ph_

Two recent papers explore a common question: What occurs when we train new generative models using data that includes content generated by earlier models? • The Curse of Recursion: Training on Generated Data Makes Models Forget Using model-generated content in training causes irreversible defects. The tails of the original content distribution disappear. Within a few generations, text becomes garbage, and authors call this effect model collapse. • Self-Consuming Generative Models Go MAD[2] The lack of sufficient fresh real data in each generation of a self-consuming loop leads to a progressive decrease in the quality and diversity of future generative models. We refer to this condition as Model Autophagy Disorder (MAD). [1] arxiv.org/pdf/2305.17493v2.p… [2] arxiv.org/pdf/2307.01850.pdf

Just received the hard copy of my book. It's a surreal feeling to finally hold it in my hands. Thanks to everyone who made it possible.🙏

Thankful to my co-authors, @jpatrickhall & James Curtis. Also grateful to Agus Sudjianto for writing the foreword for the book.

📣 Excited to announce that our book on #MachineLearning for High-risk applications has finally reached the finish line 🏁 and it is going into production! It was a mammoth task, but the final draft is now complete. Pre-release on @OReillyMedia buff.ly/3E0xOa8

Also big shoutout to my co-authors, including @jpatrickhall, who made this possible.