Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Poster Sessions | Data Blitz

Feedforward and reconstructive computations engage dissociable neural modules during memory encoding

Poster Session C - Sunday, March 8, 2026, 5:00 – 7:00 pm PDT, Fairview/Kitsilano Ballroom

Gracie Shao¹ (), Aalap Shah², Ilker Yildirim², Qi Lin³; ¹University of British Columbia, ²Yale University, ³Institute for Basic Science, Republic of Korea

How does the brain transform perceptual experiences into durable memory traces? Previous computational and behavioral work identifies at least two computational processes that explain substantial variance in the memorability of images in a dissociable manner: feedforward processing as captured by a deep convolutional neural network (DCNN) and reconstructive processing as captured by a sparse coding model (SPC). Yet whether and how these two computational processes may be dissociable neurally remains elusive. Here, we investigate how these two computational processes may map onto brain activities during memory encoding using 7T fMRI data from the Natural Scenes Dataset (NSD). Adopting computational architecture from previous work, we trained sparse coding models to reconstruct DCNN embeddings of the 73000 NSD stimuli and derived two computational signatures per image: distinctiveness and reconstruction error, corresponding to feedforward processing and reconstructive processing respectively. Behaviorally, we replicated previous findings: images with more distinctive and harder-to-reconstruct representations are retrieved more accurately and faster. We then combined these two computational signatures in regression models to predict activities in brain regions engaging in memory encoding suggested by previous literature, that is, visual regions, medial temporal lobe and frontoparietal regions. Interestingly, while distinctiveness explains the most variance in visual regions, reconstruction error explains the most variance in the medial temporal lobe and frontoparietal regions. Together, these findings bridge the explanatory gap between memory behavior and the brain with computational modeling, revealing an integrated mechanism of memory encoding driven by feedforward and reconstructive computations across dissociable neural modules.

Topic Area: LONG-TERM MEMORY: Other