In the Latest Machine Learning Research, UC Berkeley Researchers Propose an Efficient, Expressive, Multimodal Parameterization Called Adaptive Categorical Discretization (ADACAT) for Autoregressive Models

Autoregressive generative models can estimate complex continuous data distributions such as trajectory rollouts in an RL environment, image intensities, and audio. Traditional techniques discretize continuous data into various bins and approximate the continuous data distribution using categorical distributions over the bins. This approximation is parameter inefficient as it cannot express abrupt changes in density without using a significant number of additional bins. Adaptive Categorical Discretization (ADACAT) is proposed in this paper as a parameterization of 1-D conditionals that is expressive, …

autoregressive models learning machine machine learning machinelearningnews multimodal research researchers uc berkeley