Geometric determinants of time series learning and attractor stability in closed-loop circuits

Adding a feed-forward layer to a cortico-cerebellar-like closed-loop circuit enables learning periodic target signals that matched recurrent networks cannot, by stabilizing attractor states.

• Why it matters: Autonomous, structured neural activity must be learned and kept stable in distributed motor circuits, but the link between neural representation geometry and learning is unclear.
• What they did: The study compared a cortex-like recurrent neural network (RNN) with an RNN+FFL closed-loop circuit, then decomposed population activity by frequency to identify task-relevant versus task-irrelevant subspaces.
• The result: The feed-forward layer improved learning and stabilized attractors by increasing angles within and between these subspaces, making representational geometry a better predictor than global dimensionality.

AI summary of the abstract · may contain errors · not medical advice · read the paper.