Source
Cammarata et al. (2020), Curve Detectors, LIT-1575, read in full in NOTE-1198: the section “What exactly are we claiming”, and the evidence each claim is assigned.
What is asserted
The article’s four claims, in its words. Causality: “Curve detectors genuinely detect a curve feature, rather than another stimulus correlated with curves.” Generality: they “respond to a wide variety of curve stimuli”, tolerant of radius and “largely invariant to cosmetic attributes like color, brightness, and texture”. Purity: they “are not polysemantic and they have no meaningful secondary function”; weak responses to edges or angles “are a natural extension of the algorithm”. Family: they “collectively span all orientations of curves”.
What the record holds toward it
- Generality and family, measured (THEORY-1028): synthetic curves, invariance to fill and colour, rotation tuning across the ten units.
- Causality, argued: an optimized input is always a curve, and attribution of activating images through the previous layer shows the curves in them. An optimized input shows that curves can drive the unit. That curves are what drive it on natural images rests on the attribution and labelling, for strongly firing cases.
- Purity, argued for the strong cases: diversity-regularized visualizations are all curves, and blind labels sort by activation. For the weak cases, where most firing happens, the authors attribute curve labels’ presence “at least partly” to labelling error and give no precision–recall analysis.
- The mechanism, promised: “We’ll see in the next article that the algorithm used is straightforward and understandable, and we’ve since reimplemented it by hand.” That article is not held.
Why it stays Proposed
Purity is the claim that matters for whether a unit is a feature in THEORY-1013’s sense, and the article itself says only that a secondary function “would need to be rare”. The 45% of 3b:379’s contribution to expected value that is not curves or imperfect curves is not explained.
What would confirm or reject it
The promote_when above.
Note of 2026-10-27: Curve Circuits, read
“What the record holds toward it” says of the article that promises the mechanism: “That article is not held.” It was registered and read on 2026-10-27, at the owner’s request (“file the next readings”), as Curve Circuits, LIT-1629 (NOTE-1207). What it shows is THEORY-1038, and the authors’ claim of a simple, readable algorithm is THEORY-1047, Proposed.
Neither condition of the promote_when is met.
- The weak firings are not accounted for. The hand-written reimplementation is compared with the natural units on synthetic curves, on strong dataset examples and by its weights, by figures only. It is not compared on weak firings, it is grayscale, and the authors expect “many false-negatives”.
- No intervention shows curve use downstream. Substituting the reimplementation for the ten units recovered 26 of the 46 top-5 classifications lost by deleting them, on 2,000 images, a result the authors call preliminary. That is an intervention on accuracy, not one showing that downstream units read the curve units’ output as curve evidence. The downstream use is described, not measured, and its sentence names mixed3b groups as mixed4a’s (NOTE-1207).
The account stays Proposed. Its status and title are unchanged.