Towards Time-Aware Language Models: Modeling, Retrieval, and Reasoning

Par

Adam Jatowt

University of Innsbruck

Vendredi 24 avril 2026, 10:30-11:30 EST, Salle 3195

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Time is central to how we understand the world, shaping how we interpret events, construct narratives, and reason about cause and effect. In this talk, I examine how well large language models handle temporal reasoning—an essential yet underexplored aspect of language intelligence. I will discuss our recent advances in incorporating temporal awareness into language models, including approaches to time-aware representation learning, retrieval from evolving document collections, and the development of new resources for training and evaluation. I will also introduce temporal reasoning-oriented QA and retrieval benchmarks that highlight the challenges of retrieving time-sensitive information and addressing complex temporal questions. Finally, I will also briefly present our recent work on automatically generating hints to guide user reasoning and mitigate cognitive offloading.

Bio: Adam Jatowt is a Professor in the Department of Computer Science and Deputy Head of the Digital Science Center at the University of Innsbruck. He received his PhD in Information Science and Technology from the University of Tokyo and spent 14 years at Kyoto University. His research interests include natural language processing, information retrieval, knowledge management, and temporal information processing. He is a recipient of the Friedrich Wilhelm Bessel Research Award from the Alexander von Humboldt Foundation and the International Excellence Fellowship from the Karlsruhe Institute of Technology (KIT). He currently serves as an Associate Editor for the ACM Transactions on Information Systems (TOIS) and the Journal of the Association for Information Science and Technology (JASIST). Adam has received Best Paper, Best Short Paper, and Best Demo Paper awards at ECIR, as well as the Vannevar Bush Best Paper Award at JCDL. He has also been a visiting scholar at the University of California, Berkeley, KIT, the University of Toulouse, the National Institute of Advanced Industrial Science and Technology (AIST), the University of Porto, and the University of La Rochelle.

Concurrency Abstraction for Compositional Systems Verification

Par

Arthur Oliveira Vale

Yale University

Jeudi 19 février 2026, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Concurrent and distributed systems are pervasive, yet verifying their correctness remains challenging. A core difficulty is heterogeneity: verification techniques developed for one computational model rarely transfer to others. In this talk, I present compositional linearizability, a framework that reconstructs linearizability through a compositional lens. This perspective yields a general theory from which correctness criteria for specific domains can be systematically derived, as demonstrated in work on crash-aware systems and systems with liveness requirements. The framework leads to novel verification techniques that enable modular reasoning about complex concurrent objects, mechanized in the Rocq proof assistant. These results open promising new paths toward trustworthy distributed systems.

Bio: Arthur Oliveira Vale is a PhD candidate in Computer Science at Yale University, where he is advised by Zhong Shao. His doctoral work develops new foundations and techniques for stating and verifying correctness of concurrent systems, aiming toward trustworthy distributed systems. His work has appeared at POPL, in the Journal of the ACM, and at OOPSLA.

Peut-on produire du sens sans référence ? Ancrage référentiel et autonomie symbolique des grands modèles de langue

Par

Thierry Poibeau

CNRS

lundi 30 mars 2026, 15:30-16:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Résumé: Les récents progrès autour des grands modèles de langue (LLMs) posent des défis conceptuels majeurs aux théories traditionnelles du langage, de la référence et du sens. Ce séminaire examinera la manière dont ces systèmes, entraînés exclusivement sur de vastes corpus textuels, sans perception sensorielle ni contact direct avec le monde, parviennent néanmoins à produire des énoncés linguistiquement cohérents et souvent informatifs. Cette situation semble contredire l’idée selon laquelle le sens nécessite un rattachement causal ou perceptuel aux entités du monde. 

Le séminaire proposé examinera cette question : peut-on rendre compte du sens linguistique sans ancrage (grounding) référentiel classique ? Nous verrons que les LLMs incarnent une forme de linguistique latente où la compétence linguistique émerge de structures internes, de corrélation et de contexte plutôt que d’une référence directe à des objets du monde. Cette perspective rapproche le fonctionnement des LLMs des approches structuralistes du sens et invite à repenser la relation entre structure linguistique, usage discursif et référence.  

En confrontant ces pistes avec les débats contemporains sur la nécessité d’un ancrage sensorimoteur pour la compréhension, ce séminaire interrogera les frontières entre linguistique, représentation du monde et compréhension. Il s’agit non seulement de saisir comment les modèles génèrent du sens, mais aussi de mesurer les implications théoriques et épistémologiques pour les sciences du langage, l’intelligence artificielle et la philosophie du langage.

