisms after HIE is incomplete but promising non-invasive magnetic resonance imaging (MRI) technique, called quantitative susceptibility mapping (QSM), provide new brain biomarkers that can help understand how HIE affects the brain development. Yet, because cortical development of neonates is rapid and sophisticated, standard clinical neurological imaging tools, such as MRI templates, are not suited for neurodevelopmental analysis in neonates.

Therefore, we propose to implement new methods for solving the QSM reconstruction problem and improve the common MRI template by developing adaptive age-based longitudinal templates. We will adopt a data-driven strategy with deep learning in order to create a new framework for the pediatric and neurology communities.

Méthodes multi-précision pour l’optimisation et l’algèbre linéaire

Ce projet de recherche a pour but de développer des méthodes capables de basculer d’une précision machine à l’autre durant la résolution de problèmes d’optimisation de grande taille, et d’effectuer l’essentiel des opérations en basse précision où elles sont peu coûteuses et requièrent peu d’énergie.

Nos résultats préliminaires indiquent des économies énergétiques pouvant aller jusqu’à 90% sur certains problèmes.

Ces méthodes s’appliquent notamment à la biologie des systèmes, qui requiert des solutions en quadruple précision, et au machine learning, où la demi précision est de plus en plus populaire. Sur les plateformes spécialisées émergentes gérant nativement ces nouvelles précisions, comme les cartes graphiques Turing de Nvidia qui implémentent la demie-précision ou encore le processeur IBM Power9 qui implémente la quadruple précision, ces méthodes seront à même d’exploiter au maximum le bénéfice du travail en multi-précision.

À l’ère des données massives et de l’explosion de l’information, des algorithmes permettant des économies d’énergie significatives sur les plateformes adéquates sont un investissement pour l’avenir du Canada, en termes du volume de données exploité et de l’environnement.

Decomposition of information encoded in learned representations of recurrent neural networks

The human brain contains billions of neurons that communicate with each other through trillions of synapses, enabling us to learn new skills, solve complex tasks and understand intricate concepts. Everything we do such as walking, eating, communicating, and learning, is a function of these neurons firing in certain patterns, in specific locations. This sophisticated biological neural network is the outcome of millions of years of evolution. Recent advances in deep learning have proposed several artificial neural network architectures for solving complex learning tasks, by taking simplified inspiration from neural circuits in our brains. Examples of these include convolutional neural networks for image and audio processing, recurrent neural networks for sequence learning and autoencoders for dimensionality reduction. Both biological and artificial networks rely on efficient calibration of synapses (or connection weights) to match desired behaviours. This adjustment is how a network « learns », but is a complicated task that is not well understood. An important substrate of networks after learning is the internal low dimensional representation found in the joint activity of neural populations that emerge upon performing a learned task. The present research aims to explore and further investigate these internal representations, and address the question of how do structural properties of network connectivity impact the geometry, dimensionality and learning mechanisms encoded by these internal features. We plan to answer this question by leveraging multidisciplinary data exploration tools from graph signal processing, dimensionality reduction, representation learning and dynamical systems. We expect that this project will allow us to gain better understanding of how natural and artificial neural networks solve complicated tasks, which in turn will help us find methodological ways to improve existing structures, and build new models, but more from a deeper understanding rather than trial and error.

Statistical Arbitrage of Internationally Interlisted Stocks

In this project, we will investigate a novel form of statistical arbitrage that will combine artificially created financial instruments in a high-frequency world, meaning that we will operate in the millisecond timeframe. These instruments will be constructed in such a way that they will offer very interesting statistical properties that will enable us to exploit violations in the law of one price in the Canadian and American markets. This arbitraging activity is essential, since it is making them more efficient by eliminating mispricing in equities that are quoted on both markets. The novel strategy will be tested on a large basket of equities on three trading venues in North America and given that we are working in high-frequency, this means that millions of market observations are ingested and analyzed daily by our trading algorithms. In order to be proactive in the markets, to make extremely fast and accurate predictions, and because of the complex nature of financial data and its abundance, we will be relying on machine learning algorithms to guide our trading decisions.

A Novel Deep Learning Approach to High-Energy Astrophysics

Despite machine learnings recent rise to stardom in the applied sciences, the astronomy community has been reluctant to accept it. We propose to gently introduce several forms of machine learning to the community through the study of the hot gas pervasive in galaxy clusters. Currently, emission spectra from galaxy clusters are studied by fitting physical models to them and using those models to extract relavent physical parameters. Unforunately, there are several inherent pitfalls with this method. We plan to train different algorithms — from a random forest classifier to a convolutional neural network — to parse the necessary thermodynamic variables from the emission spectra. The fundamental goal of this project is to create and open-source pipeline and suite of tutorials which integrate machine learning into the study of galaxy clusters.

Learning solutions to the locomotive scheduling problem

Given a set of demands on a railway network, how should one assign locomotives to trains in order to minimize total costs and satisfy operational constraints? This question is critical for Canada’s largest railway company: Canadian National Railways. Given the size of their network, even a small relative gain in efficiency would produce significant savings. The goal of this research is to explore recent advances in machine learning in order to efficiently solve the locomotive assignment problem. The idea is to train a neural network on precomputed solutions of the problem with the aim of learning the correct configuration of locomotives for a given train. By combining both integer programming and deep learning, the computational time can be reduced by at least an order of magnitude compared to integer programming alone. This is a solution that is significantly more efficient and practical for train operators.

Cognition-Based Auto-Adaptive Website User Interface in Real Time

A personal message that is designed specifically for the need and taste of consumers has always been the goal of media outlets, retailers, and social activists. Here at Tech3Lab, we are launching this massive study of personalization in an unprecedented way: by analyzing neurophysiological and psychophysiological signals of the body to determine the best possible look and feel on websites to improve user experience and best convey the intended message.

Previously, auto-adaptive website personalization was carried out mostly by guesswork and theory, in which there is no real evidence for the parameters used. Thanks to the equipment in Tech3Lab, such as EEG, fNIRS, physiological measurement instruments, and eye tracking measures, we are able to base our adaptive system on direct signals from the body.

This interdisciplinary study of cognitive neuroscience, marketing, and data science has the potential to revolutionize the approach of designers, developers, and editors to website design by studying auto-adaptive websites using direct body measures.

Bridging the gap between structures and functions: learning interpretable graph neural representation of RNA secondary structures for functional characterization

Cells are the basic units of life and their activity is regulated by many delicate subcellular processes that are crucial to their survival. Therefore, it is important to gain more insights into the complex control mechanisms at play, both to obtain a better fundamental understanding of biology, and to help understand diseases caused by defects in these mechanisms. We are particularly interested in the regulatory roles played by RNA molecules in the post-transcriptional phase such as subcellular localization and RNA-protein interactions. RNA secondary structures, a representation of how RNA sequences fold onto themselves, can have a significant impact on the molecule’s regulatory functions through its interaction with various mediating agents such as proteins, RNAs and small molecules. Therefore, in order to fully exploit RNA secondary structures to better understanding of their functions, we propose a novel framework of an interpretable graph neural representation of RNAs, which may ultimately lead us to the design of RNA based therapeutics for disease such as neurodegenerative disorders and cancers, the success of which would crucially depend on our capability of understanding the relations between RNA structures and functions.

Combining machine learning and blackbox optimization for engineering design

The efficiency of machine learning (ML) techniques relies on many mathematical foundations, one of which being optimization and its algorithms. Some aspects of ML can be approached using the simplex method, dynamic programming, line-search, Newton or quasi-Newton descent techniques. But there are many ML problems that do not possess an exploitable structure necessary for the application of the above methods. The objective of the present proposal is to merge, import, specialize and develop blackbox optimization (BBO) techniques in the context of ML. BBO considers problems in which the analytical expressions of the objective function and/or of the constraints defining an optimization are unavailable. The most frequent situation is when these functions are computed through a time-consuming simulation. These functions are often nonsmooth, contaminated by numerical noise and can fail to produce an usable output. Research in BBO is in constant growth since the last 20 years, and has seen a variety of applications in many fields. The research projects will be bidirectional. We plan to use and develop BBO techniques to improve the performance of ML algorithms. Conversely, we plan to deploy ML strategies to improve the efficiency of BBO algorithms.

Physics-informed deep learning architecture to generalize medical imaging tasks

The field of AI has flourished in recent years; in particular deep learning has shown unprecedented performance for image analysis tasks, such as segmentation and labeling of anatomical and pathological features. Unfortunately, while dozens of deep learning papers applied to medical imaging get published every year, most methods are tested in single-center: in the rare case where the code is publicly available, the algorithm usually fails when applied to other centers, which is the “real-world” scenario. This happens because images from different centers have different features than the images used to train the algorithm (contrast, resolution, etc.). Another issue limiting the performance potential of deep learning in medical imaging is that little data and few manual labels are available, and the labels are themselves highly variable across experts. The main objective of this project is to push the generalization capabilities of medical imaging tasks by incorporating prior information from MRI physics and from the inter-rater variability into deep learning architectures. A secondary objective will be to disseminate the developed methods to research and hospital institutions via open-source software (www.ivadomed.org), in-situ training and workshops.

