nucleation

LIT-1575: Curve Detectors

2 min read

Properties6
StatusActive — read 2026-10-27 (NOTE-1198) in full, from the Distill article; worth reading as the first detailed case for one neuron family in a trained vision model, the ten curve units of InceptionV1's mixed3b, argued by feature visualization, blind hand-labelling of dataset examples, attribution, rotation tuning and synthetic stimuli (THEORY-1028). Read it knowing that the dataset analysis is of one unit, 3b:379, that the circuit, the downstream use and the hand-written reimplementation that Zoom In counts among its seven arguments are deferred again to a later article of the thread, and that its claim of purity is not settled here (THEORY-1022).
Version3
Tagsrepresentation-learning, anthology-candidate
Cited as “extends” byLIT-1629 — Curve Circuits
Cited as “paper” byNOTE-1198 — Cammarata et al. — Curve detectors
Cited as “source” byTHEORY-1022 — The curve units of InceptionV1's mixed3b detect curves as such: curves, not correlated stimuli, cause their firing, they respond to curves in general whatever their colour, fill or texture, and they have no meaningful secondary function, their weak responses to edges and angles being part of the same curve-detecting algorithm, as Cammarata, Goh, Carter, Schubert, Petrov and Olah claim, THEORY-1028 — In InceptionV1 trained on ImageNet, the ten curve units of layer mixed3b each respond most to curves of one orientation, with responses invariant to fill and colour and jointly spanning all orientations under rotation of real images; for one of them, 3b:379, which fires on about 11% of dataset positions, activation level separated one author's blind four-class labels of over 800 activation-binned images, and curves and imperfect curves made up 55% of its contribution to expected value

Graph View

Cammarata, N., Goh, G., Carter, S., Schubert, L., Petrov, M. and Olah, C. (2020). Curve detectors. Distill, 17 June 2020. doi:10.23915/distill.00024.003.

Read in full on 2026-10-27 (NOTE-1198). Two accounts were drawn from it: THEORY-1028 (Active) and THEORY-1022 (Proposed).

Why this is here

Registered and read at the owner’s request of 2026-10-27, as one of the works the circuits readings of 2026-10-26 named and did not register. Zoom In (LIT-1543, NOTE-1182) rests its first claim, that units of a trained vision network are meaningful features, on curve detectors, and defers their evidence to “later articles”; THEORY-1013 is Proposed partly for that reason. This is the first of those articles. The record’s circuits line held the programme statement and the transformer work that followed it, but none of the vision case it started from. Read against THEORY-1013’s promote_when, it supplies four of Zoom In’s seven arguments for one family, not a count over a layer: the family is ten of mixed3b’s 480 units.

Note of 2026-10-27: the next article, now held

The status note says the circuit, the downstream use and the hand-written reimplementation “are deferred again to a later article of the thread”. That article, Cammarata, Goh, Carter, Voss, Schubert and Olah’s Curve Circuits, was registered and read on 2026-10-27, at the owner’s request (“file the next readings”), as LIT-1629 (NOTE-1207), and it extends this one. It supplies the three arguments (THEORY-1038), the circuit with one weight edit tested. The thread’s earlier An Overview of Early Vision in InceptionV1, LIT-1624, was read the same day; its taxonomy confirms the curve family as 10 of mixed3b’s 480 units (THEORY-1042). The status is unchanged.

Backlinks