Ce séminaire sera en partie fondé sur les chapitres 1 et 2 du livre « Understanding Conversational AI : Philosophy, Ethics, and Social Impact of Large Language Models »  (Ubiquity Press, 2025, open access : https://ubiquitypress.com/books/m/10.5334/bde)

Bio: Thierry Poibeau est Directeur de recherche au CNRS, au sein du laboratoire LATTICE (Langues, Textes, Traitement Informatique et Cognition) de l'École normale supérieure. Il est également titulaire d’une chaire PRAIRIE-PSAI (Paris Artificial Intelligence Research Institute–Paris School of Artificial Intelligence). Ses travaux portent sur le traitement automatique des langues, les humanités numériques et l’impact de l’intelligence artificielle sur la société.

Designing Expressive Effect Systems

Par

Matthew Lutze

Aarhus University

Lundi 16 février 2026, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: An effect system is a way for a programming language to track the possible actions taken by a program before the program is ever run. Information about effects is important for both security and optimization of code, as well as supporting programmers through built-in documentation. In this talk, I will introduce the concept of effect systems through the lens of the Flix programming language. I will cover the unique features of Flix's effect system, including effect exclusion (ICFP 2023) and associated effects (PLDI 2024), and then discuss some ongoing and future work in the domain.

Bio: Matthew Lutze is a postdoctoral researcher in the Programming Languages group at Aarhus University. Before earning his PhD at Aarhus in 2026, he earned his master's degree at Université Paris Cité in 2022. His research has covered several topics in programming language design and implementation, including effect systems, type inference, set-theoretic typing, and monomorphization, and has been published at top programming language venues including PLDI, POPL, ICFP, and OOPSLA. As a principal contributor to the Flix programming language, he has leveraged his research to implement several features for Flix's type and effect system.

Implémentation systématique de compilateurs optimisants pour les langages hautement dynamiques

Par

Olivier Melançon

Université de Montréal

Vendredi 13 février 2026, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Les langages de programmation hautement dynamiques, tels que Python, JavaScript, Ruby ou R, doivent leur succès à leur grande expressivité, à leur flexibilité et à la rapidité de développement qu’ils offrent, et ont joué un rôle important dans la démocratisation de la programmation au cours des dernières décennies. Toutefois, cette expressivité est généralement associée à des problèmes de performance. En pratique, obtenir des performances compétitives exige des efforts considérables : des compilateurs complexes et des millions d’heures-personnes de développement. Il existe donc une tension apparente entre le confort des utilisateurs d’un langage et celui de ses implémenteurs. Cette présentation questionnera le caractère inévitable de ce compromis. On y présentera des approches permettant de simplifier et de systématiser l’implémentation d’un langage dynamique en dérivant un compilateur optimisant à partir d’une sémantique exécutable et de techniques d’interprétation abstraite. On y abordera également les perspectives futures de ces approches pour la conception de la prochaine génération de langages de programmation.

Bio: Olivier Melançon est candidat au doctorat à l’Université de Montréal, et membre du Laboratoire de Traitement Parallèle de l’Université de Montréal et du groupe de recherche Compilation et Optimisation des Langages Dynamiques du Centre Inria d’Université Côte d’Azur. Chercheur en implémentation et optimisation des langages de programmation, ses recherches courantes se consacrent aux techniques d’interprétation abstraite pour l’optimisation des langages dynamiques, à la génération de compilateurs à partir de sémantiques exécutables, et aux plateformes d’enseignement de l’informatique.

Synthesizing Quantum Compilers

Par

Aws Albarghouthi

University of Wisconsin-Madison

Lundi 9 février 2026, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: The promise of quantum computing has tantalized researchers for decades, and recent breakthroughs in physical implementations have brought this technology closer to reality. However, the quantum computing landscape remains highly dynamic: competing physical substrates, fault tolerance schemes, and architectures continue to emerge with no clear frontrunner. This diversity creates a significant bottleneck in the compilation pipeline – developing and maintaining separate compilers for each new device or experimental setup is both time-consuming and error-prone.

In this talk, I will present an alternative approach: automatically synthesizing device-specific quantum circuit compilers. This synthesis-based methodology enables rapid iteration while maintaining correctness guarantees. I will demonstrate how automatically synthesized compilers can achieve superior performance compared to sophisticated hand-crafted alternatives.

Bio: Aws Albarghouthi is an associate professor of computer science at the University of Wisconsin-Madison. He works in the field of programming languages and formal methods, where he develops new techniques for automatically building reliable and secure software systems. He received his PhD from the University of Toronto in 2015. He has received several paper awards for his work (PLDI, FSE, UIST, and FAST), an NSF CAREER award, and multiple awards from industry (Meta, Google, Amazon). He also received the Class of 1955 Teaching Excellence Award, one of the University of Wisconsin’s highest teaching honors.

