Source
Cammarata, Goh, Carter, Schubert, Petrov and Olah (2020), Curve Detectors, Distill, LIT-1575, read in full in NOTE-1198: Feature Visualization, Dataset Analysis, Visualizing Attribution, Human Comparison, Joint Tuning Curves, Synthetic Curves and Synthetic Angles. The figures were not rendered; the numbers are those the text states.
What was actually shown
The units. The ten units of mixed3b whose feature visualizations are curves. Every optimization for maximal firing, with or without a diversity term, gives a curve.
Orientation and the family. Each unit’s top dataset examples, rotated in 1° steps, drive that unit in a bump around one orientation, and as one unit stops firing the next starts, so that the ten together cover 360°. The article notes gaps, partly filled by double-curve units.
Synthetic stimuli. Rendered curves varied in orientation and curvature drive 3b:379 up to 24 standard deviations above its mean dataset activation, most for curves like its visualization. The response is wider in orientation at higher curvature, and unchanged by filling or colouring the shape. Angles drive it where an arm is tangent to its preferred curve.
One unit on the data. 3b:379 fires (pre-ReLU above zero) on 11% of receptive-field crops. One author labelled over 800 images drawn at random in activation bins, seeing pixels only: curve, imperfect curve, unrelated, opposing curve. The labels sort by activation level. Weighting each level by x·p(x), curves and imperfect curves account for 55% of the unit’s contribution to expected value.
What could have come out otherwise. Visualizations could have been textures or varied by seed. Rotated examples could have left the unit unmoved, or the family could have left orientations uncovered. Fill or colour could have mattered. Blind labels could have been spread evenly across activation levels. Strongly firing images could have been clocks for reasons attribution would show were not curves.
What this does not say
- Not that the units are pure. Most of 3b:379’s firings are weak, most images labelled curves fall where it fires weakly or not at all, and the other 45% of its contribution to expected value is not accounted for. The authors’ claim of purity is THEORY-1022.
- Not how the units are computed. The circuit and the hand-written reimplementation are in a later article, not read.
- Not about other models or bases. One network, its neuron basis.
- Not a count. Ten of mixed3b’s 480 units; nothing here says how much of the layer is understood, which THEORY-1013’s promote_when requires.
- The units are not robust. Small changes in orientation or curvature swing activations by several standard deviations.
Connections
- THEORY-1013. The worked example that account cites, measured. It supplies four of Zoom In’s seven arguments for one family, and none of the count the account’s promote_when asks for.
- THEORY-991. In a one-layer transformer the neurons were the uninterpretable basis and dictionary features the interpretable one. This family is a case where the neuron basis itself is clean; the article does not ask how typical that is.
Note of 2026-10-27: how the units are computed, now held
“What this does not say” puts the circuit and the hand-written reimplementation “in a later article, not read”. That article, Curve Circuits, was registered and read on 2026-10-27, at the owner’s request (“file the next readings”), as LIT-1629 (NOTE-1207). How these ten units are computed from mixed3a, the echo edit and the reimplementation are THEORY-1038, a separate account that this one leaves to it. An Overview of Early Vision in InceptionV1, LIT-1624, read the same day, confirms that the family is 10 of mixed3b’s 480 units (THEORY-1042). Nothing in either contradicts what is measured here. The title and status are unchanged.