Information Theory in Modern Science 2026, Okinawa, Japan

Title: Geometric information theory: A hub to information sciences

Abstract: In this talk, I will first present an overview of geometric information theory emphasizing that fundamental quantities such as the Kullback-Leibler divergence, Shannon entropy, and Fisher information are not merely formulas, but arise from principled geometric structures endowed with duality relations. I will further discuss the deep interplay between information measures and geometry more broadly, and explain the links of their variational characterizations with geodesics. From this perspective, many optimization problems in information theory including inference and model approximation can be interpreted as geometric projection problems on statistical manifolds. In the second part, I will present several recent developments including curved Bregman divergences, duo Bregman pseudo-divergences, generalized convexity/Legendre transforms and information measures, and maximal invariants and divergences. I will illustrate how these  geometric concepts and methods extend classical frameworks and provide new tools and opportunities  in information sciences.

Title: Information theoretical approaches to model synaptic plasticity

Abstract: We adapt our behavior through experience, a process widely believed to depend on activity-dependent synaptic plasticity. In this talk, I present a top-down theoretical framework in which synaptic plasticity rules are understood as optimizing information transmission between neurons. This perspective yields a neuro-inspired algorithm for uncovering hidden independent sources in sensory inputs, extendable to scenarios in which the sources are nonlinearly mixed. The framework also accounts for experimental observations by reproducing distinct plasticity outcomes during the up and down states of non-REM sleep. Together, these findings offer insight into how memory consolidation is shaped by brain state and the spatial scale of slow waves.

Title: Generalization bounds in machine learning, information theory and predictions

Abstract: A key question in machine learning theory is to control the generalization error of learning algorithms, that is, the discrepancy between the accuracy of the algorithm on the data it is trained on, and its ability to predict future data. Many approaches are available. Among them, PAC-Bayes bounds and mutual information bounds are deeply related, as they make use of measure of informations to quantify this gap. In a first time, I will introduce PAC-Bayes bounds and mutual information bounds. I will illustrate on examples how these bounds can be used in practice to guide algorithm training and to perform model selection.

Most PAC-Bayes bounds and information bounds were proven for independent data. In a second time, I will discuss the difficulties encountered when we try to extend these theories to time series prediction. I will introduce a new approach to derive practical bounds for Markov chains, based on the estimation of the pseudo-spectral gap of the chain.

The first part will be based on my lecture notes: User-friendly introduction to PAC-Bayes bounds (2024); the second part is a joint work with Vahe Karagulyan (ESSEC Business School): Empirical PAC-Bayes bounds for Markov chains, to appear at AISTATS 2026.

Title: Irruption Theory: An Information-Theoretic Framework for Mental Efficacy and Stochastic Decision-Making

Abstract: Cognitive neuroscience faces a fundamental measurement problem: the purpose-driven, normative features of the human mind cannot be directly observed in the brain. To address this, "irruption theory" proposes that the dynamical relevance of unobservable mental efficacy manifests indirectly as increased unpredictability - irruptions - quantifiable via information-theoretic entropy. The first part of this talk introduces the conceptual foundation of this framework, modeling mind-body transactions as an informational boundary. The second part demonstrates how this stochastic framework concretely informs current research on decision-making under uncertainty. Specifically, we will explore how memoryless exponential dynamics provide a principled, clock-free approximation to optimal stopping strategies (the 1/e-law), allowing the autonomic nervous system to solve complex temporal tasks through purely state-dependent relaxation.

Title: Theoretical Foundations and Practical Challenges of Diffusion (Generative) Models

Abstract: Diffusion models have achieved impressive performance in generating high-quality and diverse synthetic data. This talk will first review the underlying theory of discrete diffusion models, their relationships to SDEs, and the setting of classifier-free guidance for conditional diffusion models. We will then focus on two emerging challenges for these models: (i) enhancing diffusion models to produce high-fidelity and diverse samples when trained on class-imbalanced data, and (ii) quantifying generalization and memorization of diffusion models when the training data has class-imbalances. We will briefly discussion a solution to this via CORAL (Contrastive Regularized Alignment of Latents), a framework that leverages supervised contrastive losses to encourage well-separated latent class representations and present theoretical results on memorization resulting from training with class-imbalanced data.

Title: TBA

Abstract: TBA

Title: TBA

Abstract: TBA

Title: TBA

Abstract: TBA

Title: TBA

Abstract: TBA

Title: TBA

Abstract: TBA