A neurodevelopmentally-informed computational model of flexible human learning and decision making

The adolescent period is characterized by significant neurodevelopmental changes which impact on reinforcement learning and the efficiency with which such learning occurs. Our team has modelled passive-avoidance learning using a bayesian reinforcement learning framework. Results indicated that parameters estimating individual differences in impulsivity, reward sensitivity, punishment sensitivity and working memory, best predicted human behaviour on the task. The model was also sensitive to year-to-year changes in performance (cognitive development), with individual components of the learning model showing different developmental growth patterns and relationships to health risk behaviours. This project aims to expand and validate this computer model of human cognition to: 1) Better measure neuropsychological age/delay; 2) understand how learning parameters contribute to human decision making processes on more complex learning tasks; 3) simulate better learning scenarios to inform development of targeted interventions that boost human learning and decision making; and 4) inform next generation artificial intelligence models of lifelong learning.

Novel AI driven neuroprosthetics to shape stroke recovery

Stroke is the leading cause of disability in occidental countries. After stroke, patients often have abnormally low activity in the part of the brain that controls movements, the motor cortex. However, the malfunctioning motor cortex receives connections from multiple spared brain regions. Our general hypothesis is that neuroprostheses interfacing with the brain can exploit these connections to help restore adequate motor cortex activation after stroke. In theory, brain connections can be targeted using new electrode technologies, but this problem is highly complex. It cannot be done by hand, one patient at a time. We need automated stimulation strategies to harness this potential for recovery. Our main objective is thus to develop an algorithm that efficiently finds the best residual connections to restore adequate excitation of the motor cortex after stroke. In animals, we will implant hundreds of electrodes in the diverse areas connected with the motor cortex. The algorithm will learn the pattern of stimulation that is the most effective to increase activity in the motor cortex. For the first time, machine learning will become a structural part of neuroprosthetic design. We will use these algorithms to create a new generation of neuroprostheses that act as rehabilitation catalyzers.

Paths to a cleaner Northeast energy system through approximate dynamic programming

Our main research question is the design of greener energy systems for the American Northeast (Canada and USA). Some of the sub questions are as follows. How can renewable energy penetrate the markets? Are supplementary power transmission lines necessary ? Can energy storage improve the intermittency problems of wind and solar power? Which greenhouse gases (GHG) reductions are achievable ? What is the cost of such changes ? Crucially, what is the path to a better system ? To support the transition to this new energy system, our proposition is : 1. to model the evolution of the Northeast power system as a Markov Decision process (MDP), including crucial uncertainties, e.g. on technological advances and renewable energy cost; 2. to solve this decision process with dynamic programming and reinforcement learning techniques; 3. to derive energy/environmental policy intelligence from our computational results. Our methodological approach relies on two building blocks, an inter-regional energy model and a set of algorithmic tools to solve the model as an MDP.

Learning the Dynamics of the Limit Order Book

Modern financial markets are increasingly complex. A particular topic of interest is how this complexity affects how easily investors can buy or sell securities at a fair price. Many have also raised concerns that algorithms trading at high frequency could create excess volatility and crash risk. The central objective of our research agenda is to better understand the fundamental forces at play in those markets where trading speed is now measured in nanoseconds. Our project seeks to lay the groundwork, using big data, visualization, and machine learning, to answer some of the most fundamental questions in the literature on market structure. Ultimately, we envision an environment in which we could learn the behavior of the various types of agents in a given market. Once such an environment is obtained, it would allow us to better understand, for instance, the main drivers of major market disruptions. More importantly, it could allow us to guide regulators in the design of new regulations, by testing them in a highly realistic simulation setup, thereby avoiding the unintended consequences associated with potential flaws in the proposed regulation.

Data-driven Demand Learning and Sharing Strategies for Two-Sided Online Marketplaces

The proliferation of two-sided online platforms managed by a provider is disrupting the global retail industry by enabling consumers (on one side) and sellers (on the other side) to interact in exponential ways. Evolving technologies such as artificial intelligence, big data analytics, distributed ledger technology, and machine learning are posing challenges and opportunities for the platform providers with regards to understanding the behaviors of the stakeholders – consumers, and third-party sellers. In this proposed research project, we will focus on two-sided platforms for which demand-price relationship is unknown upfront and has to be learned from accumulating purchase data, thus highlighting the importance of the information-sharing environment. In order to address this problem, we will focus on the following closely connected research objectives: 1.Identify the willingness-to-pay and purchase decisions (i.e., conversion rate) of online customers based on how they respond to the design of product listing pages, online price and promotion information posted on the page, shipping and handling prices, and stock availability information. 2.Determine how much of the consumer data is shared with the sellers and quantify the value of different information sharing configurations – given the sellers’ optimal pricing, inventory (product availability), and product assortment (variety) decisions within a setting.

Interpretability of Deep Learning Approaches Applied to Omics Datasets

The high-throughput generation of molecular data (omics data) nowadays permits researchers to glance deeply into the biological variation that exists among individuals. This variation underlies the differences in risks for human diseases, as well as efficacy in their treatment. This requires combining multiple biological levels (multi-omics) through flexible computational strategies, including machine learning (ML) approaches, becoming highly popular in biology and medicine, with a particular enthusiasm for deep neural networks (DNNs). While it appears like a natural way to analyze complex multi-omics datasets, the application of such techniques to biomedical datasets poses an important challenge: the black-box problem. Once a model is trained, it can be difficult to understand why it gives a particular response to a set of data inputs. In this project, our goal is to train and apply state-of-the-art ML models to extract accurate predictive signatures from multi-omics datasets while focusing on biological interpretability. This will contribute to building the trust of the medical community in the use of these algorithms and will lead to deeper insights into the biological mechanisms underlying disease risk, pathogenesis and response to therapy.

Modélisation des surverses occasionnées par les précipitations et développement d’outils adaptés aux besoins de la Ville de Montréal

La contamination fécale des eaux de surface constitue l’une des premières causes de maladies d’origine hydrique dans les pays industrialisés et dans les pays en voie de développement. En zone urbaine, la contamination fécale provient majoritairement des débordements des réseaux d’égouts combinés. Lors de précipitations, l’eau pluviale entre dans le réseau d’égouts et se mélange à l’eau sanitaire pour être acheminée vers la station d’épuration. Si l’intensité des précipitations dépasse la capacité de transport du réseau, le mélange des eaux pluviales et sanitaires est alors directement rejeté dans le milieu récepteur sans passer par la station d’épuration. Ces débordements constituent un risque environnemental et un enjeu de santé publique. À l’heure actuelle, les caractéristiques des événements pluvieux occasionnant des surverses sont incertaines. Ce projet de recherche vise à tirer profit des données sur les surverses récemment rendues publiques par la Ville de Montréal pour caractériser les événements de précipitations occasionnant des surverses sur son territoire. Cette caractérisation permettra, d’une part, d’estimer le nombre de surverses attendues pour le climat projeté des prochaines décennies. D’autre part, elle sera utilisée pour dimensionner les mesures de mitigation, tels que les bassins de rétention et les jardins de pluie.

Reactive/learning/self-adaptive metaheuristics for healthcare resource scheduling

The goal of this research proposal is to develop state-of-the-art decision support tools to address the fundamental challenges of accessible and quality health services. The challenges to meeting this mandate are real, and efficient resource management is a key factor in achieving this goal. This proposal will specifically focus on applications related to patient flow. Analysis of the literature shows that most research focuses on single-resource scheduling and considers that demand is known; Patient and resource scheduling problems are often solved sequentially and independently. The research goal is to develop efficient metaheuristic algorithms to solve integrated patient and resource scheduling problems under uncertainty (e.g., demand, prole, and availability of resources). This research will be divided into three main themes, each of them investigating a different avenue to more efficient metaheuristics: A) learning approaches to better explore the search space; B) blackbox optimization for parameter tuning; and C) simulation-inspired approaches to control the noise induced by uncertainty.

Deciphering RNA regulatory codes and their disease-associated alterations using machine learning

The human DNA genome serves as an instruction guide to allow the formation of all the cells and organs that make up our body over the course of our lives. Much of this genome is transcribed into RNA, termed the ‘transcriptome’, that serves as a key conveyor of genetic information and provides the template for the synthesis of proteins. The transcriptome is itself subject to many regulatory steps for which the basic rules are still poorly understood. Importantly, when these steps are improperly executed, this can lead to disease. This project aims to utilize machine learning approaches to decipher the complex regulatory code that controls the human transcriptome and to predict how these processes may go awry in different disease settings.

Next generation neonatal brain segmentation built on HyperDense-Net, a fully automated real-world tool

There is growing recognition that major breakthroughs in healthcare will result from the combination of databanks and artificial intelligence (AI) tools. This would be very helpful in the study of the neonatal brain and its alterations. For instance, the neonatal brain is extremely vulnerable to the biological consequences of prematurity or birth asphyxia, resulting in cognitive, motor, language and behavioural disorders. A key difference with adults is that key aspects of brain-related functions can only be tested several years later, hindering greatly the advancement of neonatal neuroprotection. Researchers and clinicians need objective tools to immediately assess the effectiveness of a therapy that is given to protect the brain without waiting five years to see if it succeeded. Neonatal brain magnetic resonance imaging can bridge this gap. However, it represents a real challenge as this period of life represents a unique period of intense brain growth (e.g. myelination and gyrification) and brain maturation. Thus, we plan to improve our existing neonatal brain segmentation tools (i.e. HyperDense-Net) using the latest iterations of AI tools. We will also develop a validated tool to determine objective brain maturation in newborns.