Towards Principled Hyperparameter Optimization and Algorithm Selection

Par

Dravyansh Sharma

Northwestern University

Mercredi 20 août 2025, 14:00-16:00 EST, Salle 3195

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract:Popular practical approaches for effective hyperparameter optimization include Bayesian optimization and bandit-based approaches come with limited theoretical guarantees. Bayesian optimization, often using Gaussian processes, is effective for expensive evaluations but struggles with high-dimensional spaces and the performance is highly sensitive to the choice of priors and internal parameters. Bandit-based approaches make additional assumptions that capture aspects specific to hyperparameter tuning, including fixed limiting values of arm rewards (Hyperband) or increasing pull-dependent rewards with diminishing returns (rising bandits). A major blind spot in effectively using these techniques is the lack of insights on how the algorithmic performance actually varies with the hyperparameter.

A recent line of theoretically grounded work elevates hyperparameter optimization and algorithm selection to a learning problem in its own right. A growing body of research over the past decade from the learning theory community has successfully analysed how to provably tune several fundamental algorithms including decision trees, linear regression, and very recently even deep learning. The new techniques apply naturally to both hyperparameter tuning and algorithm selection. Future research areas include integration of these structure-aware principled approaches with the currently used techniques, better optimization in high-dimensional and discrete spaces, and improving scalability in distributed settings.

The content of this talk is related to a UAI 2025 tutorial and an upcoming NeurIPS 2025 tutorial.

Bio: Dravyansh (Dravy) Sharma is an IDEAL postdoctoral researcher, hosted by Avrim Blum at TTIC and Aravindan Vijayaraghavan at Northwestern University. He obtained his PhD at Carnegie Mellon University, advised by Nina Balcan. His research interests include machine learning theory and algorithms, with a focus on provable hyperparameter tuning, adversarial robustness, and learning in the presence of rational agents. His work develops principled techniques for tuning fundamental machine learning algorithms to domain-specific data, including decision trees, linear regression, graph-based learning and, most recently, deep networks. He has published several papers at top ML venues, including NeurIPS, ICML, COLT, JMLR, AISTATS, UAI and AAAI, has multiple papers awarded with Oral presentations, won the Outstanding Student Paper Award at UAI 2024, and has interned with Google Research and Microsoft Research. He has presented a tutorial at UAI 2025, has an accepted tutorial at AutoML 2025 and an accepted joint tutorial at NeurIPS 2025.

Towards Semantic Versioning of Foundation Models

Par

Bram Adams

Université de Queen

Jeudi 26 juin 2025, 14:00-16:00 EST, Salle 3195

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Developers of FMware, i.e., software products involving foundation models (FMs), are experiencing an explosion of model variants and versions to work and cope with, both closed and open source, because of competing model architectures, optimization of model size to available hardware resources, and alignment to downstream tasks. Organizations eager to integrate these FMs into innovative FMware need to select the most reliable model for their specific use case or determine any backwards compatibility issues with the currently used model version merely based on the models' names, marketing info and (if available) online documentation and performance data. While software engineers developed and adopted semantic versioning practices to deal with such challenges, these practices hardly exist for the different components of FMware (models, prompts, model chains and agents), forcing FMware organizations to resort to trial-and-error and (educated) guesswork of the impact of new FMware component versions. This talk presents empirical insights obtained by mining thousands of open-source FMs, prompts, leaderboards and Model Context Protocol (MCP) servers, trying to understand the challenges faced by FMware stakeholders, as well as identifying opportunities for more sustainable FM model versioning and evolution practices.

Bio: Bram Adams is a full professor at Queen's University (until June 2020: Polytechnique Montreal). He obtained his PhD in 2008 at Ghent University's GH-SEL lab (Belgium). His work on software release engineering and software analytics received the 2021 Mining Software Repositories Foundational Contribution Award, and has been published at premier software engineering venues such as ICSE, FSE, MSR, ICSME, EMSE and TSE. In addition to co-organizing the RELENG International Workshop on Release Engineering from 2013 to 2015 (and the 1st/2nd IEEE Software Special Issue on Release Engineering), he co-organized, amongst others, SEMLA 2018/19, FM+SE Vision 2030 and AIware 2024. He has been PC co-chair of SCAM 2013, SANER 2015, ICSME 2016 and MSR 2019; ICSE 2023 software analytics area co-chair, and general chair of MSR 2025. He is a Senior IEEE Member.

Monotone computations, bracket formulas, and liftings theorems

Par

Dmitry Sokolov

École Polytechnique Fédérale de Lausanne

Mercredi 23 avril 2025, 10:30-11:30 EST

EN LIGNE -  Zoom

Abstract: Suppose some function can be computed in time T on one computer. Can we parallelize the computational process? Namely, can we compute our function in time T/K if we have access to K computers? In this talk, we consider ideas behind one of my recent results that informally can be stated in the following way: "monotone computations cannot be efficiently parallelized". We discuss the result itself and the machinery behind it, which involves several areas: proof, circuit, and communication complexity.

