Neuron Makers · Part I · Chapter 4
For years, ImageNet scores crept down at the margins. Then in 2012 a neural network from Toronto posted 15.3% top-5 error against the best non-neural runner-up at 26.2% — a drop of nearly eleven points in a single year. After that, the floor fell out, sliding past the human benchmark within a few years.
SOURCE · NEURON MAKERS, CH. 4 (Breakthrough) + Prologue · ImageNet ILSVRC top-5 error. Exact figures cited in the chapter: AlexNet 15.3% vs runner-up 26.2% (2012). Pre-2012 baseline and post-2012 winning scores shown as the documented decline toward the ~5% human benchmark; only the 2012 pair is plotted as exact.