End-to-end novel visual categories learning via auxiliary self-supervision

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning typically requires extensive labeled data, which is often impractical.
• Existing semi-supervised methods assume overlapping classes between labeled and unlabeled data, limiting real-world applicability.
• Leveraging knowledge across disjoint labeled and unlabeled datasets presents a significant challenge.

Purpose of the Study:

• To develop a semi-supervised learning method for scenarios with disjoint class categories between labeled and unlabeled data.
• To overcome limitations of previous methods, such as multi-phase training and reliance on noisy pseudo-labels.
• To enable effective knowledge transfer and improve model performance in challenging semi-supervised settings.

Main Methods:

• Proposes a novel semi-supervised learning framework utilizing self-supervision as an auxiliary task.
• Enables simultaneous training of all model components for end-to-end learning.
• Incorporates local structure information in feature space for robust pairwise pseudo-label construction.

Main Results:

• Achieved new state-of-the-art performance on CIFAR-10, CIFAR-100, and SVHN datasets.
• Demonstrated the effectiveness of the proposed end-to-end training approach.
• Showcased the robustness of using local structure information for pseudo-labeling.

Conclusions:

• The proposed self-supervised semi-supervised learning method effectively addresses the challenge of disjoint class categories.
• End-to-end training and robust pseudo-labeling significantly enhance learning capabilities.
• The approach offers a promising direction for real-world semi-supervised learning applications.