Reproduce “Rejuvenating image-GPT as Strong Visual Representation Learners” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sucheng Ren, Zeyu Wang, Hongru Zhu, Junfei Xiao, Alan Yuille, Cihang Xie
A framework called D-iGPT rejuvenates autoregressive pretraining for visual representation learning by shifting to semantic tokens and incorporating dual supervision on visible inputs. It achieves 90.0% top-1 accuracy on ImageNet-1K using only publicly available datasets and exhibits superior robustness across various benchmarks.
Key insights
- Autoregressive pretraining, when properly adapted with semantic tokens and dual supervision, can achieve competitive or superior performance to masked image modeling, even at scale.
- Utilizing semantic tokens derived from discriminatively trained models (like CLIP) is critical for both boosting performance and significantly reducing computational pretraining costs compared to raw pixel or VQVAE targets.
- The combination of generative next-token prediction and explicit discriminative supervision on visible inputs provides a synergistic effect, leading to more robust and higher-quality visual representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.