LeNet-5 read handwritten digits on bank cheques in 1998. It had roughly weights and was trained with saturating activations on a set of around labelled digit images.
Its ideas then waited fourteen years for three things that were missing: enough data, enough compute, and non-saturating activations.
A student now rebuilds LeNet-5 exactly as published, swaps its saturating activations for ReLU, and trains it on the same digit images on a modern GPU. Training finishes in under a minute rather than days — and the accuracy comes out essentially the same as the 1998 result.
The student has all three of the missing ingredients, and the accuracy did not move. What is the right reading of that?
Select all that apply.