Robustness and generalization guarantees for Deep Neural Networks in security and safety critical applications

Despite impressive human-like performance on many tasks, deep neural networks are surprisingly brittle in scenarios outside their previous experience, often failing when new experiences do not closely match their previous experiences. This ‘failure to generalize’ is a major hurdle impeding the adoption of an otherwise powerful tool in security- and safety-critical applications, such as medical image classification. The issue is in part due to a lack of our theoretical understanding of why neural networks work so well. They are powerful tools but less interpretable than traditional machine learning methods which have performance guarantees but do not work as well in practice. This research program will aim to address this ‘failure to generalize’, by developing guarantees of generalization, using notions of the complexity of a regularized model, corresponding to model averaging. This approach will be tested in computer vision applications, and will have near-term applications to medical health research, through medical image classification and segmentation. More broadly, the data science methods developed under this project will be applicable to a wide variety of fields and applications, notably wherever reliability and safety are paramount.

Differentiable perception, graphics, and optimization for weakly supervised 3D perception

An ability to perceive and understand the world is a prerequisite for almost any embodied agent to achieve almost any task in the world. Typically, world representations are hand-constructed because it is difficult to learn them directly from sensor signals. In this work, we propose to build the components so that this map-building procedure is differentiable. Specifically, we will focus on the perception (grad-SLAM) and the optimization (meta-LS) components. This will allow us to backpropagate error signals from the 3D world back to the sensor inputs. This enables us to do many things, such as regularize sensor data with 3D geometry. Finally, by also building a differentiable rendering component (grad-Sim), we can leverage self-supervision through cycle consistency to learn representations with no or sparse hand-annotated labels. Combining all of these components together gives us the first method of world representation building that is completely differentiable and self-supervised.

Investigating Combinations of Neural Networks and Constraint Programming for Structured Prediction

L’intelligence artificielle occupe une place de plus en plus importante dans de nombreuses sphères d’activité et dans notre quotidien. En particulier, les réseaux de neurones arrivent maintenant à assimiler puis à accomplir des tâches auparavant réservées aux humains. Cependant lorsqu’une tâche nécessite le respect de règles structurantes complexes, un réseau de neurones éprouve parfois beaucoup de mal à apprendre ces règles. Or un autre domaine de l’intelligence artificielle, la programmation par contraintes, a précisément été conçue pour trouver des solutions respectant de telles règles. Le but de ce projet est donc d’étudier des combinaisons de ces deux approches à l’intelligence artificielle afin de plus facilement apprendre à accomplir des tâches sous contraintes. Dans le cadre du projet, nous nous concentrerons sur le domaine du traitement de la langue naturelle mais nos travaux pourraient aussi s’appliquer à des tâches dans d’autres domaines.

Statistical inference and modelling for distributed systems

Statistical inference requires a large toolbox of models and algorithms that can accommodate complex data structures. Modern datasets are often so large that they need to be stored on distributed systems, with the data stored across a number of nodes with limited bandwidth between them. Many complex statistical models cannot be used with such complex data, as they rely on the complete data being accessible. In this project, we will advance statistical modeling contributions to data science by creating solutions that are ideally suited for analysis on distributed systems. More specifically, we will develop spatio-temporal models as well as accurate and efficient approximations of general statistical models that are suitable for distributed data, and as such, scalable to massive data.

Deals on Wheels: Advancing Joint ML/OR Methodologies for Enabling City-Wide, Personalized and Mobile Retail

Moving forward to a smart-city future, cities in Canada and around the world are embracing the emergence of new retail paradigms. That is, retail channels can further diversify beyond the traditional online and offline boundaries, combining the best of the both. In this project, we focus on an emerging mobile retail paradigm in which retailers run their stores on mobile vehicles or self-driving cars. Our mission is to develop cross-disciplinary models, algorithms and data-verified insights for enabling mobile retail.   We will achieve this mission by focusing on three interrelated research themes: Theme 1 – Formulating novel optimization problems of citywide siting and inventory replenishment for mobile stores. Theme 2 – Developing novel learning models for personalized demand estimation. Theme 3 – Integrating Theme 1 and Theme 2 by proposing a holistic algorithmic framework for joint and dynamic demand learning and retail operations, and for discovering managerial insights. The long-term goal is to thereby advance the synergy of operations and machine learning methodologies in the broad contexts of new retail and smart-city analytics.

Humanitarian Supply Chain Analytics

Network design problems lie at the heart of the most important issues faced in the humanitarian sector. However, given their complex nature, humanitarian supply chains involve the solution of difficult analytics problems. The main research question of this project is “how to better analyze imperfect information and address uncertainty to support decision making in humanitarian supply chains?”. To this end, we propose a methodological framework combining data analysis and optimization, which will be validated through real-life applications using multiple sources of data. First, we propose to build robust relief networks under uncertainty in demand and transportation accessibility, due to weather shocks and vulnerable infrastructures. We will consider two contexts: shelter location in Haiti and food aid distribution planning in Southeastern Asia. Second, we propose to embed fair cost sharing mechanisms into a collaborative prepositioning network design problem arising in the Caribbean. Classic economics methods will be adapted to solve large-scale stochastic optimization problems, and novel models based on catastrophic insurance theory will be proposed. Finally, a simulation will be developed to disguise data collection as a serious game and gather real-time information on the behavior of decision makers during disasters to extrapolate the best management strategies.

Retail Innovation Lab: Data Science for Socially Responsible Food Choices

In this research program, we propose to investigate the use of artificial intelligence techniques, involving data, models, behavioral analysis, and decision-making algorithms, to efficiently provide higher convenience for retail customers while being socially responsible. In particular, the research objective of the multidisciplinary team is to study, implement, and validate systems for guiding customers to make healthy food choices in a convenience store setting, while being cognizant of privacy concerns, both online and in a brick-and-mortar store environment. The creation of the digital infrastructure and decision support systems that encourage people and organizations to make health-promoting choices should hopefully result in a healthier population and reduce the costs of chronic diseases to the healthcare system. These systems should also foster the competitiveness of organizations operating in the agri-food and digital technology sectors. A distinguishing feature of this research program is that it will make use of a unique asset – a new “living-lab”, the McGill Retail Innovation Lab (MRIL). It will house a fully functioning retail store operated by a retail partner with extensive sensing, data access, and customer monitoring. The MRIL will be an invaluable source of data to use in developing and validating our approaches as well as a perfect site for running field experiments.

Combining extreme value theory and causal inference for data-driven flood hazard assessment

The IPCC reports highlight an increase in mean precipitation, but the impact of climate change on streamflow is not as certain and the existing methodology is ill-equipped to predict changes in flood extremes. Our project looks into climate drivers impacting flood hazard and proposes methodological advances based on extreme value theory and causal inference in order to simulate realistic streamflow extremes at high resolution. The project will also investigate how climate drivers impact the hydrological balance using tools from machine learning for causal discovery to enhance risk assessment of flood hazard.

Bridging Data-Driven and Behavioural Models for Transportation

Transportation data is traditionally collected through travel surveys and fixed sensors, mostly on the roadways: such data is expensive to collect and has limited spatial and temporal coverage. In recent years, more and more transportation data has become available on a continuous basis from multiple new sources, including users themselves. This has fed the rise of machine learning methods that can learn models directly from data. Yet, such models often lack robustness and may be difficult to transfer to a different region or period. This can be alleviated by taking advantage of domain knowledge stemming from the properties of the flow of people moving in transportation systems with daily activities. This project aims to develop hybrid methods relying on transportation and data-driven models to predict flows for all modes at different spatial and temporal scales using multiple sources of heterogeneous data. This results in two specific objectives: 1. to learn probabilistic flow models at the link level for several modes based on heterogeneous data; 2. to develop a method bridging the flow models (objective 1) with a dynamic multi-agent transportation model at the network level. These new models and methods will be developed and tested using real transportation data.

Ultra-Low-Energy Reliable DNN Inference Using Memristive Circuits for Biomedical Applications (ULERIM)

Recent advances in machine learning based on deep neural networks (DNNs) have brought powerful new capabilities for many signal processing tasks. These advances also hold great promises for several applications in healthcare. However, state-of-the-art DNN architectures may depend on hundreds of millions of parameters that must be stored and then retrieved, resulting in a large energy usage. Thus, it is essential to reduce their energy consumption to allow in-situ computations. One possible approach involves using memristor devices, a concept first proposed in 1971 but only recently put in practice. Memristors are a very promising way to implement compact and energy-efficient artificial neural networks.  The aim of this research is to advance the state-of-the-art in the energy-efficient implementation of deep neural networks using memristive circuits and introducing DNN-specific methods to better manage uncertainty inherent to integrated circuit fabrication. These advances will benefit a large number of medical applications for which portable devices are required to perform a complex analysis of the state of the patient, and also benefit generally the field of machine learning by reducing the amount of energy required to apply it. Within this project, the energy improvements will be exploited to improve the signal processing performance of an embedded biomedical device for the advanced detection of epileptic seizures.

Flexible multivariate spatio-temporal models for health and social sciences

Health and social economic variables are commonly observed at different spatial scales of a region (e.g. districts of a city or provinces of a country), over a given period of time. Commonly, multiple variables are observed at a given spatial unit resulting in high dimensional data. The challenge in this case is to consider models that account for the possible correlation among variables across space or space and time. This project aims at developing statistical methodology that accounts for this complex hierarchical structure of the observed data. And inference procedure follows the Bayesian paradigm meaning that uncertainty about the unknowns in the model is naturally accounted for. The project is subdivided into four sub projects that range from the estimation of a social economic vulnerability index for a given city to the spatio-temporal modelling of multiple vector borne diseases. The statistical tools proposed here will help authorities with the understanding of the dynamics across space and time of multiple diseases, and assist with the decision making process of evaluating how urban policies and programmes will impact the urban environment and population health, through a lens of health equity.

