Schedule of Events | Symposia

The representation of ordinal context in visual working memory

Poster Session B - Sunday, March 30, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Jung Woo Hur1, Jacqueline M. Fulvio1, Bradley R. Postle1; 1University of Wisconsin-Madison

The binding of stimulus information to the context in which it was encountered (i.e., the “where” and “when”) is essential of visual working memory. To study this mechanism, we trained recurrent neural networks (RNNs) to perform a double-serial retrocuing (DSR) task in which two stimulus items (S1 and S2) are presented serially, followed by a first cue indicating which will be tested at the first recall epoch, then a second cue, indicating which will be tested at the second recall epoch. In order to visualize the representational geometry of the encoding of context, we applied demixed principal component analysis (dPCA) to the hidden layer during the presentation of S1 and of S2, then projected data from subsequent epochs into these “stimulus encoding” subspaces, to track representational transformations. This revealed that the initial representational geometry of S1, during its encoding, underwent a representational ‘flip’ (i.e. reflection) upon presentation of S2, and S2 was encoded into the subspace that had been occupied by S1. This representational transformation of S1 suggests that order-context (i.e., “when” an item was presented) may be encoded chronologically (i.e. most-recent, second most-recent, etc.) rather than ordinally (i.e. 1st, 2nd, etc.). Next steps for this work are to extend it to more than two items, and assess its applicability to human working memory performance.

Topic Area: EXECUTIVE PROCESSES: Working memory

CNS Account Login

CNS2025-Logo_FNL_HZ-150_REV

March 29–April 1  |  2025

Latest from Twitter