Prospective Layout-Guided Multi-Modal Online Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Real-world data often streams, necessitating efficient online multi-modal hashing for fast retrieval and low storage.
• Existing methods struggle with new data classes in streaming scenarios, limiting retrieval performance.
• Addressing the challenge of evolving data classes is crucial for robust online hashing systems.

Purpose of the Study:

• To propose a novel online multi-modal hashing method, Prospective Layout-Guided Multi-modal Online Hashing (PLG-MOH).
• To improve retrieval performance in streaming scenarios, particularly when new data classes emerge.
• To develop a hashing technique that effectively adapts to new data while retaining information from old data.

Main Methods:

• PLG-MOH establishes a Hamming space layout using hashing centers, with each center assigned to a class.
• A prospective layout-guided loss function is introduced, utilizing all hashing centers (assigned and unassigned) for training.
• The method proactively considers future data classes by incorporating unassigned centers into the training process.

Main Results:

• PLG-MOH demonstrates superior retrieval performance compared to state-of-the-art baselines in online scenarios.
• The method effectively adapts hashing functions to new arriving samples, preserving semantic similarity.
• PLG-MOH successfully retains information learned from previously encountered data.

Conclusions:

• The proposed PLG-MOH method significantly enhances retrieval performance in challenging online streaming data scenarios.
• PLG-MOH offers a robust solution for multi-modal hashing when faced with emerging data classes.
• This approach advances the field of online multi-modal hashing by incorporating prospective learning.