Statistical modelling of health trajectories and interventions

Large amounts of longitudinal health records are now collected in private and public healthcare systems. Data from sources such as electronic health records, healthcare administrative databases and data from mobile health applications are available to inform clinical and public health decision-making.  In many situations, such data enable the dynamic monitoring of the underlying disease process that governs the observations. However, this process is not observed directly and so inferential methods are needed to ascertain progression.  The objective of the project is to build a comprehensive Bayesian computational framework for performing inference for large scale health data. In particular, the project will focus on the analysis of records that arise in primary and clinical care contexts to study patient health trajectories, that is, how the health status of a patient changes across time. Having been able to infer the mechanisms that influence health trajectories, we will then be able to introduce treatment intervention policies that aim to improve patient outcomes.

Ultrasound classification of chronic liver disease with deep learning

Chronic liver disease is one of the top ten leading causes of death in North America. The most common form is nonalcoholic fatty liver disease which may evolve to nonalcoholic steatohepatitis and cirrhosis if left untreated. In many cases, the liver may be damaged without any symptoms. A liver biopsy is currently required to evaluate the severity of chronic liver disease. This procedure requires the insertion of a needle inside the liver to remove a small piece of tissue for examination under a microscope. Liver biopsy is an invasive procedure with a risk of major complications such as bleeding. Ultrasound is ideal for screening patients because it is a safe and widely available technology to image the whole liver. Our multi-disciplinary team is proposing the use of novel artificial intelligence techniques to assess the severity of chronic liver disease from ultrasound images and determine the severity of liver fat, inflammation, and fibrosis without the need for liver biopsy. This study is timely because chronic liver disease is on the rise which means that complications and mortality will continue to rise if there is no alternative technique for early detection and monitoring of disease severity.

Unified approach to graph structure utilization in data science

While deep neural networks are at the frontier of machine learning and data science research, their most impressive results come from data with clear spatial/temporal structure (e.g., images or audio signals) that informs network architectures to capture semantic information (e.g., textures, shapes, or phonemes). Recently, multiple attempts have been made to extend such architectures to non-Euclidean structures that typically exist in data, and in particular to graphs that model data geometry or interaction between data elements. However, so far, such attempts have been separately conducted by largely-independent communities, leveraging specific tools from traditional/spectral graph theory, graph signal processing, or applied harmonic analysis. We propose a multidisciplinary unified approach (combining computer science, applied mathematics, and decision science perspectives) for understanding deep graph processing. In particular, we will establish connections between spectral and traditional graph theory applied for this task, introduce rich notions of intrinsic graph regularity (e.g., equivalent to image textures), and enable continuous-depth graph processing (i.e., treating depth as time) to capture multiresolution local structures. Our computational framework will unify the multitude of existing disparate attempts and establish rigorous foundations for the emerging field of geometric deep learning, which is a rapidly growing field in machine learning.

Stage chez Synapse-C, spécialité science des données

Stage chez Radio-Canada, spécialité communication

Stage au CIRANO, spécialité science des données

Stage chez Synapse-C, spécialité science des données

Stage chez Synapse-C, spécialité communication

Stage chez Le Devoir, spécialité science des données

Stage chez Le Devoir, spécialité science des données

Stage au CIRANO, spécialité communication

Stage chez Radio-Canada, spécialité science des données

Mesurer l’expansion de l’Univers avec l’apprentissage automatique

Le taux d’expansion de l’Univers est une observable importante pour contraindre les modèles cosmologiques qui retracent l’évolution de l’Univers depuis le Big Bang. Récemment (2018), l’équipe du satellite Planck a publié une valeur dérivée des mesures du rayonnement fossile émis lorsque l’Univers n’était âgé que de 300,000 ans. La valeur trouvée contredit les mesures locales du paramètre, faites à partir de la vitesse de fuite des supernovas Ia et des céphéides se trouvant près de la Voie lactée. Nous proposons d’investiguer ce problème via une troisième méthode de mesure qui, jusqu’à maintenant, possédait une précision limitée par la faible quantité connue de quasar situé derrière une galaxie selon notre ligne de vue, telle que l’image du quasar est multipliée par l’effet de lentille gravitationnelle. La précision de cette méthode est limitée en grande partie par la reconstruction de la distribution de masse de la galaxie-lentille. Les avancées récentes des algorithmes d’apprentissage automatiques ont permis de démontrer qu’un réseau neuronal convolutionnel (CNN) pouvait accomplir la reconstruction de la lentille 10 millions de fois plus rapidement que les algorithmes conventionnels. Cette preuve de concept arrive juste à temps pour permettre l’analyse de la quantité phénoménale de données qui sera produite par les télescopes à champs larges dans la prochaine décennie. Nous devrons aussi adapter des architectures comme les machines à inférences récurrentes (RIM) pour automatiser le processus de reconstruction. Les besoins scientifiques de notre mission nécessitera d’adapter l’architecture de nos modèles pour l’estimation des incertitudes.

Microscopie de localisation par ultrasons fonctionnelle 3D (fULM)

L’imagerie fonctionnelle cérébrale permet de mieux comprendre quelles régions du cerveau sont impliquées dans différents types de tâches. Il est possible de réaliser ce type d’analyse à l’aide, par exemple, de l’imagerie par résonance magnétique, mais à une résolution spatiotemporelle limitée (de l’ordre du millimètre et de la seconde). Plus récemment, une autre technique, la microscopie de localisation 2D a permis de drastiquement augmenter la résolution spatiale des ultrasons (5 millièmes de millimètre), mais puisqu’elle requiert la détection de microbulles injectées individuelles (approuvées en clinique), sa résolution temporelle était insuffisante pour détecter l’activation du cerveau (dans l’ordre des minutes). Le laboratoire de Jean Provost a récemment développé une nouvelle technique d’imagerie appelée Microscopie de Localisation Ultrasonore Dynamique 3D (dMLU-3D), qui permet d’atteindre la même résolution spatiale en trois dimensions plutôt que deux et aussi une résolution élevée pour les phénomènes périodiques (de l’ordre de la milliseconde). La technique permet la visualisation de la microvasculature cérébrale (morphologie), mais la visualisation de l’activité cérébrale n’a pas encore été développée (fonction). La modélisation de ce qui caractérise une activation cérébrale dépend de plusieurs paramètres non linéaires dont il n’existe pas de vérité terrain à l’échelle de la microvasculature in-vivo, alors l’utilisation d’un réseau de neurones convolutionnel (CNN) s’avère pertinente à cette application. Ce projet vise à montrer qu’il est possible de faire de l’imagerie fonctionnelle (détecter l’activité ou le manque d’activité cérébrale) dans tout le cerveau de rongeur à l’aide de l’approche dMLU-3D avec une résolution spatiotemporelle encore jamais atteinte avec d’autres méthodes comparables. Des expériences seront réalisées afin de révéler et de corréler l’activité des régions visuelles thalamiques et corticales du cerveau du modèle murin suivant la présentation de stimuli visuels. Par la suite, ces résultats seront comparés avec ceux obtenus chez des modèles animaux de la schizophrénie (développemental, pharmacologique, lésionnel ou génétique) afin de vérifier l’hypothèse que ce désordre est caractérisé par une altération des connexions entre le cortex visuel et le thalamus. Ce projet serait la toute première démonstration de la faisabilité de l’imagerie fonctionnelle cérébrale par ultrasons superrésolus en 2D et en 3D, permettant la cartographie de l’activation cérébrale de la totalité du cerveau de rongeur ou d’autres petits animaux, tel le chat, pour des études pré-cliniques permettant à terme de mieux comprendre certaines pathologies et menant potentiellement à un meilleur diagnostic ou même traitement. C’est d’autant plus prometteur étant donné qu’aucune autre modalité d’imagerie peut atteindre une résolution aussi fine, avec une profondeur d’imagerie suffisante et ce, de manière non invasive.

Segmentation automatique d’images histologiques par apprentissage profond

Les axones de la matière blanche sont le prolongement des neurones, et constituent les autoroutes du système nerveux central. Une gaine lipidique, la myéline, entoure ces axones permettant la conduction plus rapide de l’influx nerveux. Des maladies neurodégénératives comme la sclérose en plaques ou encore des traumatismes menacent l’intégrité des axones myélinisés, ce qui peut mener à des déficits sensoriels ou moteurs tels que la douleur ou la paraplégie. Afin de développer de nouveaux traitements, les chercheurs en neurosciences ont besoin de quantifier avec précision la morphométrie de ces axones (taille, épaisseur de myéline, etc.). Mon laboratoire d’accueil NeuroPoly a développé le logiciel AxonDeepSeg permettant de faire la segmentation automatique de neurones sur des images histologiques par des algorithmes d’apprentissage profond. Cependant, AxonDeepSeg manque de robustesse vis-à-vis de la variabilité qui peut exister selon les paramètres d’acquisition et la qualité des images ainsi que selon les espèces. Ce projet vise donc à développer des modèles robustes de segmentation de neurones par l’adaptation et l’implémentation de méthodes innovantes de segmentation par apprentissage profond (Adaptation de domaine, MixUp, FiLM). Le potentiel de généralisation des algorithmes développés sera validé à l’aide de bases de données de microscopie incluant diverses modalités d’imagerie (optique, électronique à balayage, électronique en transmission), espèces, organes et pathologies. De plus, les modèles développés et les données générées seront rendus publics en accès libre et documentés afin de permettre à de nombreux chercheurs et cliniciens en neurosciences de les utiliser. Cet outil permettra également de faire la validation d’autres modalités d’imagerie essentielles dans la recherche sur les maladies neurodégénératives comme l’imagerie par résonance magnétique quantitative non-invasive, et ainsi augmenter la quantité de données utilisables par les chercheurs.

