How AI Learned to See - The ImageNet Collapse (2010-2017)
In 2010, the best machine vision got 1 in 4 images wrong. By 2017 it was 1 in 44 - passing the trained human in 2015. Each drop down the curve is a landmark paper, and together they're the proof-of-concept that scaling deep learning works: the bet that powers the modern AI era.
—credits5sourcesAll episodes
Data sources & references
02Russakovsky et al. 2015 — ImageNet Large Scale Visual Recognition Challenge (IJCV; arXiv:1409.0575)
03Krizhevsky, Sutskever & Hinton 2012 — ImageNet Classification with Deep CNNs (AlexNet, NeurIPS)
04He et al. 2015 — Deep Residual Learning (ResNet; arXiv:1512.03385)
05Karpathy — human top-5 error baseline (~5.1%)