Unsupervised Feature Learning via Non-parametric Instance Discrimination

The work investigates whether visual similarity among individual instances can be learned without class labels, extending the observation that supervised classifiers capture apparent category similarity. It formulates feature learning as a non-parametric classification problem at the instance level, asking features to discriminate individual instances, and applies noise-contrastive estimation to overcome the computational cost of having as many classes as training images.

Under unsupervised learning, the method surpasses the state-of-the-art on ImageNet classification by a large margin and consistently improves with more training data and stronger network architectures. Fine-tuning the learned features yields competitive results on semi-supervised learning and object detection. The model is highly compact, storing only 128 features per image (about 600MB for a million images), enabling fast nearest-neighbor retrieval at runtime.