From information theory to the mathematical foundations of reliable and collaborative intelligence

How can intelligent systems determine what information to acquire, what computations to perform, what evidence to trust, and when to communicate or act?

My research develops mathematical and algorithmic foundations for understanding how intelligent systems learn, reason, evaluate evidence, and make decisions under uncertainty and limited resources.

Building on information theory, statistics, and statistical decision theory, I investigate the fundamental limits of decision-relevant information processing: what information must be represented or acquired to solve a task, when additional computation or interaction provides genuine value, and what guarantees can be established from finite evidence.

This perspective connects three complementary research directions: the mathematical foundations of learning and reasoning; uncertainty, evaluation, and reliable decision-making; and collaborative, resource-aware intelligence.

My long-term objective is to develop a unified theoretical framework in which learning, reasoning, information acquisition, communication, and action are understood as interconnected operations, guided by their contribution to reliable decisions.

The scientific ideas and methodological advances underlying my research. These established contributions complement the ongoing and future questions outlined in the research directions above.

Connecting fundamental research in information theory and machine learning with real-world challenges.

Industrial collaboration Industrial collaboration Industrial collaboration Industrial collaboration Research award Research chair Academic partnership Research programme · Artificial intelligence Research programme · Cybersecurity Research support Research support · Health Research support

Representative papers illustrating these scientific contributions across information theory, statistical learning, and artificial intelligence. Each entry provides the publication venue, a short description, and links to the work.

DOCTOR: A Simple Method for Detecting Misclassification Errors

NeurIPS · 2021

Identifies potentially incorrect classifier predictions, including settings with restricted access to the model.

Paper ↗arXiv ↗

The Role of Mutual Information in Variational Classifiers

Machine Learning · 2023

Examines how mutual information enters variational classification objectives and what it reveals about learned predictive representations. Connects information-theoretic analysis with classifier design and interpretation.

Paper ↗arXiv ↗

A Data-Driven Measure of Relative Uncertainty for Misclassification Detection

ICLR · 2024

Develops a data-driven measure of relative uncertainty for detecting incorrect classifier predictions. Focuses on using uncertainty information to distinguish reliable from potentially erroneous decisions.

Paper ↗arXiv ↗

Beyond the Norms: Detecting Prediction Errors in Regression Models

ICML · 2024

Studies detection of unreliable predictions in regression rather than classification.

Paper ↗arXiv ↗

COSMIC: Mutual Information for Task-Agnostic Summarization Evaluation

ACL · 2024

Uses information-theoretic ideas to evaluate summaries without committing to a single downstream task.

Paper ↗arXiv ↗

When Is an Embedding Model More Promising Than Another?

NeurIPS · 2024

Compares learned representations in terms of their potential usefulness across downstream tasks.

Paper ↗arXiv ↗

Adversarial Robustness Via Fisher-Rao Regularization

IEEE TPAMI · 2023

Studies adversarial robustness through regularization based on Fisher–Rao geometry. Connects the geometry of predictive models with methods for improving their stability under perturbations.

Paper ↗arXiv ↗

Learning Task-Agnostic Representations through Multi-Teacher Distillation

NeurIPS · 2025

Investigates how knowledge from multiple teachers can be distilled into representations that remain useful across tasks. Addresses the challenge of learning reusable features without tailoring them to a single downstream objective.

Paper ↗arXiv ↗

Fundamental Limits of Membership Inference Attacks on Machine Learning Models

JMLR · 2025

Investigates fundamental limits on what membership inference attacks can reveal.

Paper ↗arXiv ↗

Collaborative Rational Speech Act: Pragmatic Reasoning for Multi-Turn Dialog

EMNLP · 2025

Extends pragmatic language modelling to collaborative, multi-turn interactions with private information.

Paper ↗arXiv ↗

Rational Retrieval Acts: Leveraging Pragmatic Reasoning to Improve Sparse Retrieval

SIGIR · 2025

Introduces pragmatic reasoning into sparse retrieval to account for the information conveyed by retrieval terms.

Paper ↗arXiv ↗

MOSAIC: Multiple Observers Spotting AI Content

Findings of ACL · 2025

Investigates the detection of AI-generated content by combining evidence from multiple observers. Explores how complementary signals can support more robust content assessment.

