Deep learning methods take images as inputs and compute hierarchical representations automatically. For example, traditional convolutional neural networks use convolution and pooling operations to compute features of decreasing sizes in classification tasks. Recent developments in deep learning lead to richer classes of methods for pixel-level predictive and generative image analysis, including generative adversarial networks, variational autoencoders, and U-net. These models require generating features of increasing sizes. While operations for downsizing features have been intensively studied for many years, the operations and challenges for upsizing features in pixel-level image analysis tasks are largely unexplored. In this talk, I will discuss a few limitations and challenges of current methods for pixel-level predictive and generative image analysis. I will then present our work on addressing these challenges using principled, fundamental techniques. I will present case studies on how to use our methods for solving real-world image analysis problems in neuroscience, biology, and medicine.

Bio:
Shuiwang Ji is an Associate Professor in the School of Electrical Engineering and Computer Science at Washington State University. He received the Ph.D. degree in Computer Science from Arizona State University in 2010. His research interests include machine learning, image analysis, and computational neuroscience. Shuiwang Ji received the National Science Foundation CAREER Award in 2014. Currently, he serves as an Associate Editor or Editorial Board Member for ACM Transactions on Knowledge Discovery from Data, IEEE Transactions on Neural Networks and Learning Systems, Data Mining and Knowledge Discovery, and BMC Bioinformatics. He was a program chair for the 2017 Bioimage Informatics Conference and serves as a senior program committee member for IJCAI, KDD, and SDM. He has served as a technical program committee member of major conferences in machine learning (ICML, NIPS), data mining (KDD, SDM, ICDM), and bioinformatics and medical image computing (MICCAI and PSB).