Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Neural generalization principles of working memory in humans and recurrent neural networks

Dongping Shi¹, Luchengchen Shu², Qing Yu³; ¹University of the Chinese Academy of Sciences, ²Institute of Neuroscience, ³Institute of Neuroscience, Chinese Academy of Sciences

Presenter: Qing Yu

A fundamental endeavor in cognitive neuroscience is to understand how information can be rapidly abstracted through shared perceptual or structural knowledge to facilitate efficiency and learning. Working memory (WM) provides a flexible mental workspace for these computations, yet how generalization is realized within WM remains largely unexplored. Here, using functional MRI (fMRI) and recurrent neural network (RNN) modeling, we investigated how stimulus and rule information generalize within WM. Across two experiments, participants performed two WM tasks with shared stimulus structure but distinct stimulus sets (location and object), either without (Experiment 1) or with (Experiment 2) explicit mapping. In each task, they flexibly switched between maintenance and manipulation of stimulus information following task rules. Leveraging multivariate decoding and state space analyses, we revealed separate neural substrates in the generalization of stimulus and rule information in WM: the posterior parietal cortex represented mnemonic information across stimulus domains, with enhanced generalization of mnemonic information during memory manipulation compared to maintenance. In contrast, frontal subregions encoded abstract rules that were generalizable across tasks. RNN simulations replicated the key generalization patterns. Together, our findings reveal the neural generalization principles of WM that enable flexible maintenance and manipulation of information for goal-directed behavior.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF