Elements of episodic memory: insights from artificial agents

Alexandria Boyle,Andrea Blomkvist
DOI: https://doi.org/10.1098/rstb.2023.0416
2024-09-16
Philosophical Transactions of the Royal Society B Biological Sciences
Abstract:Many recent artificial intelligence (AI) systems take inspiration from biological episodic memory. Here, we ask how these ‘episodic-inspired’ AI systems might inform our understanding of biological episodic memory. We discuss work showing that these systems implement some key features of episodic memory while differing in important respects and appear to enjoy behavioural advantages in the domains of strategic decision-making, fast learning, navigation, exploration and acting over temporal distance. We propose that these systems could be used to evaluate competing theories of episodic memory’s operations and function. However, further work is needed to validate them as models of episodic memory and isolate the contributions of their memory systems to their behaviour. More immediately, we propose that these systems have a role to play in directing episodic memory research by highlighting novel or neglected hypotheses as pursuit-worthy. In this vein, we propose that the evidence reviewed here highlights two pursuit-worthy hypotheses about episodic memory’s function: that it plays a role in planning that is independent of future-oriented simulation, and that it is adaptive in virtue of its contributions to fast learning in novel, sparse-reward environments. This article is part of the theme issue ‘Elements of episodic memory: lessons from 40 years of research’.
biology
What problem does this paper attempt to address?