Loading data…

About the Explorer

The Neural Manifold Explorer lets you examine how a neural population represents visual motion stimuli. It combines two views of the same population:

• Encoding manifold (left): Neurons in stimulus-response space. Each point is a neuron; position reflects similarity of response profiles across stimuli and directions. Close neurons on the manifold exhibit similar response profiles.
• Decoding manifold (centre): Stimuli in neural activity space. Each point is the population's mean response to one stimulus × direction combination, projected via PCA. Tight, separated clusters indicate the population can discriminate stimuli.
• Decoding trajectories (right): The same PCA space, showing how population activity evolves over time.

Selecting neurons. Use the tabs to define a subpopulation:

• Bounding Box: Drag sliders to select neurons inside an axis-aligned box in the embedding.
• Radius: Position the seed point, then adjust the radius.
• Metric: Rank neurons by a functional property and keep the top-N%. Click Run to select, update the decoding plots, and reveal the fraction-sweep curves. The five properties are:
  • Speed: mean absolute frame-to-frame change in activity during the early response window (up to 40 time steps from stimulus onset), averaged over directions and stimuli. High-speed neurons show large, rapid fluctuations at stimulus onset.
  • Stability: mean absolute frame-to-frame change during the late response window (last 40 time steps), averaged over directions and stimuli. Low stability means the response is still fluctuating at the end of the trial rather than settling into a steady state.
  • Curvature: RMS deviation of the early response (Gaussian-smoothed, σ = 2) from a linear fit over the early window, averaged over stimuli. High-curvature neurons have a non-linear onset — e.g. an initial transient followed by adaptation — rather than a simple linear ramp.
  • Classifiability: variance of the mean steady-state response (last 10 time steps, averaged over directions) across stimuli. High-classifiability neurons fire differently at steady-state for different stimuli and are therefore strongly stimulus-selective.
  • PC contribution: L2 norm of the neuron's loading on the first principal component of the full population response (all stimuli × directions × time steps stacked). High-PC-contribution neurons are the strongest drivers of the dominant mode of population-wide variance.

Fraction sweep (Metric tab, appears after Run): Shows how two population-level metrics change as neurons are progressively removed, keeping only the top-ranked fraction. Dark blue = high-ranked neurons; light blue = low-ranked; grey dashed = random subsampling baseline (± 1 s.d. shaded). The x-axis is inverted so the full population (fraction = 1) is on the left.

• Decoding accuracy: k-nearest-neighbour accuracy (k = 5) on stimulus identity in the PCA decoding space. Values near 1 mean the subpopulation's activity reliably separates all stimuli.
• Geometric fidelity (R²): Procrustes R² between the subpopulation's decoding manifold and the full-population manifold. Values near 1 mean the subpopulation preserves the geometric structure of the full population's stimulus representation.

A curve that stays high at small fractions means that property identifies the most informative neurons.

PSTHs. After selecting neurons, click Update Decoding ↻ to re-compute the decoding manifold for the subpopulation, or Plot PSTHs ▶ to display response heatmaps. For ≤ 5 neurons, individual per-neuron strips are shown; for larger groups, the population-averaged heatmap is displayed.

Activity over time. Select activity from the Color by dropdown to animate the encoding manifold by neural activity. Each neuron's color reflects its response to the selected stimulus at the current time frame, mapped to a viridis colorscale (dark purple = low, yellow = high), normalised across the population at each frame. Use the Stim dropdown to choose which stimulus is shown (and Dir for datasets with multiple motion directions). Drag the time slider to inspect a specific frame, or click ▶ Play to animate at 8 fps. This view lets you directly read off which region of the encoding manifold is most active at stimulus onset, during the sustained response, or at offset — and how those active sub-populations shift over time.