Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Gating by Both Relative and Absolute Match Helps to Optimize Episodic Memory Retrieval

Cody V. Dong¹, Kenneth A. Norman¹; ¹Princeton University

Presenter: Cody V. Dong

In real life, information is continuously revealed, and it can be unclear which memories (if any) are relevant to the current situation. How do we optimize the usefulness of episodic memory retrieval in this challenging situation? Prior work suggests that humans can gate episodic memory retrieval based on the absolute level of memory match and also relative memory match (i.e., does the best-matching memory match better than other memories), but it is unclear how these properties contribute to successful memory use in complex naturalistic situations. To answer this question, models were incrementally presented with a rolling summary of a story that might or might not have been previously encoded into external memory, and this rolling summary was compared to all of the story chunks stored in external memory. A separate retrieval policy network learned to gate retrieval in order to maximize retrieval sensitivity (recalling in response to familiar but not novel stories), relevance (recalling chunks from the correct story), and timeliness (recalling earlier rather than later, all other things being equal). The inputs to the retrieval policy network were the absolute and relative levels of match for the best-matching story chunks and also time (i.e., how much of the summary had been revealed so far). We found that networks that jointly consider both absolute and relative match outperform those using either alone, indicating that these two signals contain non-redundant information that the model can use to optimize memory retrieval.

Topic Area: Memory, Learning & Knowledge Structures