הסמינר בנושא בינה מלאכותית יכלול את שלושת הסמינרים הבאים :
Date: Monday, 15.6.26
Place: Building 37, room 202
Time: 11:30-13:00
Speaker: Ariel Avital
Supervisor: Prof. Klim Efremenko and Prof. Aryeh Kontorovich
Date: Monday, 15/6, talk starts at 11:30
Place: Building 37, room 202
Talk title: Total Variation Bounds for Bernoulli Products: From Expert Aggregation to Small-Parameter Regime
Abstract: This talk presents a unified perspective on two recent works concerning total variation distance between Bernoulli product measures. The first part studies the problem of aggregating binary advice from conditionally independent experts in the Naive Bayes setting. The optimal prediction error is expressed through a pointwise-min formula, and, in the balanced-prior case, is directly connected to total variation via Scheffé’s identity. This leads to sharp upper and lower bounds on aggregate expert error in terms of balanced accuracies and log-accuracy parameters. The second part focuses on the intrinsic geometry of total variation between Bernoulli product measures in sparse regimes. When all Bernoulli parameters are tiny, the distance behaves like the L1-distance between the parameter vectors. In the larger small-parameter regime, total variation is controlled, up to constants, by the discrepancy on the singleton Hamming slice. Together, these results show that Bernoulli-product TV exhibits different effective geometries in different regions of parameter space, with consequences for hypothesis testing, expert aggregation, and sparse binary models.
Speaker: Nimrod Berman
Supervisor: Dr. Omri Azencot
Date: Monday, 15/6, talk starts at 12:00
Place: Building 37, room 202
Talk title: Koopman Operator for Disentanglement and Diffusion
Abstract: Koopman operator theory provides a powerful lens for analyzing nonlinear dynamical systems by lifting them into a space where their evolution becomes linear, exposing structure through the operator's spectrum. In this talk I show how this single idea drives progress on two very different deep-learning problems. First, for sequential disentanglement, We present a structured Koopman autoencoder that separates the latent factors of a sequence into static (time-invariant) and dynamic (time-varying) components, and further into multiple independent factors, by shaping the eigenvalues of the learned Koopman matrix, where eigenvectors associated with unit eigenvalues capture the invariant content. Second, for generative modeling, I recast the multi-step reverse process of diffusion models as a dynamical system and learn a finite-dimensional Koopman operator that maps noise to data in a single step, enabling fast offline distillation. We prove that such a finite operator exists under mild conditions and that it preserves the structure of the original generative dynamics, and we show competitive one-step generation results on standard image benchmarks. Together, these results illustrate how operator-theoretic structure can unify representation learning and efficient generation. (Based on "Multifactor Sequential Disentanglement via Structured Koopman Autoencoders," ICLR 2023 Spotlight, and "One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling," NeurIPS 2025.)
Speaker: Ilan Naiman
Supervisor: Dr. Omri Azencot
Date: Monday, 15/6, talk starts at 12:30
Place: Building 37, room 202
Talk title: Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs
Abstract: Generating realistic time series data is important for many engineering and scientific applications. Existing work tackles this problem using generative adversarial networks (GANs). However, GANs are unstable during training, and they can suffer from mode collapse. While variational autoencoders (VAEs) are known to be more robust to these issues, they are (surprisingly) less considered for time series generation. In this talk, I will present Koopman VAE (KoVAE), a novel generative framework that successfully models both regular and irregular time series data. Inspired by Koopman theory, KoVAE introduces a unique model prior that represents latent conditional prior dynamics using a linear map. This innovative approach not only provides robust generative performance but also allows for the incorporation of domain knowledge via spectral constraints on the map's eigenvalues, while enabling the analysis of the system's qualitative behavior and stability. I will share results demonstrating how KoVAE outperforms state-of-the-art GAN and VAE baselines across multiple real-world and synthetic benchmarks, ultimately generating high-quality time series that significantly improve both discriminative and predictive metrics.
