LeCun et al. (1998) — DOI-10.1109/5.726791
Key takeaways
- MNIST — 70,000 labelled 28×28 handwritten digits — is introduced here, alongside LeNet-5 and the argument for learning features with gradient descent rather than engineering them
- The dataset outlived the paper’s thesis and then the paper’s difficulty. MNIST has been saturated for a long time; nothing in this record cites an MNIST accuracy number as evidence that a method is good
- It is instead used as a scale at which the expensive thing is affordable: small enough that leave-one-out retraining, exhaustive subsampling and per-example influence can actually be computed
Standing in the anthology
Carries no practice, deliberately — ADR-032.
Twenty-six documents cite it, and the interesting fact is how uniform the
role is. NOTE-181 matches influence functions against actual
leave-one-out retraining on 10-class MNIST — a comparison that exists
because retraining 60,000 times is possible here and nowhere else.
LIT-058 subsampled MNIST and ImageNet and re-ran to test whether the
optimal batch size tracks dataset size — and that experiment is what put
THEORY-013 in the attic, because the dependence it predicted “does not
depend on data set size in any consistent way”. The same experiment is what
SOTA-218 rests on. A dataset you can shrink and retrain is how a scaling
claim gets falsified rather than argued about. NOTE-180 ranks mislabelled
examples by self-influence on CIFAR-10 and MNIST. NOTE-144 uses MNIST and
Fashion MNIST as the near-saturation regime where batch-size differences fall
inside noise.
So the bound this note supplies is not “MNIST is easy”, which every citing
document already knows. It is the more useful one: MNIST results in this
record are almost all existence proofs or ground-truth checks, not
performance claims, and that is a different kind of evidence with a
different failure mode. An influence method validated against retraining on
MNIST has been validated on a small, low-dimensional, nearly-convex problem;
NOTE-181 marks its own claim “strong for convex models” for exactly
this reason, and it is the one that says so most clearly rather than the only
one it applies to.
LIT-058’s 35-workload sweep is the counter-case worth knowing about: there
MNIST and Fashion MNIST are two of seven datasets used as genuine workloads
rather than as a proving ground.
Unread — no NOTE.