An artificial recurrent network reproduces the within-/outside-manifold learning asymmetry seen in monkey motor cortex (Feulner & Clopath 2021)

Feulner & Clopath, "Neural manifold under plasticity in a goal driven learning behaviour" (PLoS Comput. Biol. 17(2):e1008621, 2021), build a recurrent neural network model of the goal-driven brain–computer-interface task and ask whether an artificial network trained by synaptic plasticity shows the same asymmetry that claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists found in monkeys. It does. Modifying recurrent weights under a learned feedback signal accounts for the observed behavioural gap: the model adapts readily when the required change lies inside the network's existing low-dimensional subspace and struggles when it lies outside.

The paper's framing is that "successful learning is naturally constrained to a common subspace," and that "learning the feedback signal from scratch was only possible for within-manifold perturbations." The asymmetry therefore need not be a special property of biological cortex; it emerges in a generic plastic recurrent network, which upgrades the brain-to-network correspondence from a loose analogy toward a mechanistic one: the same geometric constraint on adaptation appears when the substrate is silicon rather than tissue.

This is the in-silico leg of a three-domain recurrence. The biological anchor is claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists; the large-language-model analogue — adaptation confined to a low-dimensional subspace — is claim-aghajanyan-2020-fine-tuning-low-intrinsic-dimension. The connective synthesis, and the open question of whether these are one mathematical object or three analogies, is observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets.