# Alex Krizhevsky > 1986– · Computer Scientist, AI Researcher > > **Recorded contribution:** AlexNet — ImageNet breakthrough (2012); launched the deep learning revolution ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Ukrainian-Canadian computer scientist Alex Krizhevsky built the convolutional neural network later called AlexNet with Ilya Sutskever and Geoffrey Hinton. Its decisive 2012 ImageNet result combined GPUs, rectified linear units, dropout, data augmentation, and a large labeled dataset. Krizhevsky wrote the highly optimized GPU implementation that made the experiment practical. AlexNet's breakthrough was not a single novel layer but a system whose pieces reinforced each other: enough labeled images, a large convolutional network, nonlinear activations, regularization and augmentation, and a GPU implementation capable of finishing the experiment. Krizhevsky's engineering made that conjunction operational. ## 2. The problem inherited Computer vision relied heavily on handcrafted features, while earlier convolutional networks had not yet demonstrated dominant performance at ImageNet scale. ## 3. The central contribution Krizhevsky led the implementation and first authorship of AlexNet, whose ImageNet performance catalyzed modern deep-learning adoption in vision. ## 4. Reconstruct the mechanism 1. Resize and augment labeled images to expand useful variation. 2. Apply learned convolutional filters and ReLU nonlinearities through multiple layers. 3. Train millions of parameters with backpropagation on GPUs and regularize with dropout and weight decay. 4. Evaluate top-k classification error on the held-out ImageNet challenge set. ## 5. What changed downstream - AlexNet's large error reduction redirected computer vision toward deep learned features and GPU training. - It accelerated demand for larger datasets, accelerators, frameworks, and deep-network research. - Its ImageNet margin redirected computer vision toward deep learned features and accelerated investment in GPU software, larger datasets, and end-to-end representation learning. ## 6. Attribution, limits, and uncertainty - The paper has three authors and depends on ImageNet creators, annotators, GPU hardware, convolutional-network pioneers, and software infrastructure. - Image classification accuracy does not imply robustness, causal understanding, or safe deployment; the dataset and labels contain biases and errors. - The result depended on ImageNet's categories and labels, significant compute, and multiple collaborators; classification accuracy did not solve localization, causality, robustness, or social bias. ## 7. Reconstruction lab Train a small convolutional network and a handcrafted-feature baseline on the same image subset. Ablate augmentation and ReLU, and report compute and subgroup errors with accuracy. Ablate augmentation, ReLU, and GPU batching one at a time and report both accuracy and wall-clock cost, revealing how engineering and statistics jointly produced the result. ## 8. Evidence trail - [ImageNet Classification with Deep Convolutional Neural Networks](https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html) — NeurIPS - [Alex Krizhevsky](https://www.cs.toronto.edu/~kriz/) — University of Toronto - [Alex Krizhevsky](https://en.wikipedia.org/wiki/Alex_Krizhevsky) — Wikipedia contributors --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*