Data-Efficient Image Recognition with Contrastive Predictive Coding

TL;DR

This paper enhances Contrastive Predictive Coding to learn data-efficient image representations, enabling high accuracy with fewer labels and better transfer learning, surpassing supervised methods on benchmarks.

Contribution

It introduces an improved unsupervised learning method that produces representations supporting state-of-the-art classification and transfer learning with significantly fewer labeled data.

Findings

Achieves state-of-the-art linear classification accuracy on ImageNet.

Enables 2-5x reduction in labeled data for non-linear classifiers.

Surpasses supervised pre-trained models in transfer learning for object detection.

Abstract

Human observers can learn to recognize new categories of images from a handful of examples, yet doing so with artificial ones remains an open challenge. We hypothesize that data-efficient recognition is enabled by representations which make the variability in natural signals more predictable. We therefore revisit and improve Contrastive Predictive Coding, an unsupervised objective for learning such representations. This new implementation produces features which support state-of-the-art linear classification accuracy on the ImageNet dataset. When used as input for non-linear classification with deep neural networks, this representation allows us to use 2-5x less labels than classifiers trained directly on image pixels. Finally, this unsupervised representation substantially improves transfer learning to object detection on the PASCAL VOC dataset, surpassing fully supervised pre-trained ImageNet classifiers.