Bio: In 2015 Dmitry completed his Ph.D. at Steklov Institute of Mathematics at St. Petersburg in Russia under the supervision of Edward Hirsch and Dmitry Itsykson. In 2017 he was a postdoc at KTH University in Stockholm hosted by Jakob Nordstrom, a postdoc at Lund University and a visitor at University of Copenhagen. In 2020 he got an Associate Professor position at St. Petersburg State University in Russia. In 2022 he moved to EPFL as a researcher. His research interests are concentrated in complexity theory, discrete mathematics, and math logic. It mainly focuses on proof, communication, and circuit complexity.

Advancing 3D Geometry Processing through Geometric Prior

Par

Jing Ren

École polytechnique fédérale de Zurich

Vendredi 21 mars 2025, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: In the realm of 3D geometry processing, geometric priors serve as foundational constraints that are essential for shape modeling, analysis, and manipulation. However, formulating these priors is highly application-dependent, requiring domain-specific knowledge and expertise. This presentation explores techniques for formulating geometric priors across diverse tasks including shape matching, urban reconstruction, and digital fabrication. Specifically, in the context of shape matching—where the goal is to establish accurate correspondences between shapes undergoing non-isometric deformation—geometric priors play a crucial role. They help define criteria for high-quality mappings, resolve ambiguities arising from shape symmetries, and guide the optimization process to identify optimal correspondences. Additionally, encoding planarity as a geometric prior can significantly enhance the task of 3D roof modeling, enabling efficient and intuitive modeling and reconstruction. Similarly, incorporating structural information from stitching as a geometric prior can contribute to the simulation of intricate embroidered folds, paving the way for interactive design in digital fabrication processes. In this talk, I will discuss the formulation of various geometric priors and demonstrate how they enhance modeling and optimization across different applications in geometry processing.

Bio: Jing Ren is currently a senior researcher in the Interactive Geometry Lab, ETH Zurich, advised by Prof. Olga Sorkine-Hornung. She obtained her Ph.D. degree in 2021 from the Visual Computing Center, KAUST, supervised by Prof. Peter Wonka and Prof. Maks Ovsjanikov. Before that, she obtained the M.Sc. degree from Oxford University, and the B.Sc. degree from Zhejiang University. Her research focuses on shape analysis, geometry processing and digital fabrication, currently with a strong emphasis on bridging theoretical advancements with practical applications in design and manufacturing.

Connecting Proofs and Algorithms

Par

Noah Fleming

Memorial University

Mardi 18 mars 2025, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Over the past several decades an exciting interplay has emerged between algorithms and the provability of mathematical statements, whereby short proofs give rise to efficient algorithms and vice versa. This has led to state-of-the-art algorithms for many NP-hard problems. As well, it has enabled proof complexity — the study of what can be proven efficiently — to become a powerful tool for analyzing algorithms in both theory and practice. This has resulted in provable guarantees for a large number of important classes of practical algorithms including SAT solvers, integer programming solvers, and certain convex programs. These connections have also had an outsized impact on proof complexity, with tools from algorithms resolving a large number of important open questions in this area.

In this talk I will survey this interplay and present a unifying framework for much of it. I will describe how my co-authors and I have used it in order to resolve important open problems in a wide variety of areas; these include:

This talk is aimed at a broad computer science audience, and only a basic familiarity with complexity theory (P, NP, etc.) will be assumed.

The Complexity of Asynchronous Algorithms with Bounded-Size Objects

Par

Sean Ovens

Université de Waterloo

Mercredi 12 mars 2025, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: In a shared memory system, a set of processes communicate by applying operations to a set of shared objects. When designing shared memory algorithms, it is often convenient to assume that the shared objects are infinitely large. However, this is not a practical assumption, as real systems are constrained to use shared objects with constant domain sizes. My work studies the capabilities and limitations of shared memory systems with bounded-size objects. Specifically, I seek to prove new complexity lower bounds in such systems.

In this talk, I will introduce a general technique for obtaining space complexity lower bounds in systems with bounded-size objects. In particular, I will apply this technique to the well-studied consensus problem. To solve consensus, processes begin with private input values and must agree on a common output value, which is equal to one of the inputs. It was previously known that solving obstruction-free consensus using swap objects among n processes requires at least Ω(√n) objects. My work was the first in nearly thirty years to improve upon this lower bound, showing that Ω(n/b) swap objects with domain size b are needed. I will conclude my talk by highlighting a few longstanding open questions about the complexity of shared memory algorithms that could be easier to approach in a system with bounded-size shared objects.

