Paper Accepted at ICLR 2025 (Spotlight): When do GFlowNets Learn the Right Distribution?
Table of Contents
I am delighted to announce that our paper, When do GFlowNets Learn the Right Distribution?, has been accepted for presentation at ICLR 2025! This work advances our theoretical understanding of Generative Flow Networks (GFlowNets) by examining the impact of balance violations on their ability to approximate target distributions and proposing a novel metric for assessing correctness.
Update: our work has been selected as a Spotlight at ICLR 2025 🎉 (5.1% of the submitted papers are selected for a spotlight or an oral)
When do GFlowNets Learn the Right Distribution? #
Authors: Tiago Silva, Rodrigo Barreto Alves, Eliezer de Souza da Silva, Amauri H Souza, Vikas Garg, Samuel Kaski, Diego Mesquita
Conference: International Conference on Learning Representations (ICLR 2025)
Summary:
We analyze when and whether GFlowNets successfully approximate the correct target distribution, addressing their stability under balance violations and the limitations of their parameterization with graph neural networks (GNNs). Our findings show that imbalance effects are unevenly distributed across the flow network, impacting accuracy based on flow intensity. We also highlight the fundamental representation limits of GNN-based GFlowNets when learning distributions over graphs. We propose a new metric for evaluating GFlowNets to address these challenges, empirically demonstrating its superior effectiveness over existing evaluation protocols.
Links:
Reviewers received this paper well, recognizing its theoretical and empirical contributions. I look forward to presenting this work at ICLR 2025 and discussing its implications with the broader ML community.
Thanks to the co-authors for their incredible efforts in bringing this research to life!
Stay tuned for more updates from ICLR 2025!