Recommendation to Enrich Computer Vision Part Content with Latest Research

[frinkleko]

Thank you so much for making such wonderful slides.

I am here to recommand to enrich Module 4, Chapter Computer vision, Page 141-153, by incorporating insights from the newly published paper titled "Scalable Pre-Training of Large Autoregressive Image Models.

This paper follow the inital idea of iGPT, which is use an autoregressive objective to pre-train vision transformers on image patches, without any supervision or labels but using pixel-level regression loss.

Moreover, AIM (model proposed in this paper) introduces two architectural modifications - prefix attention and MLP prediction heads and discussed correlation between the pre-training objective and the downstream performance.

I think this could be great additional information for the iGPT part. If possiable, I can help you with making the slides.