neural manifold

In systems neuroscience, the low-dimensional subspace to which a neural population's activity is largely confined — spanned by a handful of covarying "neural modes" rather than filling the full space of independent single-neuron firing rates. The hypothesis (reviewed by Jazayeri & Ostojic and others, building on Gallego et al. 2017) is that this structure, not individual neurons, is the meaningful unit of population computation, and that it recurs across brain regions, behaviours, and species.

The construct is the organizing object of a whole cluster in this vault: it is what claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists shows constrains learning (cheap adaptation inside the manifold, resistance outside it), what claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry reproduces in an artificial recurrent network, and one of three domains in the low-dimensional-subspace recurrence. Kept deliberately distinct from entity-intrinsic-dimension: the manifold is the geometric object (the subspace itself), whereas intrinsic dimension is a count (how many dimensions it spans) — and the neural-manifold's dimensionality is the same type of object as representation-space intrinsic dimension, not the weight-space one (claim-jazayeri-ostojic-2021-neural-manifold-intrinsic-dimension-parametrizes-activity).

References

• claim-sadtler-2014-within-manifold-bci-learning-fast-outside-resists · entity-sadtler-2014-neural-constraints
• claim-feulner-clopath-2021-rnn-reproduces-manifold-learning-asymmetry
• claim-jazayeri-ostojic-2021-neural-manifold-intrinsic-dimension-parametrizes-activity
• observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets · entity-intrinsic-dimension
• question-low-dimensional-subspace-one-object-or-analogy