Button to scroll to the top of the page.

Summer 2019 Colloquium: Graduate Portfolio in Scientific Computation






Yu Lu TBD TBD TBD "Deep Visualization in Unsupervised Convolutional Neural Networks"

Yu Lu

Title: "Deep Visualization in Unsupervised Convolutional Neural Networks"

Abstract: Convolutional neural networks has brought the recent fast develops to computer vision, deep learning, and computational sciences. Deeper networks with multiple convolutional layers and millions of parameters are used for object recognition, object detection, semantic segmentation, etc. Unsupervised learning with deep convolutional networks is also under very active studies. Although significant unsupervised segmentation results have been achieved from deep convolutional networks like W-Net, the efficiency and data flow between layers has merely been investigated. A visual understanding of working mechanism of these networks will not only help improve their efficiency, but also inspire new approaches. In this study I show the attempt towards deep visualization in multi-layer convolutional neural networks when used for unsupervised learning tasks with images. Specifically, a W-Net is investigated in terms of the gradient update during the backpropagation, input/output of each stage, and the loss function used for segmentation. Additionally, sparsely sampled input image is considered in similar segmentation problems to improve efficiency and reduce memory consumption.