Improving Semantic Control in Discrete Latent Spaces with Transformer Quantized Variational Autoencoders

This article discusses a novel model, T5VQVAE, which improves semantic control and generation in Transformer-based Variational AutoEncoders (VAEs). By leveraging the controllability of VQVAEs, T5VQVAE guides the self-attention mechanism in T5 at the token-level. This model outperforms existing VAE models in terms of controllability and preservation of semantic information. It also shows improved inference capabilities, suggesting potential applications for downstream natural language and symbolic reasoning tasks.

Publication date: 2 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.00723