AlexNet wins the ImageNet ILSVRC 2012 challenge

Background

Image recognition means deciding what object a picture shows. Before 2012 the strongest computer-vision systems usually combined hand-designed feature extractors with a separate classifier, and researchers spent much of their effort designing those features by hand. Neural networks, which learn their own features from data, had been studied for a long time but were not the dominant approach in the field.

To compare methods fairly, the field used shared benchmarks. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) was one of them. The 2012 edition had 1,000 object classes and 1,431,167 annotated images, about 1.2 million of them for training.^[4]

What happened

The winning entry came from a team called SuperVision: Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton of the University of Toronto. Their classification system reached a top-5 error of 15.3%. Top-5 error is the share of test images for which the correct answer was not among the network's five most likely labels. The 15.3% figure used extra training data from the ImageNet Fall 2011 release; with only the supplied training data, their error was 16.4%.^[1]

The timeline was short. Test images were released on 10 July 2012, the submission deadline was 30 September 2012, and full results were released on 13 October 2012. The results were presented at the PASCAL VOC workshop at ECCV 2012 in Florence, Italy, where Fei-Fei Li gave the introduction and overview of results.^[2]

The same team also won the localization task, which requires a system to name an object and draw a box around it. Its error was 33.5%, against about 50.0% for the next team, Oxford's VGG group.^[1]

Key details

The second-best classification entry came from the University of Tokyo team ISI, with a top-5 error of 26.2%. That entry relied on hand-designed features: Fisher Vectors computed over SIFT and other descriptors. These methods summarise local patterns in an image into numbers that people designed, rather than numbers learned from data.^[1]

The winning network had 60 million parameters and 650,000 neurons, arranged in five convolutional layers and three fully connected layers. A convolutional layer slides small learned filters across the image, so early layers respond to simple edges and later layers respond to more complex shapes. The network took raw RGB pixels as input and was trained for five to six days on two NVIDIA GTX 580 GPUs. Graphics processors can carry out many multiplications at once, which made training a network this large practical.^[3]

The team's paper, "ImageNet Classification with Deep Convolutional Neural Networks," was published in the NIPS 2012 (NeurIPS) proceedings.^[3]

Impact

The gap between the winning entry and the runner-up, 15.3% against 26.2%, showed that a large neural network trained on a large labelled dataset with GPUs could far outperform the feature-based methods used by the other entries.^[1]

Because the challenge gave every method the same images and scoring rules, the results could be compared directly. Many researchers later date the start of the deep-learning boom in computer vision to this result. Deep convolutional networks trained on large labelled datasets became a standard recipe for visual tasks, and the approach later spread to other areas of AI.