PIATRA . INSTITUTE

geometries of action

exploring neural manifolds as geometry, dynamics, alignment, and decoding

based on Gallego et al., The Neural Manifold Manifesto

ClaudeMarch 2026·Initial implementation with six-panel manifold laboratory

Manifold geometry

Left: activity embedded in neural state space. Right: compare nonlinear unfolding against linear projection.

Embedded state space

Intrinsic unfoldingdistortion 1.0%

Intrinsic dim

Latent degrees of freedom

Embedding dim

Visible state-space axes

Linear faithfulness

76%

How much a linear view recovers

Smooth geometry

97%

Surface continuity under perturbation

Population activity

The manifold view does not abolish single neurons. It says isolated neurons are often insufficient summaries of the computation.

Population heatmap

12 units34 time bins

Single neuron trace

Population coherence

84%

Ensemble coordination

Recorded units

Compressed to 2D latent space

Dynamics and causal perturbation

Cooling should bias timing by slowing traversal along the manifold rather than by destroying the manifold itself.

Traversal speed

62%

Effective latent dynamics

Timing bias

Under/overestimation risk

Geometry retained

97%

Structure after perturbation

Cross-subject alignment

A strong manifold claim: related tasks induce comparable latent structure even when individual neurons differ.

Alignment

65%

Latent overlap

Shared grammar

74%

Task narrows space

Cross-map stability

62%

Robustness to noise

Clinical decoding

Weak residual activity may live on a meaningful low-dimensional structure decodable for control.

Residual signal

35%

Structured motor intent

Control quality

39%

Decoder performance

Parameter sweep

Sweep one parameter across its range while holding others constant.

Sensitivity analysis

Which parameters does decoder confidence depend on most?

sensitivity analysis · decoder confidence

residual

Δ74.100

noise

Δ7.920

curvature

Δ0.000

task constraint

Δ0.000

speed

Δ0.000

cooling

Δ0.000

alignment

Δ0.000

baseline: 39.140 · each parameter swept min→max while others held constant

This is a pedagogical model. The formulas are simplified proxies for neural population dynamics, not fitted to real electrophysiology data. Use it to build intuition about the manifold framework, not to make quantitative predictions.

2D motor sheet · phase 0%

The neural manifold hypothesis

The central claim is that neural population activity, rather than scattering freely through the high-dimensional space of all possible firing-rate combinations, is confined to a smooth, low-dimensional surface — a manifold. The intrinsic dimensionality of this surface reflects the degrees of freedom of the task, not the number of neurons recorded.

\mathbf{x}(t) \in \mathcal{M}^d \subset \mathbb{R}^N, \quad d \ll N

where $\mathbf{x}(t)$ is the population state at time $t$ , $\mathcal{M}^d$ is the $d$ -dimensional manifold, and $N$ is the number of recorded neurons.

Geometry is not projection

A linear dimensionality reduction like PCA can flatten a curved manifold, collapsing distances and distorting neighborhoods. Nonlinear methods (UMAP, diffusion maps, Isomap) attempt to unfold the intrinsic geometry. The playground's projection toggle makes this distinction visible: when curvature is high, the linear projection's distortion metric climbs while the nonlinear unfolding stays faithful.

Dynamics versus geometry

Gallego and colleagues demonstrated that cooling the striatum during an interval-timing task slows traversal speed along the manifold without substantially altering the manifold's shape. This dissociation between dynamics and geometry is central to the causal reading of the framework: the manifold constrains which states are reachable, while dynamics determine how fast they are reached.

Cross-subject invariance

If manifolds are ontologically real — as Gallego argues — then different individuals performing the same task should exhibit comparable latent structure despite having entirely different neurons. Alignment methods like canonical correlation analysis (CCA) and Procrustes rotation can quantify this overlap. The alignment slider explores how shared task constraints push two trajectories toward a common geometry.

Clinical translation

In patients with clinically complete spinal cord injuries, residual descending signals may still carry structured low-dimensional information about intended movements. Decoding this residual manifold structure is the basis for emerging neuroprosthetic interfaces where patients control virtual cursors or wheelchairs by attempting to move.

Limitations of this model

This playground uses simplified parametric formulas, not real electrophysiology data. The manifold is generated analytically rather than extracted from neural recordings via dimensionality reduction. Metric values (decoder confidence, alignment score) are proxies that capture qualitative relationships, not quantitative predictions. The model omits spike-timing correlations, trial-by-trial variability, and the multi-area distributed nature of real manifold computations.

v1March 2026

Six-panel visualization: manifold geometry, population activity, dynamics, alignment, decoder, sweep
Four presets mapping manifold framework claims: motor reach, timing + cooling, cross-subject, spinal decoding
Linear vs nonlinear projection comparison with distortion metric
Continuous trajectory animation with play/pause and speed modulation via cooling
Cross-subject alignment with configurable warping
Clinical decoder cursor driven by residual manifold structure
Parameter sweep and sensitivity analysis on decoder confidence