Analyse du transport d’urgence aérien dans les régions éloignées du Québec

Dans un objectif d’offrir des soins spécialisés à l’ensemble de sa population, le Québec peut compter sur le programme d’Évacuation aeromédicales du Québec (EVAQ) mis en place par le gouvernement. L’offre de service permet de transférer des patients depuis les différentes régions du Québec vers des centres spécialisés de Québec et Montréal afin de leur offrir les soins nécessaires, le tout entouré d’une équipe médicale adaptée à leur condition et leur niveau d’urgence. Plusieurs des services offerts par l’EVAQ ont connu une augmentation de la demande durant la dernière décennie. La présente recherche vise à bâtir un outil de simulation qui permettra de simuler différentes utilisations des ressources. L’analyse des différents scénarios permettra de faire des recommandations à l’ÉVAQ sur les actions à prendre afin d’offrir le meilleur niveau de service possible aux populations des régions. Il y a beaucoup à apprendre sur le modèle instauré au Québec, tant au niveau de la planification stratégique des appareils et des trajets, qu’au niveau de la coordination et des opérations au quotidien. La densité de population, les distances à franchir et les conditions météorologiques difficiles sont des facteurs déterminants à considérer dans leur unicité.

Optimisation de l’entraînement des réseaux de neurones profonds à partir d’extensions de l’algorithme MADS sur les hyperparamètres de type variable de catégorie

Ce projet de maîtrise vise à optimiser l’entraînement des réseaux de neurones profonds à partir d’extensions de l’algorithme MADS sur les hyperparamètres de type variable de catégorie. Ces hyperparamètres sont généralement choisis de manière arbitraire ou heuristique. Or, la plupart des algorithmes d’optimisation développés solutionnent des problèmes où les variables sont de type continu ou entier. En d’autres mots, il existe peu de méthodes d’optimisation pouvant traiter efficacement les variables de catégorie. Cependant, puisque ces variables sont discrètes, il est possible de construire et d’explorer un espace de variables discrétisées avec les méthodes d’optimisation dites recherche directe. Le projet de recherche a pour objectif d’adapter les récents développements de l’algorithme MADS (« Mesh Adaptive Direct Search ») aux variables de catégorie, notamment pour le traitement des contraintes et l’intégration d’un treillis anisotrope dynamique. Plus précisément, nous nous intéressons à optimiser plus rigoureusement les hyperparamètres des réseaux de neurones profonds, afin d’entraîner plus intelligemment les modèles d’intelligence artificielle. Plus particulièrement, les hyperparamètres étudiés seront : la fonction de perte ; les extensions et les modifications à l’algorithme de rétropropagation (ADAM, RMSProp, etc.) ainsi que les régulateurs (LASSO, « Ridge regression », etc.). Les mécanismes développés pourront également servir à modéliser la topologie des réseaux (nombres de couches, nombres de neurones, etc.) En effet, dans le cadre de l’algorithme MADS, le traitement des variables de catégorie pourraient s’étendre à des variables discrètes, dont la valeur modifie la dimension du problème. En pratique, le système résultant permettra donc, pour la première fois, d’optimiser simultanément les hyperparamètres reliés à l’entraînement et ceux reliés à la topologie.

Analyzing mouse hippocampal « time cell » activities during memory task with machine learning approaches

Previous studies have identified hippocampal “time cells” in CA1 that bridge the temporal gap between discontiguous events by firing in tiling patterns during the delay period of memory tasks, such as alternative maze (Pastalkova et al., 2008) and object-odor pairing tasks (MacDonald et al., 2011). However, recent findings have argued that this tiling might be an analysis artifact due to cell-sorting because it also appears in tasks with no memory load (Salz et al., 2016). To address this discrepancy, our collaborators have collected calcium recordings in mouse hippocampal CA1 region during trial unique, nonmatch-to-location (TUNL) task (Talpos et al., 2010) and showed tiling patterns. Our objective is to use computational methods to determine if these patterns are meaningful. To do this, we will first train decoders on the calcium recordings to decode sample for each trial, with temporal sequences preserved (i.e. sorted tiling columns) or shuffled (i.e. randomized columns). If the tiling patterns are indeed meaningful, we would expect to see higher accuracy of the decoder in the preserved sequences. Our next step is to construct a simulated reinforcement learning agent on simulated TUNL task to see whether there exists a consistent tiling pattern in the activities of the neural networks of the reinforcement learning agent. If so, it would suggest that these patterns play a role in preserving information about the sample location during the delay period as a solution to the task. If not, it would suggest that the tiling patterns previously observed in memory tasks could merely be a ubiquitous artifact. Our findings would have a significant impact on the current view of hippocampal “time cells” as well as the functional segregation of the brain.

Jumpy, Hierarchical and Adversarial Variational Video Prediction

This project is in the context of intelligent transportation systems. To improve road user detection and tracking, we want to predict their position in future frames using video prediction. However, predicting high fidelity videos over long time scale is notoriously difficult. Current video prediction models either diverges from real samples after a few frames or fail to capture stochasticity in the videos, resulting in bad prediction performance for long videos. In order to overcome this difficulty, new models with ability to do jumpy or hierarchical video prediction are proposed by the AI community. In this proposal, we propose to further develop these ideas and explore new models for stochastic video prediction that is able to do jumpy predictions in a hierarchical manner. We mainly want to explore two research problems: (1) How to do stochastic jumpy video predictions. (2) How to combine jumpy prediction with temporal abstraction.

Emergent Behaviour in Multi-Agent Reinforcement Learning

This project aims for the investigation of intricate emergent behaviours in large scale multi-agent reinforcement learning (MARL). Of particular concern are the behaviours of agents in settings where they are tightly interdependent to the point of nearly composing a single entity. Such settings will draw strong inspiration from biological systems, and be achieved either through a shared common reward or through complex and necessary interactions. Because large interconnected populations of agents present a novel collection of settings complete with new challenges, this project will force a rethinking of well-established reinforcement learning practices, all while probing the limits of their scalability. Furthermore, enabling MARL systems that simultaneously achieve large population scales and appropriate complexity will allow for better modelling of intricate phenomena that have been out of reach of previous artificial intelligence methods. This would potentially result in far-reaching benefits in other scientific disciplines, thus broadening the range of applications of reinforcement learning and simultaneously opening it to easier idea cross-pollination from other fields. These settings will be studied empirically by analyzing the behaviour of existing MARL algorithms, and by comparing and contrasting them to new approaches that allow for more complex interactions between agents. The analysis of the results will be performed quantitatively on the basis of standard reinforcement learning and game theoretic methodology, and qualitatively in light of the principles of behavioural biology. The implementation of the environments and the MARL models will be done with modularity and concurrency in mind and the code-base will then be openly released.

Stylistic Controls for Neural Text Generation

Deep learning-based approaches to text generation have proven effective in recent years, with many models able to generate realistic text, often exhibiting higher-order structure. While these models produce high-quality samples, there is usually little control provided over what is specifically generated. Recently, work has begun in this area, but much remains to be explored. This application proposes research towards controllable text generation by implementing a variety of stylistic controls that can be used to influence what is sampled from a neural language model. In my previous work, we developed a conditional generative model for music. We demonstrated that we could control for a variety of characteristics during generation by providing the model with an additional externally-specified input called the control signal. For instance, in this work, we trained a model using a composer-based control signal. This signal identified the composer of each piece on which the model was trained. After training, we used the control signal to produce samples of music in the style of specific composers, for instance, Bach and Beethoven. Based on my previous work with music, we are now interested in implementing a similar set of controls for generating text. Such a set of stylistic controls would extend the practical utility of text generated from neural language models. We plan to explore generation methods involving supervised controls and latent (disentangled) controls.

Utilisation d’approches d’apprentissage automatique pour l’identification d’une signature cérébrale de l’expression faciale de la douleur

L’expression faciale est un outil important pour communiquer diverses informations, notamment la manifestation d’un état de douleur, la présence d’une menace immédiate dans l’environnement et un éventuel besoin d’aide. Les dimensions sensorielle (intensité) et affective (caractère déplaisant) de la douleur peuvent être encodées dans les mouvements faciaux. Les techniques d’analyse jusqu’à présent utilisées pour examiner la relation entre l’expression faciale et l’activité cérébrale lors de l’expérience de la douleur possèdent plusieurs limitations statistiques par rapport à l’évaluation de l’activité cérébrale spatialement distribuée. L’objectif principal du projet proposé est de mieux comprendre les mécanismes neuronaux qui sous-tendent l’expression faciale de la douleur. Des données d’imagerie par résonance magnétique fonctionnelle (IRMf) seront utilisées pour analyser les changements dans l’activité cérébrale en réponse à des stimuli douloureux (mais non dommageables). Plus spécifiquement, ce projet vise à utiliser des approches d’apprentissage automatique (c’est-à-dire l’analyse de modèles multivariés) pour développer une signature cérébrale de l’expression faciale de la douleur afin de prédire les changements faciaux en réponse à des stimuli douloureux dans différents contextes : douleur phasique (stimulation courte), douleur tonique (stimulation longue), et modulation des dimensions sensorielle et affective de la douleur. En bref, ce projet permettra de résoudre certaines lacunes des analyses univariés précédemment utilisées afin de déterminer avec une meilleure précision les bases neurales de l’expression faciale de la douleur et de faire progresser de manière significative notre compréhension des mécanismes cérébraux qui sous-tendent la communication non verbale.

