Hashing with Mutual Information.

Area of Science:

• Computer Science
• Machine Learning
• Information Theory

Background:

• Binary vector embeddings are crucial for efficient nearest neighbor retrieval in high-dimensional data.
• Supervised hashing aims to learn compact binary codes for effective data representation and retrieval.
• Existing methods may struggle with ambiguity in learned neighborhood structures, impacting retrieval performance.

Purpose of the Study:

• To develop a novel supervised hashing method for learning high-quality binary embeddings.
• To leverage mutual information optimization for reduced ambiguity in the Hamming space.
• To enhance the performance of nearest neighbor retrieval in large-scale databases.

Main Methods:

• Proposing a supervised hashing technique centered on optimizing mutual information.
• Employing deep neural networks and minibatch stochastic gradient descent for optimization.
• Developing a formulation that maximizes the utilization of supervision signals.

Main Results:

• Optimizing mutual information effectively reduces ambiguity in the learned neighborhood structure.
• The proposed method demonstrates superior performance in learning high-quality binary embeddings.
• Experiments on benchmarks including ImageNet validate the method's effectiveness.

Conclusions:

• Optimizing mutual information is a promising direction for supervised hashing.
• The developed method offers significant improvements in nearest neighbor retrieval accuracy.
• This approach is effective for practical applications like image and video retrieval.