CaCo: Both Positive and Negative Samples are Directly Learnable via Cooperative-Adversarial Contrastive Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Contrastive learning is a key self-supervised method for unsupervised representation training.
• Effectively selecting positive and negative samples is crucial for learning discriminative encoders and avoiding trivial features.

Purpose of the Study:

• To propose a principled method for end-to-end learning of positive and negative samples in contrastive learning.
• To enhance the discriminative power of encoders by adaptively learning samples.

Main Methods:

• Developed a method for directly learning positive and negative samples simultaneously with the encoder.
• Employed cooperative and adversarial learning strategies by minimizing and maximizing contrastive loss, respectively.
• Ensured samples continuously track query anchor representations across mini-batches.

Main Results:

• Achieved 71.3% and 75.3% top-1 accuracy on ImageNet1K pre-training ResNet-50 over 200 and 800 epochs, respectively.
• Demonstrated strong performance without relying on multi-crop or advanced augmentation techniques.
• Further improved accuracy to 75.7% when utilizing Multi-Crop strategy.

Conclusions:

• The proposed end-to-end sample learning significantly boosts contrastive learning performance.
• This principled approach offers a more robust and effective way to train discriminative encoders.
• The method provides a strong baseline for future research in self-supervised representation learning.