Agent-centric learning in minds and machines
File(s)
Author(s)
Lee, Sebastian Alexander
Type
Thesis
Abstract
Despite the remarkable abilities of modern deep learning systems in a range of domains, they remain sample inefficient and rigid; they exemplify task-centric intelligence, in contrast to the agent-centric intelligence typical of biological agents. Addressing this divide between minds and machines is critical to the endeavour of designing more general machine intelligence. This thesis seeks to take steps in this direction in the context of various learning paradigms including continual learning and reinforcement learning—both in understanding the properties of this divide from a theoretical perspective, and in practically advancing algorithms to narrow this gap.
Version
Open Access
Date Issued
2024-07-16
Date Awarded
01/01/2025
License URL
Advisor
Clopath, Claudia
Saxe, Andrew
Publisher Department
Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)