DELBO: Efficient Score Algorithm for Feature Selection on Latent Variables of VAE.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Variational Autoencoders (VAEs) are powerful generative models, but their performance can be limited by suboptimal feature selection in latent spaces.
• Existing feature selection methods often struggle with the complex latent variable structures inherent in VAEs.

Purpose of the Study:

• To develop an efficient feature selection algorithm for VAEs and their variants.
• To enhance VAE model optimization by effectively weighting and selecting important latent variables.
• To extend the proposed method for application in classification tasks.

Main Methods:

• Introduction of the difference of evidence lower bounds (DELBO) framework.
• Development of an efficient score algorithm for latent variable feature selection.
• Proposal of marginalization approximation algorithms for VAE optimization.
• Theoretical analysis for mean-field and full-covariance Gaussian posteriors.
• Extension of DELBO to a generalized version for classification tasks.

Main Results:

• DELBO-based algorithms demonstrate superior performance compared to 9 other feature selection methods across 7 public datasets for generative tasks.
• Experimental validation shows significant improvements in VAE model optimization.
• The generalized DELBO achieved satisfactory results on 5 new public datasets for classification tasks.

Conclusions:

• The proposed DELBO framework offers an effective approach to feature selection in VAEs.
• The method enhances generative capabilities and extends successfully to classification tasks.
• DELBO provides a robust and scalable solution for VAE latent space optimization.