Latent Space Map for Visual Utilization of Generated Data.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Data Science

Background:

• Generated Samples (GSs) are crucial for data-centric applications but selecting desirable ones from generative models is challenging.
• Current methods lack guidance for locating specific or high-quality GSs within the vast Generative Latent Space (GLS).

Purpose of the Study:

• To develop a visualization tool, Latent Space Map, for effectively navigating and selecting desirable GSs from a GLS.
• To address challenges in projecting high-dimensional, unbounded GLS and filtering low-quality GSs.

Main Methods:

• A Monte Carlo-based method was used to capture a manifold for projection, focusing on high-quality GSs.
• A manifold projection method was designed to handle the high-dimensional and unbounded nature of GLS, ensuring display accuracy and pattern perception.
• A system integrating Latent Space Map was developed for GS selection and refinement.

Main Results:

• Latent Space Map successfully projects the GLS onto a plane, enabling users to identify regions rich in desirable GSs.
• The projection method achieves high display accuracy and effective pattern perception for user observation.
• The integrated system demonstrates usability and effectiveness in real-world cases and experiments.

Conclusions:

• Latent Space Map provides a valuable tool for domain experts to efficiently select and refine Generated Samples.
• The proposed projection methodology effectively addresses the complexities of visualizing high-dimensional latent spaces.
• This approach enhances the utility of generative models in data-centric applications by improving sample selection.