Семинар MTML Lab «Спектральные оценки и GFlowNet»

На заседании разберем две статьи: 
1. "Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy", (https://openreview.net/pdf?id=F0JzotXYgC) NeurIPS 2025 (oral). 
TL;DR: In this paper, sharp spectral-norm bounds for noisy low-rank approximation are derived, improving prior results by up to square root of n. Applied to DP-PCA, the proposed method resolves an open problem and matches empirical error via a novel contour bootstrapping technique.
2. GFlowNet Foundations, (https://www.jmlr.org/papers/volume24/22-0364/22-0364.pdf) JMLR 2023.
TL;DR: Generative Flow Networks (GFlowNets) have been introduced as a method to sample a diverse set of candidates in an active learning context, with a training objective that makes them approximately sample in proportion to a given reward function. In this paper, the authors show a number of additional theoretical properties of GFlowNets, including a new local and efficient training objective called detailed balance for the analogy with MCMC.