Human Inductive Biases
Sreejan discusses an experimental framework that contrasts human and machine task distributions, revealing that RL agents trained on human tasks generalize better to machine tasks. The key lies in using compressive abstractions in training, which mirror human-like descriptions and help instill inductive biases into machines. This approach challenges traditional views on architecture and suggests that useful biases can be harvested from language and programs rather than being explicitly designed.In this clip
From this podcast
Machine Learning Street Talk (MLST)
#97 SREEJAN KUMAR - Human Inductive Biases in Machines from Language
Related Questions