Self-Supervised Learning Across Domains

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human learning effectively combines supervised and unsupervised tasks.
• Supervised learning alone is insufficient for comprehensive knowledge acquisition.
• Autonomous learning aids in discovering invariances and improving generalization.

Purpose of the Study:

• To apply a human-like learning approach to object recognition across different domains.
• To improve model generalization and robustness in visual domain adaptation.

Main Methods:

• A multi-task learning model was developed, integrating supervised and self-supervised signals.
• The model learns semantic labels via supervised learning.
• Self-supervised learning on the same images enhances understanding of object shapes and spatial relationships, acting as a regularizer.

Main Results:

• The combined supervised and self-supervised approach yielded competitive results compared to complex domain generalization methods.
• The method demonstrated effectiveness in standard, predictive, and partial domain adaptation scenarios.
• The self-supervised task improved focus on object shapes and part correlations.

Conclusions:

• Combining supervised and self-supervised learning is a powerful strategy for object recognition and domain adaptation.
• This approach offers a more effective and potentially simpler solution than existing complex methods.
• The model shows promise for challenging domain adaptation tasks.