Source
Olah, Cammarata, Schubert, Goh, Petrov and Carter (2020), An Overview of Early Vision in InceptionV1, Distill, LIT-1624, read in full in NOTE-1206: the five layer sections, every group’s caption and neuron grid, and the taxonomy files of its Resources section.
What was actually shown
The partition. The article lists, for each of conv2d0 (64 units), conv2d1 (64), conv2d2 (192), mixed3a (256) and mixed3b (480), named groups with the units in each and each group’s share of the layer: 3, 8, 14, 18 and 34 groups. Every unit of each layer appears in exactly one group, and the printed shares agree with the counts to rounding.
The flagged groups. A group is counted as flagged when its caption says it is a catch-all (“Catch-all category”, “Units that don’t fit in another category”, “don’t fall in any other category”), “broad, not very well defined”, “low confidence”, “haven’t studied them very carefully”, “didn’t know what else to do with”, “don’t really understand”, or “likely best thought of as a polysemantic neuron”. By that rule:
| layer | units | flagged | share |
|---|---|---|---|
| conv2d0 | 64 | 9 | 14% |
| conv2d1 | 64 | 31 | 48% |
| conv2d2 | 192 | 14 | 7% |
| mixed3a | 256 | 122 | 48% |
| mixed3b | 480 | 215 | 45% |
| all | 1,056 | 391 | 37% |
The curve family is 10 of mixed3b’s 480 units (2.1%). The fourteen mixed3b groups the article presents as “particularly interesting and well understood” hold 118 units (24.6%).
What could have come out otherwise. Units could have been left out or counted twice, the shares could have disagreed with the grids, and the flagged share could have been small or concentrated in one layer. None of these happened.
What this does not say
- Not that 63% of early vision is understood. An unflagged group is a label given by eye, in the authors’ words often after “a few minutes of cursory investigation”, and the article states no criterion for membership.
- Not that the flagged units are meaningless. A catch-all says the authors have no name for them, not that the units have no function.
- Not about other models or bases. One network, its neuron basis.
- The rule is this record’s, applied to the authors’ words; another reader might count “Gabor-like” (conv2d1, “We haven’t studied them very carefully”) or a borderline caption differently. The flagged list is in NOTE-1206.
Connections
- THEORY-1013. Its promote_when asks for a fraction of units meeting a stated criterion over whole layers. This is the denominator, and the authors’ own uncertainty over it; the criterion is still missing.
- THEORY-1028 and THEORY-1022. The curve family of the census is the one those accounts describe.