Denoising Autoregressive Representation Learning

Yazhe Li, Jorg Bornschein, Ting Chen

In this paper, we explore a new generative approach for learning visual representations. Our method, DARL, employs a decoder-only Transformer to predict image patches autoregressiv…