Search Papers | Poster Sessions | All Posters
Poster B39 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
A Normative Account of the Influence of Contextual Familiarity and Novelty on Episodic Memory Policy
Qihong Lu1 (), Kenneth Norman2, Daphna Shohamy1; 1Columbia University, 2Princeton University
How do humans decide when to retrieve and when to encode episodic memories (EMs)? Empirical results show that seeing a familiar stimulus biases subjects toward retrieval, while seeing a novel stimulus biases subjects toward encoding, even though that stimulus is incidental to the task. From a normative standpoint, it is unclear why the familiarity of incidental stimuli should bias EM. We hypothesized that these biases could arise because the EM policy – whether to retrieve or encode at a given moment – is learned in an environment where stimulus familiarity is autocorrelated in time. We present an EM-augmented neural network that learns an EM policy using reinforcement learning. Learning to encode was facilitated by allowing the reward obtained by retrieval to propagate back to reinforce the action of encoding this memory. As our model learns in an autocorrelated environment, empirically observed effects of familiarity emerged. This is because, in an environment where familiar stimuli tend to precede other familiar stimuli, familiarity indicates that relevant EMs are present, making retrieval more useful. Novelty encourages encoding for the same reason. Our results suggest that the influences of familiarity and novelty are adaptive features of human EM policy in response to autocorrelated environments.
Keywords: episodic memory familiarity & novelty neural network reinforcement learning