Unsupervised Transfer Learning via Adversarial Contrastive Training

2024-08-16

arXiv.org

AI Summary

To address the bias-sample risk arising from covariance alignment in label-efficient transferable representation learning, this paper proposes the Adversarial Contrastive Training (ACT) framework. ACT is the first method to jointly integrate adversarial perturbations with contrastive learning, providing end-to-end theoretical guarantees under over-parameterized settings and revealing that data augmentation efficiency is the dominant factor governing downstream few-shot generalization error. The method operates in a fully self-supervised manner—requiring no human annotations—and is designed for generic supervised downstream tasks. Extensive experiments across multiple benchmark datasets demonstrate state-of-the-art performance under both linear-probe and k-nearest-neighbor (k-NN) evaluation protocols, with significant improvements in few-shot classification accuracy.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Game Theory and Economic Paradigms: Adversarial Learning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Learning a data representation for downstream supervised learning tasks under unlabeled scenario is both critical and challenging. In this paper, we propose a novel unsupervised transfer learning approach using adversarial contrastive training (ACT). Our experimental results demonstrate outstanding classification accuracy with both fine-tuned linear probe and K-NN protocol across various datasets, showing competitiveness with existing state-of-the-art self-supervised learning methods. Moreover, we provide an end-to-end theoretical guarantee for downstream classification tasks in a misspecified, over-parameterized setting, highlighting how a large amount of unlabeled data contributes to prediction accuracy. Our theoretical findings suggest that the testing error of downstream tasks depends solely on the efficiency of data augmentation used in ACT when the unlabeled sample size is sufficiently large. This offers a theoretical understanding of learning downstream tasks with a small sample size.

Problem

• Learning unbiased representations from unlabeled data
• Addressing biased sample risk in self-supervised learning
• Achieving strong classification with limited downstream labels

Innovation

• Adv-SSL uses adversarial self-supervised learning for unbiased representations
• It employs minimax optimization to create well-separated embedding clusters
• The method achieves strong few-shot learning without additional costs