5/1(수) [KAIST Bio-IT Healthcare Initiative III] Google DeepMind's Neuroscience-inspired AI 세미나 시리즈로 Google DeepMind의 Dr. Zeb Kurth-Nelson을 모시고 Inspiring Talk를 개최하였습니다.
Title: Distributions from Dopamine and Factorized Replay
- Abstract : I'll present two new studies. 1) In the classic TD theory of dopamine, the brain learns about the mean of probabilistic rewards. Recent AI research has gotten large performance gains from modifying the TD algorithm to learn about the full distribution of rewards. We looked in single unit dopamine recordings and found evidence that the brain uses this modified algorithm. 2) There are fast spontaneous neural sequences, with properties similar to rodent hippocampal replay, in humans in non-spatial tasks. In a new experiment, we found that these sequences don't simply recapitulate experience, but are re-organized according to abstract knowledge. Each replay event is also interleaved with replay of a factorized code, which we argue could facilitate rapid transfer of abstract knowledge to new experiences.