This is the conference paper by Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton, all then at the University of Toronto, published in Advances in Neural Information Processing Systems 25 (NIPS 2012). It describes the convolutional neural network subsequently known as AlexNet, and reports the ImageNet results that are conventionally cited as the point at which deep neural networks displaced hand-engineered feature pipelines in computer vision.
Reported results
The paper reports training on the 1.2 million high-resolution images of the ImageNet LSVRC-2010 contest across 1,000 classes, achieving top-1 and top-5 error rates of 37.5% and 17.0%. A variant of the model entered in the ILSVRC-2012 competition achieved a winning top-5 test error rate of 15.3%, against 26.2% for the second-best entry. The NIPS proceedings abstract page states the LSVRC-2010 figures slightly differently, as 1.3 million images with top-1 and top-5 error rates of 39.7% and 18.9%; the paper text carries the 1.2 million / 37.5% / 17.0% figures.
Architecture and training
The network has 60 million parameters and 650,000 neurons, arranged as five convolutional layers, some followed by max-pooling layers, and three fully connected layers ending in a 1,000-way softmax. The paper attributes its speed and accuracy to three elements it treats as the practical contributions:
- Non-saturating neurons. Rectified linear units (ReLUs) in place of saturating nonlinearities, which the authors report train several times faster.
- GPU implementation. A single GTX 580 carries 3 GB of memory, which bounded the network size, so the model was spread across two GTX 580 3 GB GPUs. Training took five to six days.
- Dropout. A then-recent regularization method applied to the fully connected layers to control overfitting, which the paper reports as effective.
The framing of the introduction is that recognition performance is bounded jointly by dataset size, model capacity, and overfitting control, and that datasets had until recently been on the order of tens of thousands of labeled images — a scale at which the capacity of a network like this one could not be used.
Position in the record
The paper is the opening event of the deep-learning era as that period is conventionally dated, and its author list connects it to later frontier-lab work: Sutskever went on to co-found OpenAI and later Safe Superintelligence, and Hinton has since become a prominent voice on AI risk. Its central practical claim — that scaling data, model capacity, and specialized compute together produces discontinuous capability gains — is the empirical antecedent of the argument developed formally in Scaling Laws.
Provenance
Retrieved July 31, 2026 from the NeurIPS proceedings, both the abstract page and the full paper PDF, using the vault's bin/fetch-source.py; Firecrawl was rate-limited during this cycle. Author list, venue, and abstract figures match between the two renderings, with the LSVRC-2010 discrepancy noted above.
Relationships
- related: Geoffrey Hinton, Ilya Sutskever — two of the three authors.
- supports: Scaling Laws — the earliest widely cited empirical instance of capability gains from joint scaling of data, parameters, and compute.
- related: Compute Governance — the GPU-bound training account is the origin point of compute as the binding input to frontier capability.