ImageNet Classification with Deep CNNs (AlexNet)
Krizhevsky, Sutskever, Hinton · NeurIPS 2012
TL;DR

AlexNet showed that a deep convolutional network trained on GPUs could shatter the ImageNet benchmark, kicking off the deep-learning era in computer vision.

Key contributions

  • A deep 8-layer CNN trained end-to-end on 1.2M images across 1000 classes
  • ReLU activations that train several times faster than tanh/sigmoid
  • Dropout as a powerful regularizer against overfitting
  • A practical multi-GPU training recipe with data augmentation

How it works

  1. 1Five convolutional layers (some followed by max-pooling) feed three fully-connected layers ending in a 1000-way softmax.
  2. 2ReLU non-linearities keep gradients healthy and speed convergence dramatically.
  3. 3Overlapping max-pooling and local response normalization improve generalization.
  4. 4Dropout in the dense layers and aggressive augmentation (crops, flips, PCA color jitter) combat overfitting on 60M parameters.

Results

  • Top-5 error of 15.3% on ILSVRC-2012 vs. 26.2% for the next best entry
  • A decisive margin that made deep CNNs the obvious path forward

Why it mattered

AlexNet is widely credited with igniting the modern deep-learning revolution. Almost every vision system since builds on the CNN template it popularized, and the ReLU + dropout + augmentation recipe is still standard practice.

Key takeaways

  • Depth + data + compute unlocks representation learning
  • ReLU and dropout are simple ideas with outsized impact
  • Benchmarks can catalyze whole fields when a method wins decisively

Study the topic

Read the original paper