---
id: imagenet-alexnet
kind: breakthrough
title: ImageNet–AlexNet — Deep Learning's Computer-Vision Inflection
updated: 2026-10-05
sources:
  - id: alexnet-paper
    url: https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
    title: ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky, Sutskever & Hinton, 2012)
    checked: 2026-10-05
  - id: imagenet-challenge
    url: https://image-net.org/challenges/LSVRC/2012/
    title: ImageNet Large Scale Visual Recognition Challenge 2012
    checked: 2026-10-05
  - id: hinton-profile
    url: https://www.cs.toronto.edu/~hinton/
    title: Geoffrey Hinton — University of Toronto profile
    checked: 2026-10-05
  - id: fei-fei-profile
    url: https://profiles.stanford.edu/fei-fei-li
    title: Fei-Fei Li — Stanford profile
    checked: 2026-10-05
related:
  - breakthroughs/transformer.md
  - breakthroughs/alphafold.md
contributors:
  - people/alex-krizhevsky.md
  - people/ilya-sutskever.md
  - people/geoffrey-hinton.md
  - people/fei-fei-li.md
---
## What was blocked
Computer vision had strong hand-engineered pipelines, but learning robust hierarchical features from data at scale remained difficult. Progress was constrained by the interaction among labeled data, model capacity, computation, and comparable evaluation—not by a single missing invention.

## What changed
AlexNet's authors trained a deep convolutional network on ImageNet's 1.3-million-image, 1,000-class training set, using two GPUs, ReLUs, dropout, and data augmentation. At the 2012 ImageNet challenge, the system substantially outperformed competing entries; the paper reports top-5 error of 15.3% versus 26.2% for the second-best entry. The paper authors were **Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton**. [Sources: alexnet-paper, imagenet-challenge]

ImageNet and the ImageNet Large Scale Visual Recognition Challenge, led by a broader research community including Fei-Fei Li's Stanford group, supplied an unusually large labeled dataset and a public benchmark. AlexNet was one team's system and result; benchmark creation and scientific infrastructure are distinct contributions. [Sources: imagenet-challenge, fei-fei-profile]

## Why it mattered — interpretation
The inflection was a system-level convergence: data + GPUs + deep convolutional architecture + training tricks + a competitive benchmark. AlexNet made the performance gap visible enough to shift research and industry attention toward learned representations at scale. It did not originate deep learning, convolutional networks, GPUs, or ImageNet individually.

## Practical leverage
For a learning system, identify whether the constraint is data quality/quantity, compute, architecture, or measurement. A benchmark can coordinate effort and expose progress, but only if its labels and test setting reflect real use. When building a narrow vision feature today, begin with a pretrained model and representative evaluation rather than reproducing AlexNet-era training from scratch.

## Limits and open questions
ImageNet label coverage and taxonomy are imperfect; benchmark performance is not the same as robust perception or deployment value. The 2012 result does not isolate the causal contribution of each ingredient. High compute and large labeled datasets were material advantages. Later models, datasets, and practices changed the field substantially.

## Sources
- [alexnet-paper] Full paper landing page: https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
- [imagenet-challenge] Official 2012 challenge page: https://image-net.org/challenges/LSVRC/2012/
- [hinton-profile] Official profile: https://www.cs.toronto.edu/~hinton/
- [fei-fei-profile] Stanford profile: https://profiles.stanford.edu/fei-fei-li