Bio: Sean Ovens is a postdoctoral researcher at the University of Waterloo working with Trevor Brown. He earned his PhD in Summer 2023 from the University of Toronto, where he was supervised by Faith Ellen. His research interests include impossibility results and complexity lower bounds for distributed algorithms, concurrent data structures, and performance profilers and visualizations for multithreaded applications. Sean's work has been published in the Journal of the ACM, DISC, and PODC, where he received best paper awards in both 2022 and 2024.

From Geometry Processing to Topological Defects and Beyond

Par

David Palmer

Harvard University

Mercredi 26 février 2025, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Practical problems from computer graphics, computer vision, and computational engineering reveal surprising connections to the physics of crystals, knot theory, minimal surfaces, and algebraic geometry. These mathematical tools help us devise more robust and efficient algorithms. Conversely, computational exploration with these tools can provide mathematical insight and elucidate new theoretical questions. Covering applications in meshing, neural surface representation, medical imaging, and biology, I will lay out a path forward in interdisciplinary applied geometry.

Bio: David is an NSF postdoctoral fellow at Harvard, working at the intersection of applied geometry, optimization, and physics. He completed his PhD in computer science at MIT, in the lab of Justin Solomon. He is grateful to his exceptional collaborators as well as for the support of the Hertz and MathWorks Fellowships.

Par

Tommaso Toffoli

Résumé:

Cette conférence aborde les aspects humains et scientifiques de ma carrière et du monde en constante évolution qui l'entoure. Ce ne sera pas une conférence conventionnelle (formules au tableau noir et tout le reste), car je raconterai une histoire, dans laquelle il y aura de nombreuses personnes (dont moi), des lieux et des époques différentes, de nombreux thèmes scientifiques, des rencontres et des collaborations fascinantes, ainsi que des potins et des aventures (dont certains conviendraient peut-être mieux à l'occasion d'un bon repas bien arrosé).

Biographie :

Le professeur Toffoli a commencé sa carrière en tant que physicien expérimental, avec la conception (et la fabrication) d'un détecteur à grand angle (1/8 de sphère) de mésons cosmiques. Arrivé aux États-Unis grâce à une bourse Fulbright en 1969, il a obtenu un doctorat en communication et en informatique à l'Université du Michigan. Il y poursuit des recherches sur les automates cellulaires, qui l'amènent rapidement à se tourner vers l'idée du calcul réversible, en même temps que Charles H. Bennett et Ed Fredkin.


Il a passé 20 ans au MIT Lab for Computer Science, d'abord comme chercheur puis comme directeur de recherche. En 1981, en collaboration avec Ed Fredkin et Rolf Landauer, il a organisé la conférence d'Endicott House sur « la physique et le calcul », qui a rassemblé plusieurs scientifiques de renom, dont Richard Feynman, Archibald Wheeler, Freeman Dyson et Carl Petri, et où Feynman a proposé l'idée de l’ordinateur quantique. Avec Norman Margolus, il a conçu une série d'automates cellulaires de plus en plus performants, qu'ils ont utilisés pour créer des « mondes artificiels » de toutes sortes et étudier leur « physique ».


Il a ensuite passé 20 ans au département d'électricité et d'informatique de l'Université de Boston. Aujourd'hui, il occupe un poste honoraire au département de physique de l'Université de Boston et rédige un livre sur l'évolution en tant que mécanisme de nature générale, dont la vie biologique est un exemple canonique.

Intervenants :

Etienne Elie (Y Square, Accelerate USA)
Olivier Milon (Hydro Québec)

Résumé:

La réindustrialisation de l'Amérique du Nord fait face à des défis tels que la pénurie de main-d'œuvre qualifiée, la compétitivité mondiale et la nécessité de durabilité. La recherche opérationnelle offre des solutions concrètes grâce à une approche multimodèle qui optimise l'allocation des ressources, la formation des travailleurs, l'automatisation des processus et la gestion des chaînes d'approvisionnement. En intégrant des outils d'analyse avancés et des méthodologies adaptées, la recherche opérationnelle permet de bâtir une industrie moderne, productive et durable, capable de répondre aux besoins économiques et environnementaux actuels. Un survol de l'application de la recherche operationnelle dans l'industrie technologique, en particulier dans le contexte de la Silicon Valley, sera présenté, ainsi que les impacts sur la consommation énergétique.

