ÇмúÇà»ç

 

 

Recently, score-based models, and denoising diffusion probabilistic models (DDPMs) have gained wide interest as a new class of generative model that achieves surprisingly high sample quality without adversarial training. Among many works, Song et al. (2021b) generalized discrete score-matching procedures to a continuous stochastic differential equation (SDE), which in fact also subsumes diffusion models into the same framework. Score-based models perturb the data distribution according to the forward SDE by injecting Gaussian noise, arriving at a tractable distribution (e.g. isotropic Gaussian dis- tribution). In order to sample from the data distribution, one can train a neural network to estimate the gradient of the log data distribution (i.e. score function, ∇x log p(x)), and use it to solve the reverse SDE numerically. However, diffusion models are inherently slow to sample from, needing few thousand steps of iteration to generate images from pure Gaussian noise. In this tutorial, we review the basic mathematical principles of score-based diffusion models and its applications to various computer vision tasks, such as image synthesis, translation, inverse problems, etc. Then, recent research activities to accelerate the sampling process are reviewed.

 

Jong Chul Ye is a Professor of the Kim Jaechul Graduate School of Artificial Intelligence (AI) of Korea Advanced Institute of Science and Technology (KAIST), Korea. He received the B.Sc. and M.Sc. degrees from Seoul National University, Korea, and the Ph.D. from Purdue University, West Lafayette. Before joining KAIST, he worked at Philips Research and GE Global Research in New York. He has served as an associate editor of IEEE Trans. on Image Processing, and an editorial board member for Magnetic Resonance in Medicine. He is currently an associate editor for IEEE Trans. on Medical Imaging, and a Senior Editor of IEEE Signal Processing Magazine. He is an IEEE Fellow, was the  Chair of IEEE SPS Computational Imaging TC,  and IEEE EMBS Distinguished Lecturer. He was a General Cochair (with Mathews Jacob) for IEEE Symp. On Biomedical Imaging (ISBI) 2020. His research interest is in machine learning for biomedical imaging and computer vision. 

Abs: 

Deep learning has produced significant breakthroughs in several fields such as computer vision, natural language processing, and speech recognition. Such research is based on neural networks designed for Euclidean-structured data such as images, text, or acoustic signals. In this tutorial, I will introduce geometric deep learning (GDL): a framework to generalize neural networks for non-Euclidean structured data such as graphs and manifolds. I will also introduce how GDL helps to solve drug discovery as a framework for handling molecular graph and protein 3D structures. In particular, I will briefly introduce using GDL for molecular property prediction, molecular design, retrosynthesis, and protein structure prediction (such as AlphaFold) 

 

Bio: