Browse Teach Curate

Discrete Diffusion Models for Text Generation in Deep Learning

Discrete diffusion adapts the denoising-diffusion paradigm to categorical data such as language tokens, defining a forward corruption process over a discrete state space (for example masking or uniform token noise) and learning to reverse it, enabling non-autoregressive generation of text.

Sonny Video

Discrete Diffusion Modeling Using Concrete Score Estimation in Natural Language Processing

Discrete diffusion modeling addresses the challenge of applying continuous Gaussian noise-based diffusion processes to discrete symbolic modalities like natural language by transforming tokens into p…