La prévision de la demande chez Hydro Québec est un service d'aide à la décision technico-économique puisqu'il permet de maximiser les opportunités d'exportation tout en minimisant les risques des mouvements d'énergie. Il s'inscrit donc à la fois dans le cadre de la conformité NERC des systèmes énergétiques vis-à-vis de leur fiabilité et aussi dans la dynamique de marché nord est américain de l'électricité. S'il est possible de trouver un lien de causalité entre la météorologie et la charge consommée par les clients alors on peut construire un modèle d'inférence d'une prévision de la demande de puissance en électricité au Québec à partir de la prévision météorologique aux stations. Mais il faut également prendre en compte le signal comportemental du secteur domiciliaire et celui stratégique d'affaire du secteur industriel pour rencontrer des critères de performance opérationnelles de ces prévisions. Le projet Amélioration de Prévision de la Demande vise à concrétiser ce service dans un système infonuagique innovant, offrant une grande capacité de calculs fondés sur des approches neuronales pour traiter des données en temps réel. L'état des travaux illustrés de résultats numériques sera présenté.

Les exposés seront suvis d'une discussion-débat entre les intervenants et l'auditoire.

L'événement sera tenu en français.

Quantum annealing: an overview of old and new results

Par

Andrew King

D-Wave Quantum Systems Inc

Jeudi 28 novembre 2024, 15:30-16:30 EST, Salle 5340

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Quantum annealing (QA) is a heuristic optimization method which is analogous to simulated annealing (SA), but which involves tuning of quantum fluctuations in the place of thermal fluctuations (temperature).  In this talk I will discuss the early origins of QA, which motivated the development of the first programmable QA processors.  The first benchmarking tests of QA led to questions and controversies, and I will discuss these and the subsequent progress in characterizing and refining QA performance.  Finally, I will cover the recent state of the art in QA for simulation and optimization as demonstrated by D-Wave, IBM, QuEra, and others.

Hidden Capabilities and Counterintuitive Limits in Large Language Models

Par

Peter West

University of Washington

mercredi 27 mars 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Massive scale has been a recent winning recipe in natural language processing and AI, with extreme-scale language models like GPT-4 receiving most attention. This is in spite of staggering energy and monetary costs, and further, the continuing struggle of even the largest models with concepts such as compositional problem solving and linguistic ambiguity. In this talk, I will propose my vision for a research landscape where compact language models share the forefront with extreme scale models, working in concert with many pieces besides scale, such as algorithms, knowledge, information theory, and more.

The first part of my talk will cover alternative ingredients to scale, including (1) an inference-time algorithm that combines language models with elements of discrete search and information theory and (2) a method for transferring useful knowledge from extreme-scale to compact language models with synthetically generated data. Next, I will discuss counterintuitive disparities in the capabilities of even extreme-scale models, which can meet or exceed human performance in some complex tasks while trailing behind humans in what seem to be much simpler tasks. Finally, I will discuss implications and next steps in scale-alternative methods.

Designing Timing Predictable and High-Performance Compute Systems for Cyber-Physical Systems

Par

Anirudh Kaushik

University of Waterloo

lundi 25 mars 2024, 10:30-11:30 EST, Salle 3195

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Cyber-physical systems (CPS) are interconnected systems of sensors and computing systems that sense and collect information about the physical world, and in turn operate on this sensed information to interact with and influence the physical world. Application domains of CPS include healthcare, agriculture, transportation, and aerospace. As we continue to entrust more parts of our lives to CPS, it is imperative that the underlying hardware and software computing systems on which these CPS rely on are designed correctly for safe operation. In safety-critical CPS domains, producing correct values (logical correctness) and the time it takes to produce the correct values (timing predictability) constitute as correct design. This notion of time as a first-class design principle for computing system design introduces challenges to conventional design processes and methodologies.

In this talk, I will discuss my research in designing timing-predictable and high-performance hardware computing systems for safety-critical CPS. Specifically, I will highlight the challenges and my research contributions towards facilitating timing-predictable coherent shared data communication in multi-core compute systems. I will discuss my future research vision on developing efficient compute systems for CPS, which will be crucial to sustaining the tremendous potential benefits that CPS offer and accelerating their deployment and outreach in our lives and society at large.

Designing Adaptable Interfaces that Support Scientific Communication

Par

Damien Masson

University of Toronto

vendredi 22 mars 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Interfaces to consume and produce scientific information are typically designed as one-size-fits-all: regardless of people's preferences, a scientific article forces a single presentation, and a tool forces a single mode of operation.  In this talk, I argue that this idea of people adapting to interfaces should be replaced by interfaces adapting to people, their expertise, preferences, and tasks. Towards this vision, I describe my efforts designing and evaluating "adaptable interfaces" to support scientific communication. Specifically, I demonstrate adaptable interfaces that help readers better understand scientific material, scientists interactively present their findings, and authors more effectively create visual and textual content. Informed by these projects and cognitive theories, I conclude with a framework to inform the design of future adaptable interfaces.

