Deep learning

Assembled from 27 blocks · 2 evidence · 35 related

1. Story
2. The family of methods in which many-layered artificial neural networks learn their own representations from data instead of being given hand-designed features. None of the core ideas are recent — backpropagation, convolution and gradient descent were all in the literature by the 1980s — and what changed in the 2010s was not the mathematics but three supplies arriving together: labelled data at a scale nobody had previously assembled, general-purpose graphics processors that made the arithmetic affordable, and training practices that let very deep networks converge at all. The 2012 ImageNet result is the conventional marker for that convergence, and the decade since has been an unusually clean demonstration that a single method family transfers: the same optimisation machinery that classified photographs now predicts protein structure, generates text and forecasts weather. What deep learning does not supply is an account of why any particular trained network behaves as it does, which is why almost every claim about a deep-learning system is empirical rather than derived.
3. Knowledge
4. Deep learning
5. ImageNet
6. The AI accelerator
7. Connections
8. The AI accelerator
9. ImageNet
10. The transformer architecture
11. AlphaFold
12. Machine-learned weather forecasting
13. Machine learning in medical imaging
14. A deep convolutional network wins the ImageNet challenge
15. AlphaGo defeats Lee Sedol in Seoul
16. Deep learning
17. The transformer architecture
18. Training compute
19. Taiwan
20. A deep convolutional network wins the ImageNet challenge
21. Deep learning
22. Fei-Fei Li
23. ImageNet is assembled and released
24. A deep convolutional network wins the ImageNet challenge
25. Evidence
26. An eight-layer convolutional network — five convolutional and three fully connected layers — trained on graphics processors with dropout regularisation achieved a top-5 error rate of 15.3% at ILSVRC 2012 against 26.2% for the second-placed entry, learning hierarchical features rather than using hand-designed ones. V55 verification basis: the paper was not fetched. The 15.3% and 26.2% figures and the architecture description were carried in multiple retrieved summaries and in a hosted copy of the paper itself, but the attribution in those summaries is to secondary explainers as often as to the paper, and a curator should confirm both error rates against the ILSVRC 2012 results table before publication.
27. AI benchmark scores rose rapidly across language, reasoning, coding and mathematics in 2025 while benchmarks themselves faced growing questions about reliability: resolution on SWE-bench Verified rose from about 60% to near 100% within a year; accuracy on OSWorld rose from roughly 12% to 66.3%; frontier models gained about thirty percentage points in one year on Humanity's Last Exam; and robots succeed at only about 12% of real household tasks. V55 verification basis: the report was not fetched, but these figures were returned by a search restricted to hai.stanford.edu and are attributed there to the Index itself. Model-by-model scores, leaderboard ratings and adoption percentages that appeared alongside them in retrieved synthesis were NOT separately confirmed and are excluded from this pack.