Discrete Latent Variable

Discrete latent variable models aim to represent complex data using unobserved, categorical variables, improving model interpretability and efficiency in handling high-dimensional or multimodal data. Current research focuses on developing novel training algorithms, such as improved gradient estimation techniques (e.g., adaptive methods and control variates) and incorporating discrete latents into existing architectures like diffusion models, variational autoencoders, and hidden Markov models. These advancements are impacting diverse fields, enabling improved generative modeling, causal discovery from discrete data, and more robust and fair machine learning applications.