PhD Proposal Exam
Tuesday May 16th, 10:10 am – Ramsay Wright Building, Rm. 432
Annik Yalnizyan Carson (Richards lab)
“Episodic Control: The Role of Memory Systems in Decision Making”
Reinforcement learning (RL) is an area of machine learning concerned with optimal behavioural control. RL provides a normative framework in which to understand how the brain can learn to make decisions for maximizing subjective reward in the absence of an explicit teaching signal. Currently, even agents using state-of-the-art control systems in RL tasks are data inefficient and challenged by nonstationary environmental conditions, including changes in statistics of reward probability and transitions between states, which biological agents handle with relative ease. It has been proposed that storing information about experienced episodes in a memory cache — modeled after the activity of the hippocampus in animals — can help bootstrap learning in RL systems to improve the speed of learning and ability to cope with nonstationary environments. My research proposes three different representations for episodic memories stored in such a system and aims to resolve which provides the greatest benefit to RL systems when used in conjunction with a standard controller. Furthermore I aim to resolve how these representations can account for features of animal behaviour, and which of these representations — if any — are likely to explain how episodic memory is represented in the hippocampus.
Ramsay Wright is a wheelchair accessible building.