eVAE: Evolutionary Variational Autoencoder.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Variational Autoencoders (VAEs) face challenges with representation inference and task fitting due to surrogate loss.
• Current methods for balancing these aspects involve coefficient tuning, leading to tradeoff uncertainty and inflexible hyperparameters.
• Existing approaches lack dynamic regulation and require manual hyperparameter adjustments, limiting their adaptability.

Purpose of the Study:

• To introduce an evolutionary Variational Autoencoder (eVAE) that dynamically addresses the tradeoff uncertainty in VAEs.
• To integrate variational information bottleneck (VIB) theory with evolutionary neural learning for improved VAE performance.
• To provide a novel evolutionary paradigm for optimizing VAE critical factors and overcoming deep learning integration issues.

Main Methods:

• Developed an evolutionary VAE (eVAE) by integrating a variational genetic algorithm (VGA) into the VAE framework.
• Implemented variational evolutionary operators: variational mutation (V-mutation), crossover, and evolution.
• Designed a training mechanism that dynamically updates the evidence lower bound (ELBO) tradeoff without additional constraints or hyperparameter tuning.

Main Results:

• eVAE effectively addresses the KL-vanishing problem in text generation, achieving low reconstruction loss.
• Demonstrated improved image generation quality, producing sharp images with disentangled factors.
• Achieved superior disentanglement, generation performance, and generation-inference balance compared to competing methods.

Conclusions:

• eVAE offers a synergistic and dynamic approach to managing the learning tradeoff in VAEs.
• The evolutionary paradigm in eVAE mitigates premature convergence and random search issues in deep learning.
• eVAE presents a more balanced and effective solution for VAE tasks, enhancing both generation and inference capabilities.