Quantum functional programming: A new paradigm for quantum computing with implications for machine learning and materials science

Par

Hlér Kristjansson

University of Waterloo

mercredi 13 mars 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Quantum computing has been promised to enable significant speed-up in certain computational problems and has the potential to revolutionise the simulation of physical systems, with implications in fields ranging from machine learning to materials discovery. So far, the majority of quantum algorithms are based on constructing a circuit of quantum logic gates, which is used to transform the state of a quantum system. Yet, we know from classical computing that not only states (or in other words, variables) can be transformed; functions of those states can themselves be used as inputs of transformations, enabling the modular concatenation of functions of functions. In this talk, I will introduce a recent paradigm for the quantum analogue of functional programming, known as higher-order quantum computation. I will present my results on how this paradigm can be used to develop protocols for improving quantum communication, as well as new types of algorithms for the simulation of physical systems, with implications in materials discovery. Finally, I will discuss future directions of this approach, including how the same functional programming techniques could be used to build a fully quantum generalisation of neural networks in machine learning. This talk is aimed at a general computer science audience, and no prior knowledge of quantum information is assumed.

Internet-Scale Consensus in the Blockchain Era

Par

Joachim Neu

Stanford University

lundi 11 mars 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Blockchains have ignited interest in Internet-scale consensus as a vital building block for decentralized applications and services that promise egalitarian access and robustness to faults and abuse. While the study of consensus has a 40+ year tradition, the new Internet-scale setting requires a fundamental rethinking of models, desiderata, and protocols. An emergent key challenge is to simultaneously serve clients with different requirements regarding the two fundamental aspects of security, liveness ("good things happen") and safety ("bad things don't happen"). For different instances of this theme, I explore the optimal liveness-safety tradeoff, and present the first protocols that achieve it. Results from this line of work have found adoption in the Ethereum blockchain that powers an ecosystem worth $500bn+.

Human-centered Natural Language Processing

Par

Ines Arous

Université de Montréal

jeudi 7 mars 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Natural Language Processing (NLP) has entered the public consciousness with widespread popularity, through platforms such as chatGPT. While responses from such a system are coherent, it may rely on inaccurate evidence, magnify ingrained biases, and lead to harmful outcomes. This alarming reality forces us to confront a fundamental research question: Can we rewire the very fabric of modern NLP models to align with human values? To tackle this question, my research focuses on a comprehensive reevaluation of the entire NLP pipeline. I posit that there is no one-size-fits-all solution to rectify the biases and ethical dilemmas plaguing NLP. Instead, I propose human-AI collaborative approaches where humans actively participate in all NLP pipeline stages from data annotation to model training and evaluation. In this talk, I will describe the steps to address this goal. First, I introduce a method integrating human annotation into the model's training process. Next, I explore enhancing models’ explainability by leveraging humans' rationale and evaluating them according to domain-specific requirements. Finally, I discuss the research opportunities for developing explainable and trustworthy NLP.

Algorithms, complexity, and quantum many-body physics

Par

Cunlu Zhou

University of New Mexico

mercredi 28 février 2024, 10:30-11:30 EST, Salle 3195

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: In this talk, I will discuss three results about algorithms, complexity, and quantum many-body physics. The first one is about the so-called Hamiltonian Variational Ansatz, which is used in Variational Quantum Algorithms (VQAs) for approximating the ground states of condensed matter physics models. The second one is about a Quantum Phase Estimation (QPE) algorithm based on compressed sensing that is suitable for early fault-tolerant quantum computers. The last one is about an SU(2) symmetric semidefinite programming (SDP) hierarchy for the so-called Quantum MaxCut problem, which connects the Heisenberg model in condensed matter physics with optimization algorithms and fundamental computational complexity questions such as the so-called Quantum PCP Conjecture. The talk is going to be rather high level, and no special background in quantum computation or theoretical computer science is assumed.

Machine learning for science: tackling climate and health challenges

Par

Alex Hernandez-Garcia

Université de Montréal

lundi 26 février 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Science plays a fundamental role in tackling the most pressing challenges for humanity, such as the climate crisis and the threat of pandemics or antibiotic resistance. Meanwhile, the increasing capacity to generate large amounts of data, the progress in computer engineering and the maturity of machine learning methods offer an excellent opportunity to be put at the service of scientific progress. In this talk, I will present an overview of my postdoctoral work during the last three years on machine learning research with an impact on climate and health: materials discovery, molecular modelling, biological sequence design, climate modelling and climate impacts visualisation. In particular, I will focus on my work on machine learning to accelerate scientific discoveries. I will present our recent algorithm for multi-fidelity active learning with GFlowNets, designed to efficiently explore combinatorially large, high-dimensional and mixed spaces (discrete and continuous), inspired by challenges in materials and drug discovery. Finally, I will discuss in more detail the challenging but impactful case of crystal structure generation. I will offer an introduction to GFlowNets and explain how we have adapted this method to incorporate domain knowledge from crystallography, physics and chemistry in the form of hard constraints, to efficiently discover new materials with desirable properties. I will conclude with a discussion of remaining challenges and my research plans to tackle them.

