Learning to Scale Logits for Temperature-Conditional GFlowNets

TL;DR

This paper introduces Logit-GFN, a new architecture for temperature-conditional GFlowNets that improves training speed and stability by scaling logits with a learned function of temperature, enhancing exploration and generalization.

Contribution

The paper proposes Logit-GFN, a novel method that addresses numerical challenges in training temperature-conditional GFlowNets by using a learned temperature-dependent logit scaling function.

Findings

Logit-GFN accelerates training of temperature-conditional GFlowNets.

It improves offline generalization and mode discovery in various tasks.

Empirical results validate the effectiveness across biological and chemical applications.

Abstract

GFlowNets are probabilistic models that sequentially generate compositional structures through a stochastic policy. Among GFlowNets, temperature-conditional GFlowNets can introduce temperature-based controllability for exploration and exploitation. We propose \textit{Logit-scaling GFlowNets} (Logit-GFN), a novel architectural design that greatly accelerates the training of temperature-conditional GFlowNets. It is based on the idea that previously proposed approaches introduced numerical challenges in the deep network training, since different temperatures may give rise to very different gradient profiles as well as magnitudes of the policy's logits. We find that the challenge is greatly reduced if a learned function of the temperature is used to scale the policy's logits directly. Also, using Logit-GFN, GFlowNets can be improved by having better generalization capabilities in offline learning and mode discovery capabilities in online learning, which is empirically verified in various biological and chemical tasks. Our code is available at \url{https://github.com/dbsxodud-11/logit-gfn}