A Scalable Optimization Mechanism for Pairwise Based Discrete Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Maintaining pairwise relationships in high-dimensional data within low-dimensional binary spaces is crucial for efficient data representation.
• Existing methods often struggle with non-convex and non-smooth quartic optimization problems inherent in binary code learning.

Purpose of the Study:

• To develop a novel optimization mechanism for directly solving the quartic problem in pairwise-based hashing.
• To introduce a scalable and smooth algorithm for binary code learning that improves convergence.
• To extend the proposed optimization to various other pairwise hashing algorithms.

Main Methods:

• A novel alternative optimization mechanism is proposed to linearize the quartic problem by introducing a linear regression model.
• A scalable symmetric discrete hashing algorithm is developed, updating binary codes batch by batch sequentially.
• A greedy symmetric discrete hashing algorithm is introduced to further enhance smoothness by updating individual bits.

Main Results:

• The proposed optimization mechanism effectively linearizes and solves the challenging quartic problem.
• Batch-by-batch sequential learning significantly benefits the convergence of binary code learning.
• Extensive experiments show superior performance of the proposed algorithms over state-of-the-art methods in similarity and ranking retrieval tasks.

Conclusions:

• The novel optimization approach and the proposed scalable symmetric discrete hashing algorithms offer a significant advancement in binary code learning.
• The batch-wise learning strategy is key to achieving improved convergence and performance.
• The developed methods demonstrate broad applicability and effectiveness across various pairwise-based hashing algorithms and retrieval tasks.