Masked Unsupervised Self-training for Zero-shot Image Classification

State-of-the-art computer vision models are mostly trained with supervised learning using human-labeled images, which limits their scalability due to the expensive annotation cost. While self-supervised representation learning has achieved impressive progress, it still requires a second stage of finetuning on labeled data. On the other hand, models pre-trained with large-scale text-image supervision (e.g., CLIP) have enabled zero-shot transfer to downstream image classification tasks. However, the zero-shot performance of CLIP-like models are often insufficient for real-world adoption. In this paper, we aim to leverage the abundant unlabeled data to improve the performance of a pre-trained zero-shot classifier on downstream tasks. We propose Masked Unsupervised Self-Training (MUST), a new approach which leverages two different and complimentary sources of supervision: pseudo-labels and raw images. MUST jointly optimizes three objectives to learn both class-level global feature and pixel-level local feature and enforces a regularization between the two. We demonstrate the efficacy of MUST on 8 downstream tasks across a variety of domains, where it improves upon CLIP by a large margin and narrows the performance gap between unsupervised and supervised classification. For instance, MUST achieves a zero-shot top-1 accuracy of 77.7% on ImageNet using ViT-B, +9.4% higher than CLIP. Our code is available at https://github.com/salesforce/MUST.