Les premières images d’exoplanètes orbitant autour de naines blanches, d’étoiles à neutrons et de trous noirs

Les binaires X, formés d’une étoile orbitant autour d’un objet compact stellaire compact (naine blanche, étoile à neutrons ou trou noir), sont des laboratoires fantastiques pour comprendre la physique dans des conditions extrêmes. Au cours des dernières décennies, les binaires X ont fait l’objet d’une multitude d’études dans diverses longueurs d’onde, conduisant à des avancées remarquables dans le domaine de la physique de l’accrétion, ainsi que dans la compréhension de la formation de jets de particules relativistes dans de puissants champs magnétiques. Les binaires X sont aussi d’excellents laboratoires pour comprendre les explosions de type supernova ainsi que l’effet de ces explosions sur le système et son environnement. En effet, la présence d’une étoile à neutrons ou d’un trou noir dans ces systèmes implique directement que l’étoile (et ses potentielles planètes) survivent à ces explosions. Plusieurs études montrent que les planètes et les naines brunes peuvent exister dans une multitude d’environnements, tels que celles qui orbitent très proche de leur étoile hôte (Jupiters chaudes) ou celles qui orbitent à des distances de centaines d’unités astronomiques de l’étoile. Ces découvertes montrent que la formation et la survie des planètes sont mal comprises. Par conséquent, ce projet amène un nouveau point de vue, soit celui des conditions extrêmes. Bref, on pourra étudier plusieurs binaires X et des données des télescopes NIRC2/KECK (visible) et NOEMA (millimétrique) ont déjà été acquises en 2018, et d’autres demandes de temps sont en cours. Selon une analyse préliminaire, la présence d’objets astrophysiques est confirmée, et donc ce projet garantit des résultats surprenants pour la communauté de l’astrophysique.

Addressing the retrosynthesis problem using a graph-to-graph translation network

Retrosynthesis analysis, which aims to identify a set of reactant graphs to synthesize a target molecule, is a fundamental problem in computational chemistry and is of central importance to the organic synthesis planing as well as drug discovery. The problem is challenging as the search space of all possible transformations is very huge. For decades, people have been seeking to assist chemists in retrosynthesis analysis with modern computing algorithms. Most existing machine learning works on this task rely on reaction templates that define the subgraph patterns of a set of chemical reactions, which require expensive graph isomorphism and suffer from poor generalization on unseen molecule structures.

To address the above limitations, in this project, we formulate the retrosynthesis prediction as a graph-to-graph translation task, i.e., translating a product graph to a set of reactant graphs, and propose a novel template-free approach to tackle the problem. We will show that our method excludes the need of domain knowledge, and scales well to large datasets. We will also empirically verify the superiority of our method on the benchmark data set.

A Multi-agent Decision and Control Framework for Mixed-autonomy Transportation System

As the autonomous vehicle becomes more and more popular. Recently, there has been a new emphasis on traffic control in the context of mixed-autonomy, where only a fraction of vehicles are connected autonomous vehicles and interacting with human-driven vehicles. As in a mixed autonomy system, there are several challenges. The first challenge is how to encourage different agents’ cooperation so as to maximize the total returns of the whole system. For example, when there is a gap in front of the adjacent line of the autonomous vehicle, if the autonomous vehicle cuts in immediately, the surrounding vehicle in the adjacent line will also decrease its speed sharply, which will end up a shock wave in traffic flow. Instead, if the autonomous vehicle learns to cooperate with other agents, it will adjust its speed steadily and try to mitigate the negative impact on the whole system. The second challenge is how to improve the communication efficiency in multi-agent system. As autonomous vehicle has different characteristics with human-driven agent, for example, their reacting time and action may be different. Therefore, how to formalize personalized policy for each agent is also worth to explore. The third challenge is how to explore expert knowledge (e.g. green wave, max pressure, actuated control) in transportation domain to improve the training efficiency and performance. Our overall goal of this project is to design effective decision and control framework for an efficient and safe mixed autonomy system by mitigating the shockwave and improving the transportation efficiency. To address the aforementioned problems, we will develop a novel multi agent decision framework based on deep reinforcement learning to improve the decision making and control performance of the agents in mixed autonomy system.

Exploiting Experiences and Priors in Semantic Visual Navigation

Robotics has always been anticipated to revolutionize the world. However, despite the significant progress over the past few decades, robots have yet to be able to reliably navigate within an unstructured indoor environment. Semantic visual navigation is the task of navigating within a possibly unknown environment using only visual sensors, such as asking a household robot agent to “go to the kitchen”. Traditional “modular” methods combine a Simultaneous Localization and Mapping (SLAM) component with separate search, planning, and control modules. However, these methods do not scale well to large environments, and require significant engineering efforts. Alternatively, end-to-end “learning” solutions produce agent policies that directly infer actions from camera frames, by applying Deep Reinforcement Learning (DRL) techniques on large-scale datasets. Nevertheless, these policies tend to be reactive, do not explicitly exploit scene geometry, and are not data efficient. Furthermore, both modular and learning-based approaches do not sufficiently exploit knowledge from past task instances to improve subsequent search performance in both repeated environments as well as unseen yet similar environments. Our project explores the learning and use of spatial-semantic priors for more efficient semantic visual navigation. We aim to devise a framework that learns, updates, and exploits a topological-semantic map between discovered locations and objects within. We hypothesize that these advances will result in agents that generalize better to unseen similar environments, as well as becoming increasingly more efficient during repeated search queries within the same environment.

Statistical Modeling Framework to Understand Dynamic Traffic Patterns from Video Data

Video-based traffic monitoring systems, as the backbone of modern Intelligent Transportation Systems (ITS), is playing an essential role in sensing traffic conditions and detecting abnormal events/incidents. Semantically understanding traffic scenes and automatically mining the traffic patterns from video data of a static camera can help with traffic situation analysis and anomaly events warning. Given a video of a dynamic traffic scene with several different behaviors happening simultaneously, we want the ITS to learn and understand: “How many typical traffic patterns are in the video? How to semantically interpret these patterns? What are the rules governing the transitions between these patterns?”In this project, we will mainly focus on traffic patterns recognition and anomaly detection from video data, we will: (i) construct representation learning model to extract efficient features; and (ii) develop an unsupervised learning framework based on Bayesian nonparametrics to automatically learn the traffic patterns.

Mesurer l’expansion de l’Univers avec l’apprentissage automatique

Le taux d’expansion de l’Univers est une observable importante pour contraindre les modèles cosmologiques qui retracent l’évolution de l’Univers depuis le Big Bang. Récemment (2018), l’équipe du satellite Planck a publié une valeur dérivée des mesures du rayonnement fossile émis lorsque l’Univers n’était âgé que de 300,000 ans. La valeur trouvée contredit les mesures locales du paramètre, faites à partir de la vitesse de fuite des supernovas Ia et des céphéides se trouvant près de la Voie lactée. Nous proposons d’investiguer ce problème via une troisième méthode de mesure qui, jusqu’à maintenant, possédait une précision limitée par la faible quantité connue de quasar situé derrière une galaxie selon notre ligne de vue, telle que l’image du quasar est multipliée par l’effet de lentille gravitationnelle. La précision de cette méthode est limitée en grande partie par la reconstruction de la distribution de masse de la galaxie-lentille. Les avancées récentes des algorithmes d’apprentissage automatiques ont permis de démontrer qu’un réseau neuronal convolutionnel (CNN) pouvait accomplir la reconstruction de la lentille 10 millions de fois plus rapidement que les algorithmes conventionnels. Cette preuve de concept arrive juste à temps pour permettre l’analyse de la quantité phénoménale de données qui sera produite par les télescopes à champs larges dans la prochaine décennie. Nous devrons aussi adapter des architectures comme les machines à inférences récurrentes (RIM) pour automatiser le processus de reconstruction. Les besoins scientifiques de notre mission nécessitera d’adapter l’architecture de nos modèles pour l’estimation des incertitudes.

Microscopie de localisation par ultrasons fonctionnelle 3D (fULM)

