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

Divergent working memory dynamics in biological and recurrent neural networks

Daria Kussovska¹, Robert Kim², Nuttida Rungratsameetaweemana¹; ¹Columbia University, ²Cedars-Sinai Medical Center

Presenter: Daria Kussovska

Persistent activity has long been considered a primary mechanism for maintaining information in working memory (WM). However, most supporting evidence relies on trial-averaged neural responses, which can obscure temporal dynamics during maintenance and mask contributions of heterogeneous cell types at the single-trial level. Here, we analyzed single-trial spiking activity from intracranial recordings of neurosurgical patients performing a WM task (Kyzar et al., 2024). We trained biologically plausible recurrent neural network (RNN) models on an analogous task and directly compared maintenance dynamics between the neural data and the trained RNNs. We found that RNN models relied on sustained persistent firing to maintain task-relevant information, whereas human neurons exhibited intermittent activity during maintenance. These results extend the persistent activity framework, suggesting that human WM maintenance involves intermittent coding strategies that differ from the emergent solutions in network models.

Topic Area: Memory, Learning & Knowledge Structures