Paper ↗arXiv ↗

How Much Were You Told? Measuring External Information in Peer Reviews

Preprint · 2026

Investigates how much information in a peer review comes from sources beyond the reviewed manuscript. Proposes an information-oriented perspective on measuring external content in scientific evaluations.

Paper ↗

Collaborative Information Bottleneck

IEEE Trans. on Information Theory · 2019

Studies fundamental complexity–relevance trade-offs for distributed sources and cooperative information exchange.

Paper ↗arXiv ↗

The Three-Terminal Interactive Lossy Source Coding Problem

IEEE Trans. on Information Theory · 2017

Investigates the information-theoretic limits of interactive lossy source coding.

Paper ↗arXiv ↗

Secure Multiterminal Source Coding With Side Information at the Eavesdropper

IEEE Trans. on Information Theory · 2013

Studies secure compression of correlated information across multiple terminals when an eavesdropper has side information. Characterizes trade-offs involving communication, reconstruction, and information leakage.

Paper ↗arXiv ↗

A Differential Entropy Estimator for Training Neural Networks

ICML · 2022

Develops a differential-entropy estimation approach designed for use in neural-network training. Connects continuous information measures with optimization objectives in representation learning.

Paper ↗arXiv ↗

MolRGen: A Training and Evaluation Setting for De Novo Molecular Generation with Reasonning Models

Preprint · 2026

Introduces a training and evaluation setting for de novo molecular generation using reasoning models. Examines how molecular generation can be studied through structured tasks and assessment protocols.

Paper ↗arXiv ↗

View all publications →

Current PhD students and researchers

JF

Judeson Fernando Anthony

Uncertainty Estimation in Vision-Based Perception Systems for Autonomous Driving

🇩🇪 Technische Hochschule Ingolstadt

Co-supervised with Torsten Schön

GM

Guillermo Martin

Belief Shaping and Strategic Communication in Multi-Agent AI Systems

🇨🇦 🇫🇷 ÉTS Montréal · Université Paris-Saclay

Co-supervised with Christian Desrosiers · Benedikt Bollig

TH

Thomas Hennequin

AI-based Control Method for Aircraft Flight Assistance

🇨🇦 ÉTS Montréal

Co-supervised with Philippe Feyel

CC

Cléa Chataigner

Human-In-The-Loop Auditing for Generative AI Systems

🇨🇦 🇫🇷 McGill University · Université Paris-Saclay

Co-supervised with Golnoosh Farnadi

AA

Antoine Andurao

Foundation Models for 3D Medical Imaging

🇫🇷 Université Paris-Saclay

Co-supervised with Maria Vakalopoulou

MS

Matteo Sammut

Toward a Theoretical Framework for Auditing Generative AI Models

🇨🇦 🇫🇷 ÉTS Montréal · Université Paris-Dauphine

Co-supervised with Yan Chevaleyre · Rafael Pinot

CR

Chiara Roverato

Diffusion Models for Tabular Data

🇨🇦 McGill University

Co-supervised with Mark Coates

AM

Andrew McDonald

Building Foundation Models for Tabular Data

🇫🇷 Université Paris-Saclay

Co-supervised with Spyros Angelopoulos

LE

Lautaro Estienne

Smart Conversations: Transforming Human-Machine Interaction

🇫🇷 Université Paris-Saclay

Co-supervised with Nona Naderi

ZD

Zahra Dehghani Tafti

Machine Unlearning

🇨🇦 ÉTS Montréal

Co-supervised with Mohammad Hadi Shateri

AS

Arthur Satouf

Information Retrieval

🇫🇷 Université Paris-Saclay

Co-supervised with Habiboulaye Amadou-Boubacar · Benjamin Piwowarski

MD

Matthieu Dubois

Detecting AI Generated Text

🇫🇷 Sorbonne Université

Co-supervised with François Yvon

PF

Philippe Formont

Computational Tools for Molecule Discovery

🇫🇷 🇨🇦 Université Paris-Saclay · ÉTS Montréal

Co-supervised with Ismail Ben Ayed

Former PhD students and postdoctoral researchers (25 alumni)

MD

Maxime DarrinMistral AIPhD · 2025Metrics and Methods for Generative AI Model Evaluation🇫🇷 🇨🇦 Université Paris-Saclay · McGill University

EA

Eric AubinaisKTH Royal Institute of TechnologyPhD · 2025Conformal Learning with Applications to Machine Learning🇫🇷 Université Paris-Saclay

ED

Eduardo Dadalto Camara GomesHelsingPhD · 2024Securing the use of AI🇫🇷 Université Paris-Saclay

FG

Federica GraneseInriaPhD · 2023Information-Theoretic Methods for Safety AI🇫🇷 🇮🇹 Institut Polytechnique de Paris · Università degli Studi di Roma

MB

Malik BoudiafEquallPhD · 2023Information measures for low-shot learning🇨🇦 ÉTS Montréal

MP

Marine PicotDigeizPhD · 2023Artificial Intelligence in Smart Grids🇫🇷 🇨🇦 Université Paris-Saclay · McGill University

GD

Ganesh Del GrossoSituation.shPhD · 2023Leakage of Sensitive Training Data from Deep Neural Networks🇫🇷 École Polytechnique

PC

Pierre ColomboCoherePostdoc · 2022–2023Securing the use of AI🇫🇷 Research location

MH

Mohammad Hadi ShateriÉTS MontréalPhD · 2022Deep Learning and Privacy in Smart Grids🇨🇦 McGill University

MR

Marco RomanelliHofstra UniversityPostdoc · 2020–2022Securing the use of AI🇫🇷 Research location

MV

Matias VeraCONICETPhD · 2020Learning via Information-Bottleneck Methods🇦🇷 Universidad de Buenos Aires

CF

Clément FeutryGroupe BPCEPhD · 2019Two Sides of Relevant Information: Anonymized Representation through Deep Learning and Predictor Monitoring🇫🇷 Université Paris-Saclay

FM

Francisco MessinaUniversidad de Buenos AiresPostdoc · 2018–2020Artificial Intelligence in Smart Grids: Robust Learning🇨🇦 Research location

ZL

Zuxing LiTongji UniversityPostdoc · 2018–2019Distributed Compression and Learning on Power Lines🇫🇷 Research location

KW

Kai WanHuazhong University of Science and TechnologyPhD · 2018Fundamental Limits of Index Coding

GK

Gil KatzTenGivingPhD · 2017Distributed Statistical Inference with Communication Constraints🇫🇷 Research location

MM

Maggie MhannaGooglePhD · 2017Distributed Hypothesis Testing with Privacy Constraints🇫🇷 Research location

GP

Georg PichlerTU WienPhD · 2017Information-Theoretic Clustering🇦🇹 Research location

Chao HeEricssonPhD · 2016Cooperation and Feedback in Wireless Networks🇫🇷 Research location

MB

Meryem BenammarISAE-SUPAEROPhD · 2016Broadcast Channels: Side Information and Security Constraints🇫🇷 Research location

GB

German BassiEricsson ResearchPhD · 2015Interference Relay Networks: Feedback and Security Constraints

AA

Andres AltieriQualcommPhD · 2014Fundamental Limits of Cooperative Large Wireless Networks🇦🇷 🇫🇷 Research locations

JV

Joffrey VillardSSyntelierPhD · 2012Task-Oriented Source Coding: Secure Transmission, Detection🇫🇷 Research location

AB

Arash BehboodiQualcomm AI ResearchPhD · 2012Cooperative Networks with Channel Uncertainty🇫🇷 Research location

PL

Patricia LayecNokia Bell LabsPhD · 2009Information-Theoretic Bounds and Practical Coding Schemes for MIMO Communications Under Imperfect CSIR and Partial and/or Imperfect CSIT🇫🇷 Research location

I develop international research partnerships and collaborative programmes connecting the mathematical foundations of learning and information theory with reliable artificial intelligence. My activities combine institutional leadership, competitive research programmes, and service to the international scientific community.

Montréal, Canada · Paris-Saclay, France

ILLS Executive Assistant — Katia Goyette-Desjardins (she/her)
Secrétaire de direction ILLS
Email: secretariat-ills@etsmtl.ca