GA-VAE: Enhancing Local Feature Representation in VQ-VAE Through Genetic Algorithm-Based Token Optimization

This paper introduces GA-VAE, a fine-tuning framework that enhances local feature representation in pre-trained Vector Quantized-VAE (VQ-VAE) models through genetic algorithm-based optimization. While VQ-VAE models have shown promise in learning discrete latent representations for complex data distributions, their performance is often limited by underutilized tokens in the codebook, particularly those representing local features, resulting in incomplete feature capture. Our proposed Codebook Genetic Algorithm (CGA) selectively optimizes these underutilized tokens by applying genetic operations to enhance their representation capability while maintaining the well-learned features of frequently used tokens. The framework identifies tokens with low utilization rates and systematically evolves them through carefully designed genetic operations, promoting the emergence of meaningful latent vectors that better capture local features. Experimental results demonstrate that our method effectively improves the representation of local features while maintaining computational efficiency, leading to more balanced codebook utilization and enhanced reconstruction quality across various datasets.