Krizhevsky (2009) — technical report
Key takeaways
- CIFAR-10 and CIFAR-100 are introduced here as labelled subsets of the Tiny Images collection: 60,000 colour images at 32×32, ten classes with 6,000 each, or a hundred classes with 600 each
- The size is the point, in both directions. It is small enough that a full training run is cheap, which is why it is where methods get tried first; and it is small enough that a result on it is weak evidence about anything at scale, which is why it is where methods get over-claimed
- The report itself is about learning features with restricted Boltzmann machines. The datasets are infrastructure it built along the way, and they outlived the method entirely
Standing in the anthology
Carries no practice, deliberately — ADR-032.
CIFAR-10 appears 144 times across the record and CIFAR-100 eleven, in
sixty-nine documents: twenty-seven literature notes, twenty-eight readings,
twelve practices and two explanations. Only ILSVRC (LIT-423) is
cited more.
What it bounds is the thing the record most often needs to say about a
practice, and the practices already know it. SOTA-222 states that its
claim is pitched at ImageNet scale while its experiments are CIFAR-10.
NOTE-025 records a whole line of gradient-norm work whose evidence is
“VGG16, ResNet18 or ResNet50 on CIFAR-10, CIFAR-100 or TinyImageNet” with
“no transformer and no language result”. SOTA-188 marks its results as
image-scale only. That pattern — the method was developed here, the claim is
made for somewhere else — is common enough that the CIFAR-only
result is close to a recognised evidential category in this record, and it
now has a document to point at.
A source that nothing can check. This is a university technical report:
no arXiv id, no DOI, only a URL, which ADR-009 permits as the last
preference and describes as “a string nothing can check”. That is the honest
state of this citation and the reason it is worth naming — the url here
already redirects once, and the lint cannot tell anyone if it stops
resolving. Among the most-cited instruments in this record, it is the one
with the weakest identifier.
Unread — no NOTE.