Monkeys learn within-manifold brain–computer-interface mappings within hours but resist outside-manifold mappings on the same timescale (Sadtler et al. 2014)

Sadtler, Quick, Golub, Chase, Yu & Batista et al., "Neural constraints on learning" (Nature 512:423–426, 2014), used a closed-loop intracortical brain–computer interface (BCI) in which rhesus macaques drove a cursor by modulating motor-cortex population activity. Because the experimenters controlled the decoder mapping neural activity to cursor motion, they could require the animal either to produce activity patterns that lay inside the population's pre-existing low-dimensional structure — its "intrinsic manifold" — or ones that lay outside it. Within-manifold mappings were learned readily over a session; outside-manifold mappings were not.

The load-bearing finding is that the manifold is not merely a description of population activity but a constraint on what can be learned: "On a timescale of hours, it seems to be difficult to learn to generate neural activity patterns that are not consistent with the existing network structure," and, more generally, "the existing structure of a network can shape learning." A low-dimensional subspace that summarizes normal activity thus doubles as a learning prior — cheap adaptation inside it, near-resistance outside it on short timescales.

This is the biological anchor of a cross-domain recurrence. The same within-/outside-manifold asymmetry is reproduced in an artificial recurrent network by claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry, and an analogous low-dimensional-adaptation result appears in large language models via claim-aghajanyan-2020-fine-tuning-low-intrinsic-dimension. The synthesis tying these together — and the caution that the analogy is not (yet) a shared mathematical object — is observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets. The originating thread, low-dimensional confinement of an LLM's own self-monitoring, is claim-llm-neural-metacognition-incomplete.