Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Feedback-gated synaptic modulation enables context-dependent computation in recurrent neural networks

Jingcheng Shi¹, Drew B. Headley¹; ¹Rutgers University

Presenter: Jingcheng Shi

Recurrent neural networks (RNNs) are widely used to model sequential processing in cortical circuits, yet they typically omit a key feature: contextual feedback from higher-order regions that modulates synaptic integration at the dendritic level. Such modulation is achieved through synaptic clustering and nonlinear dendritic integration, enabling input-specific regulation of individual synaptic contributions. Here we introduce Gating on Weights with Feedback (GaWF), a biologically inspired RNN architecture in which feedback signals dynamically gate individual synaptic weights—contrasting with conventional gated RNNs (LSTMs, GRUs), where gates are driven by local network activity and act on hidden unit states. We evaluate GaWF on a dynamic visual task requiring continuous tracking of a target digit amid randomly moving distractors, called the Cluttered Tracking MNIST task (CT-MNIST). GaWF generalizes substantially better than conventional gated RNNs, which overfit the training sequences. Interpretability analysis of the learned gating signals reveals selective enhancement of input connections associated with the target’s spatial location, suggesting a mechanism for feedback-driven, input-specific modulation analogous to gain control in cortical systems.

Topic Area: Methods, Tools, Theory & Neural Coding