Conference blog.

I am attending Thirty-seventh Conference on Neural Information Processing Systems (Neurips2023) now and I am happy to share my experience with you all. I will write a short summary every day.

Workshop

Finally, the workshop session has started. It was a very difficult choice for me to pick one particular workshop to attend as many of them are very interesting. At the end, I decided to learn something new (and popular) and attended the workshop on diffusion models.

The talk, which is the highlight of my day, is The Emergence of Reproducibility and Consistency in Diffusion Models. TLDR: Given different diffusion models trained on the same dataset, the same Gaussian noise input would generate almost the same images for all models, meaning the diffusion models are just optimization of interpolations of the iamge distribution and the Gaussian pixel distribution. This kind of reproducibility somehow hints the true nature of diffusion models and perhaps any generative models in the future.