Constant Bourdrez : "Learning To Sample From Diffusion Models Via Inverse Reinforcement Learning" Hanna Benarroch : "Certified Per-instance Unlearning using Individual Sensitivity Bounds"

Emmanuel Sérié: Harness over harnesses — demo of cosmon orchestrating Claude Code agents (noogram.org) Olivier Cappé: demo of Antigravity for doing useful things Gabriel Peyré: demo of Codex for making video games and doing math

Solving most image processing tasks, such as denoising, deblurring, inpainting, etc, require explicitly or implicitly a prior probability model of natural images. Classically, both learning (by maximizing likelihood) and using (via Bayes' rule) such prior models are intractable due to the curse of dimensionality. Diffusion models take a different approach, where the prior density is replaced by a family of score vector fields across noise levels. They have led to impressive success in generative modeling, but the learned density is not explicit nor is it readily usable as a prior […]

Understanding how deep neural networks learn useful internal representations from data remains a central open problem in the theory of deep learning. We introduce Neural Low-Degree Filtering (Neural LoFi), a stylized limit of gradient-based training in which hierarchical feature learning becomes an explicit iterative spectral procedure. In this limit, the dynamics at each layer decouple: given the current representation, the next layer selects directions with maximal accessible low-degree correlation to the label. This yields a tractable surrogate mechanism for deep learning, together with a natural kernel-space interpretation. Neural LoFi provides […]

Foundation models are a key platform which implicitly encodes world knowledge. In this talk we first focus on vision-language models, such as CLIP, and investigate their geometric behavior and logic behind the high-dimensional feature encoding. For instance, we find that as image or text become more rare and distinct they are encoded further from the center of the embedding. We explain why InfoNCE loss leads to that behavior. We also find out empirically that each modality can be well modeled statistically as admitting a multivariate Gaussian distribution. This finding is […]