Journal
IEEE TRANSACTIONS ON IMAGE PROCESSING
Volume 32, Issue -, Pages 2228-2236Publisher
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TIP.2023.3266169
Keywords
Task analysis; Mutual information; Entropy; Probability distribution; Self-supervised learning; Classification algorithms; Neural networks; Unsupervised learning; image classification
Ask authors/readers for more resources
This paper presents Twist, a self-supervised representation learning method that classifies large-scale unlabeled datasets in an end-to-end manner. The authors use a siamese network with a softmax operation to generate twin class distributions for augmented images. By maximizing the mutual information between input images and output class predictions, Twist avoids collapsed solutions and achieves state-of-the-art performance on various tasks. On the semi-supervised classification task, Twist outperforms previous methods by 6.2% improvement in top-1 accuracy using 1% ImageNet labels with a ResNet-50 backbone. Codes and pre-trained models are available at https://github.com/bytedance/TWIST.
We present Twist, a simple and theoretically explainable self-supervised representation learning method by classifying large-scale unlabeled datasets in an end-to-end way. We employ a siamese network terminated by a softmax operation to produce twin class distributions of two augmented images. Without supervision, we enforce the class distributions of different augmentations to be consistent. However, simply minimizing the divergence between augmentations will generate collapsed solutions, i.e., outputting the same class distribution for all images. In this case, little information about the input images is preserved. To solve this problem, we propose to maximize the mutual information between the input image and the output class predictions. Specifically, we minimize the entropy of the distribution for each sample to make the class prediction assertive, and maximize the entropy of the mean distribution to make the predictions of different samples diverse. In this way, Twist can naturally avoid the collapsed solutions without specific designs such as asymmetric network, stop-gradient operation, or momentum encoder. As a result, Twist outperforms previous state-of-the-art methods on a wide range of tasks. Specifically on the semi-supervised classification task, Twist achieves 61.2% top-1 accuracy with 1% ImageNet labels using a ResNet-50 as backbone, surpassing previous best results by an improvement of 6.2%. Codes and pre-trained models are available at https://github.com/bytedance/TWIST
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available