L’imagerie fonctionnelle cérébrale permet de mieux comprendre quelles régions du cerveau sont impliquées dans différents types de tâches. Il est possible de réaliser ce type d’analyse à l’aide, par exemple, de l’imagerie par résonance magnétique, mais à une résolution spatiotemporelle limitée (de l’ordre du millimètre et de la seconde). Plus récemment, une autre technique, la microscopie de localisation 2D a permis de drastiquement augmenter la résolution spatiale des ultrasons (5 millièmes de millimètre), mais puisqu’elle requiert la détection de microbulles injectées individuelles (approuvées en clinique), sa résolution temporelle était insuffisante pour détecter l’activation du cerveau (dans l’ordre des minutes). Le laboratoire de Jean Provost a récemment développé une nouvelle technique d’imagerie appelée Microscopie de Localisation Ultrasonore Dynamique 3D (dMLU-3D), qui permet d’atteindre la même résolution spatiale en trois dimensions plutôt que deux et aussi une résolution élevée pour les phénomènes périodiques (de l’ordre de la milliseconde). La technique permet la visualisation de la microvasculature cérébrale (morphologie), mais la visualisation de l’activité cérébrale n’a pas encore été développée (fonction). La modélisation de ce qui caractérise une activation cérébrale dépend de plusieurs paramètres non linéaires dont il n’existe pas de vérité terrain à l’échelle de la microvasculature in-vivo, alors l’utilisation d’un réseau de neurones convolutionnel (CNN) s’avère pertinente à cette application. Ce projet vise à montrer qu’il est possible de faire de l’imagerie fonctionnelle (détecter l’activité ou le manque d’activité cérébrale) dans tout le cerveau de rongeur à l’aide de l’approche dMLU-3D avec une résolution spatiotemporelle encore jamais atteinte avec d’autres méthodes comparables. Des expériences seront réalisées afin de révéler et de corréler l’activité des régions visuelles thalamiques et corticales du cerveau du modèle murin suivant la présentation de stimuli visuels. Par la suite, ces résultats seront comparés avec ceux obtenus chez des modèles animaux de la schizophrénie (développemental, pharmacologique, lésionnel ou génétique) afin de vérifier l’hypothèse que ce désordre est caractérisé par une altération des connexions entre le cortex visuel et le thalamus. Ce projet serait la toute première démonstration de la faisabilité de l’imagerie fonctionnelle cérébrale par ultrasons superrésolus en 2D et en 3D, permettant la cartographie de l’activation cérébrale de la totalité du cerveau de rongeur ou d’autres petits animaux, tel le chat, pour des études pré-cliniques permettant à terme de mieux comprendre certaines pathologies et menant potentiellement à un meilleur diagnostic ou même traitement. C’est d’autant plus prometteur étant donné qu’aucune autre modalité d’imagerie peut atteindre une résolution aussi fine, avec une profondeur d’imagerie suffisante et ce, de manière non invasive.

Segmentation automatique d’images histologiques par apprentissage profond

Les axones de la matière blanche sont le prolongement des neurones, et constituent les autoroutes du système nerveux central. Une gaine lipidique, la myéline, entoure ces axones permettant la conduction plus rapide de l’influx nerveux. Des maladies neurodégénératives comme la sclérose en plaques ou encore des traumatismes menacent l’intégrité des axones myélinisés, ce qui peut mener à des déficits sensoriels ou moteurs tels que la douleur ou la paraplégie. Afin de développer de nouveaux traitements, les chercheurs en neurosciences ont besoin de quantifier avec précision la morphométrie de ces axones (taille, épaisseur de myéline, etc.). Mon laboratoire d’accueil NeuroPoly a développé le logiciel AxonDeepSeg permettant de faire la segmentation automatique de neurones sur des images histologiques par des algorithmes d’apprentissage profond. Cependant, AxonDeepSeg manque de robustesse vis-à-vis de la variabilité qui peut exister selon les paramètres d’acquisition et la qualité des images ainsi que selon les espèces. Ce projet vise donc à développer des modèles robustes de segmentation de neurones par l’adaptation et l’implémentation de méthodes innovantes de segmentation par apprentissage profond (Adaptation de domaine, MixUp, FiLM). Le potentiel de généralisation des algorithmes développés sera validé à l’aide de bases de données de microscopie incluant diverses modalités d’imagerie (optique, électronique à balayage, électronique en transmission), espèces, organes et pathologies. De plus, les modèles développés et les données générées seront rendus publics en accès libre et documentés afin de permettre à de nombreux chercheurs et cliniciens en neurosciences de les utiliser. Cet outil permettra également de faire la validation d’autres modalités d’imagerie essentielles dans la recherche sur les maladies neurodégénératives comme l’imagerie par résonance magnétique quantitative non-invasive, et ainsi augmenter la quantité de données utilisables par les chercheurs.

Analyse du transport d’urgence aérien dans les régions éloignées du Québec

Dans un objectif d’offrir des soins spécialisés à l’ensemble de sa population, le Québec peut compter sur le programme d’Évacuation aeromédicales du Québec (EVAQ) mis en place par le gouvernement. L’offre de service permet de transférer des patients depuis les différentes régions du Québec vers des centres spécialisés de Québec et Montréal afin de leur offrir les soins nécessaires, le tout entouré d’une équipe médicale adaptée à leur condition et leur niveau d’urgence. Plusieurs des services offerts par l’EVAQ ont connu une augmentation de la demande durant la dernière décennie. La présente recherche vise à bâtir un outil de simulation qui permettra de simuler différentes utilisations des ressources. L’analyse des différents scénarios permettra de faire des recommandations à l’ÉVAQ sur les actions à prendre afin d’offrir le meilleur niveau de service possible aux populations des régions. Il y a beaucoup à apprendre sur le modèle instauré au Québec, tant au niveau de la planification stratégique des appareils et des trajets, qu’au niveau de la coordination et des opérations au quotidien. La densité de population, les distances à franchir et les conditions météorologiques difficiles sont des facteurs déterminants à considérer dans leur unicité.

Optimisation de l’entraînement des réseaux de neurones profonds à partir d’extensions de l’algorithme MADS sur les hyperparamètres de type variable de catégorie

Ce projet de maîtrise vise à optimiser l’entraînement des réseaux de neurones profonds à partir d’extensions de l’algorithme MADS sur les hyperparamètres de type variable de catégorie. Ces hyperparamètres sont généralement choisis de manière arbitraire ou heuristique. Or, la plupart des algorithmes d’optimisation développés solutionnent des problèmes où les variables sont de type continu ou entier. En d’autres mots, il existe peu de méthodes d’optimisation pouvant traiter efficacement les variables de catégorie. Cependant, puisque ces variables sont discrètes, il est possible de construire et d’explorer un espace de variables discrétisées avec les méthodes d’optimisation dites recherche directe. Le projet de recherche a pour objectif d’adapter les récents développements de l’algorithme MADS (« Mesh Adaptive Direct Search ») aux variables de catégorie, notamment pour le traitement des contraintes et l’intégration d’un treillis anisotrope dynamique. Plus précisément, nous nous intéressons à optimiser plus rigoureusement les hyperparamètres des réseaux de neurones profonds, afin d’entraîner plus intelligemment les modèles d’intelligence artificielle. Plus particulièrement, les hyperparamètres étudiés seront : la fonction de perte ; les extensions et les modifications à l’algorithme de rétropropagation (ADAM, RMSProp, etc.) ainsi que les régulateurs (LASSO, « Ridge regression », etc.). Les mécanismes développés pourront également servir à modéliser la topologie des réseaux (nombres de couches, nombres de neurones, etc.) En effet, dans le cadre de l’algorithme MADS, le traitement des variables de catégorie pourraient s’étendre à des variables discrètes, dont la valeur modifie la dimension du problème. En pratique, le système résultant permettra donc, pour la première fois, d’optimiser simultanément les hyperparamètres reliés à l’entraînement et ceux reliés à la topologie.

Analyzing mouse hippocampal « time cell » activities during memory task with machine learning approaches

Previous studies have identified hippocampal “time cells” in CA1 that bridge the temporal gap between discontiguous events by firing in tiling patterns during the delay period of memory tasks, such as alternative maze (Pastalkova et al., 2008) and object-odor pairing tasks (MacDonald et al., 2011). However, recent findings have argued that this tiling might be an analysis artifact due to cell-sorting because it also appears in tasks with no memory load (Salz et al., 2016). To address this discrepancy, our collaborators have collected calcium recordings in mouse hippocampal CA1 region during trial unique, nonmatch-to-location (TUNL) task (Talpos et al., 2010) and showed tiling patterns. Our objective is to use computational methods to determine if these patterns are meaningful. To do this, we will first train decoders on the calcium recordings to decode sample for each trial, with temporal sequences preserved (i.e. sorted tiling columns) or shuffled (i.e. randomized columns). If the tiling patterns are indeed meaningful, we would expect to see higher accuracy of the decoder in the preserved sequences. Our next step is to construct a simulated reinforcement learning agent on simulated TUNL task to see whether there exists a consistent tiling pattern in the activities of the neural networks of the reinforcement learning agent. If so, it would suggest that these patterns play a role in preserving information about the sample location during the delay period as a solution to the task. If not, it would suggest that the tiling patterns previously observed in memory tasks could merely be a ubiquitous artifact. Our findings would have a significant impact on the current view of hippocampal “time cells” as well as the functional segregation of the brain.

Jumpy, Hierarchical and Adversarial Variational Video Prediction

This project is in the context of intelligent transportation systems. To improve road user detection and tracking, we want to predict their position in future frames using video prediction. However, predicting high fidelity videos over long time scale is notoriously difficult. Current video prediction models either diverges from real samples after a few frames or fail to capture stochasticity in the videos, resulting in bad prediction performance for long videos. In order to overcome this difficulty, new models with ability to do jumpy or hierarchical video prediction are proposed by the AI community. In this proposal, we propose to further develop these ideas and explore new models for stochastic video prediction that is able to do jumpy predictions in a hierarchical manner. We mainly want to explore two research problems: (1) How to do stochastic jumpy video predictions. (2) How to combine jumpy prediction with temporal abstraction.

Emergent Behaviour in Multi-Agent Reinforcement Learning

This project aims for the investigation of intricate emergent behaviours in large scale multi-agent reinforcement learning (MARL). Of particular concern are the behaviours of agents in settings where they are tightly interdependent to the point of nearly composing a single entity. Such settings will draw strong inspiration from biological systems, and be achieved either through a shared common reward or through complex and necessary interactions. Because large interconnected populations of agents present a novel collection of settings complete with new challenges, this project will force a rethinking of well-established reinforcement learning practices, all while probing the limits of their scalability. Furthermore, enabling MARL systems that simultaneously achieve large population scales and appropriate complexity will allow for better modelling of intricate phenomena that have been out of reach of previous artificial intelligence methods. This would potentially result in far-reaching benefits in other scientific disciplines, thus broadening the range of applications of reinforcement learning and simultaneously opening it to easier idea cross-pollination from other fields. These settings will be studied empirically by analyzing the behaviour of existing MARL algorithms, and by comparing and contrasting them to new approaches that allow for more complex interactions between agents. The analysis of the results will be performed quantitatively on the basis of standard reinforcement learning and game theoretic methodology, and qualitatively in light of the principles of behavioural biology. The implementation of the environments and the MARL models will be done with modularity and concurrency in mind and the code-base will then be openly released.

