Summary of recent tweets

Nando de Freitas, an AI researcher, has been actively tweeting about various topics recently. In one of his tweets, he discusses the effectiveness of parameterizing neural nets to predict logits and training them using the cross-entropy loss function. He also mentions that this setup works for diffusion models as well by using score interpolation instead of score matching. Nando provides a link to an article for further reference on this topic. Additionally, Nando raises questions about redundancy in language and its role in error correction for learning auto-regressive language models. He expresses his curiosity regarding whether there is a formal proof or experiment supporting the intuition behind this concept. Moreover, Nando shares a tweet where he highlights the importance of parsing tables in PDFs and introduces the use of the Table Transformer model combined with GPT-4V to extract table images and synthesize answers. He provides links to a notebook guide and documentation explaining these techniques. Furthermore, Nando retweets several posts related to recent advancements in AI technology. These include topics such as seamless speech-to-text conversion, energy-based transformers combining attention and associative memory for error correction during inference, open-access video generation models, motion-conditioned image animation for video editing, diffusion models achieving ImageNet linear-probe classification success, and large multi-view datasets showcasing human experts performing skilled activities. Based on these tweets analyzed so far, it seems that Nando de Freitas is actively engaged in exploring various trends and advancements within the field of AI research.