【Team】 Imperfect Information Learning Team
【Date】2026/June/24(Wednesday) 15:00-16:00(JST)
【Speaker】Talk by Hanlin Yu, University of Helsinki
Title: Comparing Distributions and Representations: Geometry, Density Ratios, and Energy-Based Models
Abstract:
Many machine learning problems can be understood as comparing probability distributions, model representations, or unnormalized energies. In this talk, I will present principled and scalable tools for such comparisons. I will first discuss how Riemannian geometry provides a natural local comparison operator in probabilistic models, latent representations of neural networks and empirical data manifolds. I will then turn to density ratio estimation, where the goal is to directly learn the differences in densities from empirical samples. In order to solve density ratio estimation through infinitesimal classification along probability paths, I will introduce Conditional Time Score Matching and its vectorized variant, allowing fast and scalable estimation of such ratios. Finally, I will discuss how to learn Energy-based Models using spatialtemporal differences. Through comparing across space and time, the approach unifies existing temporal and spatial algorithms while avoiding their failure modes. Together, these results suggest a unified perspective, where robust machine learning can be achieved by constructing reasonable comparisons.
