Minimizing Reconstruction Bias Hashing via Joint Projection Learning and Quantization.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Approximate Nearest Neighbor (ANN) search is crucial for high-dimensional data.
• Traditional hashing methods often separate projection and quantization, limiting performance.
• Learning similarity-preserving hash codes faces NP-hard discrete constraints.

Purpose of the Study:

• To propose a novel hashing method that jointly learns projection and quantization.
• To address the limitations of existing two-stage hashing paradigms.
• To develop a method that adaptively balances information loss during hashing.

Main Methods:

• Developed Minimal Reconstruction Bias Hashing (MRH) by jointly optimizing projection and quantization.
• Introduced a lower bound analysis to guide optimization.
• Designed an effective ternary search algorithm to solve the optimization problem.
• Proposed adaptive projection dimensionality adjustment.

Main Results:

• MRH significantly outperforms state-of-the-art methods on eight benchmarks.
• The joint optimization approach effectively minimizes reconstruction bias.
• Adaptive dimensionality adjustment balances information loss between stages.

Conclusions:

• The proposed MRH framework offers a novel perspective on hashing methods.
• Jointly learning projection and quantization is more effective than sequential approaches.
• MRH provides insights for designing loss functions in deep learning models for hashing.