Stylistic Controls for Neural Text Generation

Deep learning-based approaches to text generation have proven effective in recent years, with many models able to generate realistic text, often exhibiting higher-order structure. While these models produce high-quality samples, there is usually little control provided over what is specifically generated. Recently, work has begun in this area, but much remains to be explored. This application proposes research towards controllable text generation by implementing a variety of stylistic controls that can be used to influence what is sampled from a neural language model. In my previous work, we developed a conditional generative model for music. We demonstrated that we could control for a variety of characteristics during generation by providing the model with an additional externally-specified input called the control signal. For instance, in this work, we trained a model using a composer-based control signal. This signal identified the composer of each piece on which the model was trained. After training, we used the control signal to produce samples of music in the style of specific composers, for instance, Bach and Beethoven. Based on my previous work with music, we are now interested in implementing a similar set of controls for generating text. Such a set of stylistic controls would extend the practical utility of text generated from neural language models. We plan to explore generation methods involving supervised controls and latent (disentangled) controls.

Utilisation d’approches d’apprentissage automatique pour l’identification d’une signature cérébrale de l’expression faciale de la douleur

L’expression faciale est un outil important pour communiquer diverses informations, notamment la manifestation d’un état de douleur, la présence d’une menace immédiate dans l’environnement et un éventuel besoin d’aide. Les dimensions sensorielle (intensité) et affective (caractère déplaisant) de la douleur peuvent être encodées dans les mouvements faciaux. Les techniques d’analyse jusqu’à présent utilisées pour examiner la relation entre l’expression faciale et l’activité cérébrale lors de l’expérience de la douleur possèdent plusieurs limitations statistiques par rapport à l’évaluation de l’activité cérébrale spatialement distribuée. L’objectif principal du projet proposé est de mieux comprendre les mécanismes neuronaux qui sous-tendent l’expression faciale de la douleur. Des données d’imagerie par résonance magnétique fonctionnelle (IRMf) seront utilisées pour analyser les changements dans l’activité cérébrale en réponse à des stimuli douloureux (mais non dommageables). Plus spécifiquement, ce projet vise à utiliser des approches d’apprentissage automatique (c’est-à-dire l’analyse de modèles multivariés) pour développer une signature cérébrale de l’expression faciale de la douleur afin de prédire les changements faciaux en réponse à des stimuli douloureux dans différents contextes : douleur phasique (stimulation courte), douleur tonique (stimulation longue), et modulation des dimensions sensorielle et affective de la douleur. En bref, ce projet permettra de résoudre certaines lacunes des analyses univariés précédemment utilisées afin de déterminer avec une meilleure précision les bases neurales de l’expression faciale de la douleur et de faire progresser de manière significative notre compréhension des mécanismes cérébraux qui sous-tendent la communication non verbale.

Les premières images d’exoplanètes orbitant autour de naines blanches, d’étoiles à neutrons et de trous noirs

Les binaires X, formés d’une étoile orbitant autour d’un objet compact stellaire compact (naine blanche, étoile à neutrons ou trou noir), sont des laboratoires fantastiques pour comprendre la physique dans des conditions extrêmes. Au cours des dernières décennies, les binaires X ont fait l’objet d’une multitude d’études dans diverses longueurs d’onde, conduisant à des avancées remarquables dans le domaine de la physique de l’accrétion, ainsi que dans la compréhension de la formation de jets de particules relativistes dans de puissants champs magnétiques. Les binaires X sont aussi d’excellents laboratoires pour comprendre les explosions de type supernova ainsi que l’effet de ces explosions sur le système et son environnement. En effet, la présence d’une étoile à neutrons ou d’un trou noir dans ces systèmes implique directement que l’étoile (et ses potentielles planètes) survivent à ces explosions. Plusieurs études montrent que les planètes et les naines brunes peuvent exister dans une multitude d’environnements, tels que celles qui orbitent très proche de leur étoile hôte (Jupiters chaudes) ou celles qui orbitent à des distances de centaines d’unités astronomiques de l’étoile. Ces découvertes montrent que la formation et la survie des planètes sont mal comprises. Par conséquent, ce projet amène un nouveau point de vue, soit celui des conditions extrêmes. Bref, on pourra étudier plusieurs binaires X et des données des télescopes NIRC2/KECK (visible) et NOEMA (millimétrique) ont déjà été acquises en 2018, et d’autres demandes de temps sont en cours. Selon une analyse préliminaire, la présence d’objets astrophysiques est confirmée, et donc ce projet garantit des résultats surprenants pour la communauté de l’astrophysique.

Addressing the retrosynthesis problem using a graph-to-graph translation network

Retrosynthesis analysis, which aims to identify a set of reactant graphs to synthesize a target molecule, is a fundamental problem in computational chemistry and is of central importance to the organic synthesis planing as well as drug discovery. The problem is challenging as the search space of all possible transformations is very huge. For decades, people have been seeking to assist chemists in retrosynthesis analysis with modern computing algorithms. Most existing machine learning works on this task rely on reaction templates that define the subgraph patterns of a set of chemical reactions, which require expensive graph isomorphism and suffer from poor generalization on unseen molecule structures.

To address the above limitations, in this project, we formulate the retrosynthesis prediction as a graph-to-graph translation task, i.e., translating a product graph to a set of reactant graphs, and propose a novel template-free approach to tackle the problem. We will show that our method excludes the need of domain knowledge, and scales well to large datasets. We will also empirically verify the superiority of our method on the benchmark data set.

A Multi-agent Decision and Control Framework for Mixed-autonomy Transportation System

As the autonomous vehicle becomes more and more popular. Recently, there has been a new emphasis on traffic control in the context of mixed-autonomy, where only a fraction of vehicles are connected autonomous vehicles and interacting with human-driven vehicles. As in a mixed autonomy system, there are several challenges. The first challenge is how to encourage different agents’ cooperation so as to maximize the total returns of the whole system. For example, when there is a gap in front of the adjacent line of the autonomous vehicle, if the autonomous vehicle cuts in immediately, the surrounding vehicle in the adjacent line will also decrease its speed sharply, which will end up a shock wave in traffic flow. Instead, if the autonomous vehicle learns to cooperate with other agents, it will adjust its speed steadily and try to mitigate the negative impact on the whole system. The second challenge is how to improve the communication efficiency in multi-agent system. As autonomous vehicle has different characteristics with human-driven agent, for example, their reacting time and action may be different. Therefore, how to formalize personalized policy for each agent is also worth to explore. The third challenge is how to explore expert knowledge (e.g. green wave, max pressure, actuated control) in transportation domain to improve the training efficiency and performance. Our overall goal of this project is to design effective decision and control framework for an efficient and safe mixed autonomy system by mitigating the shockwave and improving the transportation efficiency. To address the aforementioned problems, we will develop a novel multi agent decision framework based on deep reinforcement learning to improve the decision making and control performance of the agents in mixed autonomy system.

Exploiting Experiences and Priors in Semantic Visual Navigation

Robotics has always been anticipated to revolutionize the world. However, despite the significant progress over the past few decades, robots have yet to be able to reliably navigate within an unstructured indoor environment. Semantic visual navigation is the task of navigating within a possibly unknown environment using only visual sensors, such as asking a household robot agent to “go to the kitchen”. Traditional “modular” methods combine a Simultaneous Localization and Mapping (SLAM) component with separate search, planning, and control modules. However, these methods do not scale well to large environments, and require significant engineering efforts. Alternatively, end-to-end “learning” solutions produce agent policies that directly infer actions from camera frames, by applying Deep Reinforcement Learning (DRL) techniques on large-scale datasets. Nevertheless, these policies tend to be reactive, do not explicitly exploit scene geometry, and are not data efficient. Furthermore, both modular and learning-based approaches do not sufficiently exploit knowledge from past task instances to improve subsequent search performance in both repeated environments as well as unseen yet similar environments. Our project explores the learning and use of spatial-semantic priors for more efficient semantic visual navigation. We aim to devise a framework that learns, updates, and exploits a topological-semantic map between discovered locations and objects within. We hypothesize that these advances will result in agents that generalize better to unseen similar environments, as well as becoming increasingly more efficient during repeated search queries within the same environment.

Statistical Modeling Framework to Understand Dynamic Traffic Patterns from Video Data

Video-based traffic monitoring systems, as the backbone of modern Intelligent Transportation Systems (ITS), is playing an essential role in sensing traffic conditions and detecting abnormal events/incidents. Semantically understanding traffic scenes and automatically mining the traffic patterns from video data of a static camera can help with traffic situation analysis and anomaly events warning. Given a video of a dynamic traffic scene with several different behaviors happening simultaneously, we want the ITS to learn and understand: “How many typical traffic patterns are in the video? How to semantically interpret these patterns? What are the rules governing the transitions between these patterns?”In this project, we will mainly focus on traffic patterns recognition and anomaly detection from video data, we will: (i) construct representation learning model to extract efficient features; and (ii) develop an unsupervised learning framework based on Bayesian nonparametrics to automatically learn the traffic patterns.