Safe and Responsible AI Development Via Climate Applications

Par

Tegan Rajkumar-Maharaj

Schwartz-Reisman Institute for Technology and Society, U. of Toronto

vendredi 23 février 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Machine learning (ML)-based artificial intelligence (AI) systems are increasingly deployed in real-world settings, but we lack a rigorous science to understand or predict their behavior in these settings.  Even when we can formalize the problem we’re addressing in quite clear statistical terms (e.g. supervised learning on a fixed dataset), there is much we still do not understand about how and why deep nets are able to generalize as well as they do, why they fail when they do, how they will perform on out-of-distribution data. This complicates the already-difficult problem of designing safe and responsible methods and norms of AI development -- where do we even begin?  My answer to that question is to begin by helping to address our ongoing climate crisis. There is no one magic solution to all the problems of climate change, but there are ways almost every form of knowledge and expertise (including ML) can help. And developing safe and responsible AI via important real-world problems ensures that our methodology is grounded and practically useful by design. In this talk I give an overview of my research at the intersection of responsible AI and climate change, with a focus on recent work in deep risk mapping.

Understanding and Aligning Large Scale Deep Learning

Par

David Krueger

University of Cambridge

mercredi 21 février 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Large scale deep learning systems are poised to have a transformative impact on society, and may pose an existential risk to humanity.  The safety of these systems depends on our ability to understand them and harness their capabilities.  I will talk about my work on methods for doing this, and the limitations of existing approaches.  I will cover my recent and ongoing work on alignment failure modes, and the limitations of fine-tuning (such as reinforcement learning from human feedback) as an alignment approach.  A key theme is the need to better understand how deep learning systems learn and generalize, in order to predict and steer their behavior.  I will also discuss how my technical work informs and is informed by AI governance.

Flow models with applications to cell and molecular biology

Par

Alexander Tong

Université de Montréal

vendredi 16 février 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Generative flow models learn a (possibly stochastic) mapping between source and target distributions. Common paradigms include diffusion models, score matching models, and continuous normalizing flows. In this talk I will first present methods for improved training of flow models using flow matching objectives using ideas from optimal transport. I will then show how these improved methods can be applied to the tasks of (1) modelling cell dynamics, which allow us to better understand disease programs – leading to a new potential therapeutic pathway for triple-negative breast cancer and (2) generative protein design, with applications to biologic drug discovery.

A Bayesian Perspective on Causal Discovery

Par

Tristan Deleu

Valence Labs and Université de Montréal

mercredi 14 février 2024, 10:30-11:30 EST, Salle 6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Discovering the structure of a causal model purely from data is plagued with problems of identifiability. In general, shy of any assumptions about how the data was generated, multiple equivalent models may explain our observations equally well even if they could entail widely different causal conclusions. As a consequence, choosing an arbitrary element among these equivalent models may result in making unsafe decisions if it is not aligned with how the world truly works. It is therefore essential to keep a notion of epistemic uncertainty about our possible candidates in order to mitigate the risks posed by these misaligned models, especially when the data is limited. In this talk, I will introduce a new class of probabilistic models called Generative Flow Networks (GFlowNets) that provides a general framework for modeling probability distributions over discrete and compositional objects, such as the structure of a causal model, which is represented as a directed acyclic graph (DAG). GFlowNets treat generation as a sequential decision making problem, where each sample is constructed piece by piece. I will highlight how they connect to various domains of machine learning and statistics, including variational inference and reinforcement learning. Finally, I will show how GFlowNets can be used for causal discovery from a Bayesian perspective, to model the whole posterior distribution over causal models with arbitrary mechanisms, given a dataset of observations.

Democracy and the Pursuit of Randomness

Par

Ariel Procaccia

Harvard University, É-U

jeudi 7 décembre 2023, 14:00-15:00 EST Salle AA-6214

Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour

Abstract: Sortition is a storied paradigm of democracy built on the idea of choosing representatives through lotteries instead of elections. In recent years this idea has found renewed popularity in the form of citizens’ assemblies, which bring together randomly selected people from all walks of life to discuss key questions and deliver policy recommendations. A principled approach to sortition, however, must resolve the tension between two competing requirements: that the demographic composition of citizens’ assemblies reflect the general population and that every person be given a fair chance (literally) to participate. I will describe our work on designing, analyzing and implementing randomized participant selection algorithms that balance these two requirements. I will also discuss practical challenges in sortition based on experience with the adoption and deployment of our open-source system, Panelot.