Planning forward: Deep incremental hashing by gradually defrosting bits

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep incremental hashing methods face limitations in accommodating new classes due to fixed code lengths.
• Existing approaches are computationally and storage inefficient, especially in early learning sessions with few classes.

Purpose of the Study:

• To introduce Bit Defrosting Deep Incremental Hashing (BDIH) to address the limitations of existing incremental hashing techniques.
• To enable effective accommodation of new classes while maintaining performance on existing ones and improving efficiency.

Main Methods:

• Proposing a bit-defrosting code learning framework with bit-defrosting center generation and center-based code learning.
• Mapping classes into small subspaces by freezing hash bits in initial sessions and progressively expanding subspaces by defrosting bits in subsequent sessions.
• Learning globally discriminative hash codes guided by hash centers while preserving backward compatibility.

Main Results:

• BDIH achieves comparable performance on old classes using fewer bits, reserving more space for new classes.
• The method demonstrates superior retrieval accuracy and storage efficiency compared to existing methods in long-sequence incremental learning.
• Successful reservation of adequate space for future class extensions without compromising existing class performance.

Conclusions:

• BDIH effectively tackles the challenges of accommodating new classes in deep incremental hashing.
• The proposed bit-defrosting strategy offers a more efficient and scalable solution for incremental learning scenarios.
• BDIH significantly enhances both retrieval accuracy and storage efficiency in long-term incremental learning.