Symmetry-Induced Non-Identifiability in Neural Circuit Inference

Seungwon Yu, Jaeho Yang, Kijung Yoon

Keywords: neural circuit inference, symmetry-induced non-identifiability, positional encoding

TL;DR: We show that neural connectivity inference can be dominated by symmetry-aligned positional structure rather than spike activity. In ring attractor networks, positional encoding alone recovers stable connectivity.

Abstract: Inferring synaptic weights from neural activity is commonly framed as a data-driven problem, yet inferred connectivity can become largely independent of spike inputs even when models are trained on them. Using a graph-based inference framework, we show that positional encoding alone can recover stable connectivity structure, with inferred weights remaining consistent when spike inputs are removed at test time. We interpret this as symmetry-induced degeneracy, where invariances render multiple connectivity configurations observationally equivalent and can dominate over activity-dependent evidence. These findings highlight the strong influence of inductive bias and suggest that neural circuit inference should be understood as learning within a symmetry-constrained weight space.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 44