Interactive explorer for the paper How 'Neural' is a Neural Foundation Model? (Bertram et al., 2026) — full code here.
The paper asks whether the internal representations of a neural foundation model (FNN) trained to predict neural activity in mouse visual cortex actually resemble biological neural population geometry — or whether good predictive performance can coexist with fundamentally different representational structure. We analyse this by comparing the manifold each population traces out in response space: the shape, dimensionality, and topology of how stimuli are encoded across layers of the FNN versus in retinal ganglion cells, primary visual cortex (V1), and the fly visual system model FlyVis.
Each link below opens a fully self-contained, interactive 3-D manifold viewer (no installation required). The viewer shows the encoding manifold (neurons in response space), the decoding manifold (stimuli in population-activity space), and temporal decoding trajectories. You can select subpopulations by region, radius, or functional metric, animate neural activity over the trial, and compare the representational geometry of any chosen subset to the full population.
How to use: Use the Bounding Box, Radius, or Metric tabs to define a neuron subpopulation. Click Update Decoding to recompute the decoding manifold and trajectories for that subset — 10 quantitative metrics then appear below the plots. Use Dec. color to switch decoding plots between stimulus and orientation coloring. Select ▶ Activity (animate) in the Color-by dropdown to animate neural responses frame-by-frame.