Semantic Cluster Unary Loss for Efficient Deep Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Hashing methods map similar data to binary hashcodes for efficient storage and retrieval.
• Deep hashing methods show promise in information retrieval but often suffer from inefficient training due to pairwise or triplet losses.
• Existing methods involve high computational complexity (O(n2) or O(n3)) for training.

Purpose of the Study:

• To propose a novel, efficiently trainable deep hashing algorithm.
• To reduce the computational complexity of deep hashing training.
• To improve the semantic representation of hashcodes for better retrieval.

Main Methods:

• Introduced a Unary Upper Bound of triplet loss, reducing complexity to O(n).
• Developed the Semantic Cluster Deep Hashing (SCDH) algorithm utilizing a Semantic Cluster Unary Loss (SCUL).
• Demonstrated SCDH's extensibility to semi-supervised learning settings.

Main Results:

• The proposed unary loss bridges classification-based unary loss and triplet loss.
• SCDH generates hashcodes forming compact clusters with similar semantic information.
• Experimental results on large-scale datasets show superior performance compared to state-of-the-art hashing algorithms.

Conclusions:

• The novel unary loss significantly enhances the efficiency of deep hashing training.
• SCDH offers an effective approach for semantic information retrieval through clustered hashcodes.
• The method is adaptable to semi-supervised learning, broadening its applicability.