Structure is in the zoom: probing neural symmetry through dimensionality scaling
Organizations: ZEHEN Labs, Milwaukee, WI, USA
Abstract
Scaling exponents of neural dimensionality are read as properties of the code, and used to compare representations and benchmark models against cortex. Yet one recording of 11,311 mouse V1 neurons gives 0.25, 0.30 or 0.35, depending only on how it is subsampled. We show that such an exponent measures the population and the probe together. We split it exactly into a floor, which random subsets of the same sizes give, and a shift, which a declared axis of stimulus classes adds. The usual random-subsample exponent is almost all floor. Stimulus direction lifts the exponent in every grating recording by reading both symmetry sectors of the code: orientation, which repeats every half turn, and direction, which does not. The lift lives in trial-to-trial fluctuations that differ between directions, while the class means, which linear decoders read, pull it down. Two measured terms, how much the classes' fluctuations overlap and how the means sit against them, account for the shift. They give its sign where it changes: in Allen Neuropixels recordings, whose class means mostly raise the exponent, and in macaque V1. In an XY lattice whose classes stay on one orbit, temperature flips the sign as the two terms say, and in GPT-2 the shift ranks the indirect-object circuit's name-mover and backup heads. Reported with its path, its floor and its shift, an exponent says how much structure a probe finds and what carries it. Our open tool, rung, reports all three for any labelled neural recording or language-model layer: https://github.com/adilamin89/structure-in-the-zoom
Figures & tables
| What the appendix shows | Main text | |
|---|---|---|
| A | The split is exact for ladders that share rung sizes | Eq. 1 |
| B | Random-stimulus ladders sit at their floor; the floor and the shift’s threshold are sampling crossovers | Sections 2 , 3 |
| C | Per-recording shifts with resampling intervals, permutations, origins and random class orders | Section 3 |
| D | Neighbouring-neuron ladders under both centerings | Section 3 |
| E | Accumulation order: the fluctuation overlaps’ harmonics name the order the fluctuations favour | Section 6 |
| F | Ising and XY lattices: samplers, shifts and the two-term account in 72 ensembles | Section 8 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Type | rec. | Mean | Mean floor % | Mean |
|---|---|---|---|---|
| Drifting gratings | 3 | 0.203 | 97.4% | |
| Local gratings | 3 | 0.284 | 102.1% | |
| Low-contrast gratings | 2 | 0.340 | 101.7% | |
| Static, orientation-biased | 2 | 0.135 | 98.4% | |
| Spontaneous | 8 | 0.555 | 100.7% | |
| All | 18 | 0.381 | 100.3% |
| Recording | Neurons | 95% interval | ||||
|---|---|---|---|---|---|---|
| Drifting GT1 | 20,616 | |||||
| Drifting GT2 | 14,034 | |||||
| Drifting GT3 | 11,311 | |||||
| Local GT1 | 22,482 | |||||
| Local GT2 | 17,750 | |||||
| Local GT3 | 15,640 |
| Recording | Control | |||
|---|---|---|---|---|
| Drifting GT1 | ||||
| Drifting GT2 | ||||
| Drifting GT3 | ||||
| Local GT1 | ||||
| Local GT2 | ||||
| Local GT3 |
| Recording | Order effect | ||
|---|---|---|---|
| Drifting GT1 | |||
| Drifting GT2 | |||
| Drifting GT3 | |||
| Local GT1 | |||
| Local GT2 | |||
| Local GT3 |
| shift kept at | balance at , relative | |||||||
|---|---|---|---|---|---|---|---|---|
| Recording | , | orient. | direct. | random | anat. | orient. | direct. | anat. |
| Drifting GT1 | ||||||||
| Drifting GT2 | ||||||||
| Drifting GT3 | ||||||||
| Local GT1 | ||||||||
| Local GT2 | ||||||||
| Area | Populations | Mean | Positive | Mean |
|---|---|---|---|---|
| VISp | 32 | 32/32 | ||
| VISpm | 20 | 19/20 | ||
| VISam | 26 | 25/26 | ||
| VISal | 24 | 24/24 | ||
| VISrl | 27 | 25/27 | ||
| VISl | 24 | 20/24 |
| Features | (conv1/2/3) | (conv3) |
|---|---|---|
| Plain CNN | / / | |
| Equivariant, pre-pool | / / | |
| Equivariant, invariant | / / |
| direction ladder | order effect | |||||
|---|---|---|---|---|---|---|
| Recording | within-class | class means | total | within-class | class means | |
| Drifting GT1 | ||||||
| Drifting GT2 | ||||||
| Drifting GT3 | ||||||
| Local GT1 | ||||||
| Local GT2 | ||||||
| Recordings (mice) | Neurons or units | Classes | Samples |
|---|---|---|---|
| Stringer et al., two-photon calcium imaging, mouse V1 | |||
| Gratings: drifting 3, localized 3, low-contrast 2 (3) | 11,311–22,565 | 8 directions | 3,278–5,170 |
| Static gratings 2 (2); spontaneous 8 (6) | 9,392–18,982 | ; none | 4,470–5,356; 5,000 |
| Allen Brain Observatory, Neuropixels spike counts, ten visual areas | |||
| 32 sessions (32), 167 area populations | 51–421 | 8 directions | 251–600 |
| Smith and Kohn, Utah arrays, anesthetized macaque V1 | |||
Neural Representation Geometry
Neural representation geometry studies how learned features are arranged and how that structure relates to model behavior and performance.
See all 270 papers