nucleation

THEORY-986: Analogous features and circuits form across neural networks trained on different architectures, data and tasks, so that a feature or circuit found in one model can be expected in others, as Olah, Cammarata, Schubert, Goh, Petrov and Carter propose under the name universality

2 min read

Properties4
StatusProposed
Version2
Tagsrepresentation-learning
SourceLIT-1543 — Zoom In: An Introduction to Circuits

Graph View

Source

Olah et al. (2020), Zoom In, LIT-1543, read in full in NOTE-1182: Claim 3.

What is asserted

“Analogous features and circuits form across models and tasks.” In its strongest form, a “periodic table of visual features” observed and catalogued across models; in its weakest, only some lessons transfer.

What the record holds toward it

  • Two low-level feature families (curve and high-low frequency detectors) seen in four vision architectures, a Places365 model and conv nets trained from scratch. Seen, not characterized as the same features.
  • Prior correlational work the essay cites (convergent learning, SVCCA, CKA), which it judges “highly suggestive” but not definitive, since correlated features can differ.
  • Nothing about circuits across models. No circuit is compared between two models.

Why it stays Proposed

The authors’ own words: “the only evidence we can offer today is anecdotal: we simply have not yet invested enough in the comparative study of features and circuits to give confident answers”. They also state the test, which the promote_when takes over.

Connections

  • CLAIM-263 proposes a test of whether two trained networks compute the same way. Universality is the hypothesis that, for analogous models, that test would mostly find shared structure.
  • THEORY-1013 is the companion claim that features and circuits are the units in the first place; universality presupposes it.
  • The essay notes that universality is not needed for circuits research, but would decide which kind makes sense.

Note of 2026-10-27: Curve Detectors’ cross-model evidence

Curve Detectors (LIT-1575, NOTE-1198), read on 2026-10-27, adds one sighting of the kind the promote_when says cannot settle this: “combing”, excitation of curve units by short lines perpendicular to the curve, is seen in the curve units of AlexNet, InceptionV1, InceptionV3 and a Places365 model, and in the optimized stimulus of one macaque V4 neuron, from feature visualizations only. The article defers “prevalence across models” to a later part of the thread. Nothing here compares features or circuits across models rigorously, so the account stays Proposed.

Backlinks