Machine Learning Street Talk (MLST) #063 - Prof. YOSHUA BENGIO - GFlowNets, Consciousness & Causality
6 snips
Feb 22, 2022 
Yoshua Bengio, a Turing Award recipient and a leader in AI, dives into the fascinating world of GFlowNets, which he believes can revolutionize machine learning by generating diverse training data. The discussion covers the balance between exploration and exploitation in decision-making, particularly in drug discovery and gaming. Bengio also addresses the philosophical implications of consciousness in AI, urging a cautious perspective on claims of AI sentience. His reflections on the evolution of thought in neural networks reveal a journey shaped by key insights into causal representation learning.
AI Snips
Chapters
Transcript
Episode notes
GFlowNets: A Swiss Army Knife for Probabilistic Modeling
- GFlowNets offer a powerful new framework for generic learnable inference in probabilistic machine learning.
- They can potentially replace MCMC sampling and estimate intractable quantities like partition functions.
Visualizing GFlowNets with a Galton Board
- The Galton board analogy helps visualize GFlowNets: beads flowing through pegs, representing flow gates, into buckets.
- GFlowNets optimize these "flow gates" to match any desired distribution, like a reward function in reinforcement learning.
GFlowNets, Entropy, and Uncertainty
- GFlowNets translate a reward function into a sampling mechanism, maintaining entropy like the reward function itself.
- They can estimate entropy, crucial for choosing actions minimizing uncertainty, unlike traditional reinforcement learning.