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

Orthogonal Neural Subspaces in Working Memory Circuits Support Interference Protection and Flexible Updates

Liu Yuezhang¹, Zhang Ziyao², Jarrod A. Lewis-Peacock³, Xue-Xin Wei¹; ¹University of Texas at Austin, ²University of California, Berkeley, ³UT Austin

Presenter: Liu Yuezhang

Flexible cognitive behavior requires simultaneously protecting internal representations from distraction and selectively updating them when new information becomes relevant. How neural circuits balance this stability-flexibility trade-off remains unclear. Here, we test whether the orthogonalization of neural subspaces underlies this balance. In both macaque peri-arcuate cortex (PAC) and task-optimized recurrent neural networks (RNNs), we find that neurons exhibit heterogeneous tuning shifts between encoding and delay periods. Analyses of neural geometry reveal that sensory and memory representations occupy orthogonal subspaces, reducing interference. When new sensory information needs to be incorporated, updating is achieved via a rotation transferring representations from the sensory into memory subspace. These findings provide a mechanistic account of how neural circuits reconcile stability and flexibility in working memory.

Topic Area: Memory, Learning & Knowledge Structures