Speaker: Manuel Fernando Mollón Laorca.
Abstract:
Since the publication of Neural Discrete Representation Learning in 2018, Vector Quantized Variational Autoencoders (VQ-VAEs) have gained significant attention for their ability to bridge continuous and discrete representations. In particular, their integration with transformer architectures has opened new ways for modeling complex data types, merging continuous modalities like images or biological signals with discrete data such as text or symbolic sequences. In this talk, we revisit the original VQ-VAE framework, highlight its key innovations, and explore practical ways to incorporate it into modern machine learning pipelines.