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

Depth-of-processing-like computations explain visually-evoked activity in the human medial temporal lobe

Poster Session A - Saturday, March 7, 2026, 3:00 – 5:00 pm PST, Fairview/Kitsilano Ballroom
Also presenting in Data Blitz Session 1 - Saturday, March 7, 2026, 10:30 am – 12:00 pm PST, Salon ABC.

Aalap Shah¹ (), Yuchang Tian¹, Qi Lin², Runnan Cao³, Shuo Wang³, Ilker Yildirim¹; ¹Yale University, ²Institute for Basic Science, Republic of Korea, ³Washington University in St. Louis

Decades of research have shown that spontaneous visually-evoked activity in the Medial Temporal Lobe (MTL) is consequential for later memory performance. Yet, the computational basis of such visually evoked processing in the MTL remains unclear. Existing modeling work has focused on standard visual object recognition systems and their variants fine-tuned on human memory performance, without addressing how visual processing influences the strength of memory traces. In contrast, here we take a reverse-engineering approach that explicitly links a priori computational principles with neurobiological plausibility. We propose that visually evoked processing in the MTL can be understood through the lens of Craik and Lockhart’s depth-of-processing hypothesis—a foundational account of the perception-to-memory interface. Drawing on recent modeling work, we used compression-based reconstruction error from autoencoders as an image-computable signature of depth-of-processing, with the idea that images with harder-to-reconstruct representations evoke greater processing in the MTL (and vice-versa). We computed reconstruction error from two autoencoders corresponding to the extrema of category decodability within the compressed code: category-agnostic and category-informed models. We analyzed single-cell intracranial MTL recordings in human participants (n=15), comprising 362 hippocampal and 446 amygdala neurons, while they passively viewed 500 real-world images spanning 50 categories. Reconstruction error yielded an interpretable, functional window onto MTL activity, revealing a robust double dissociation: category-informed reconstruction error correlated significantly only with hippocampal activity, whereas category-agnostic reconstruction error correlated significantly only with amygdala. Moreover, reconstruction error outperformed alternatives in each MTL subregion. These findings suggest depth-of-processing as an algorithmic-level account of stimulus-driven visual processing in the MTL.

Topic Area: LONG-TERM MEMORY: Other