|
Abstract
|
Deep learningcontinues to dominate machine learning.It is now widely used in many research areas in science and engineering, and has major industrial impacts. Deep learning methods have achieved remarkable results in a variety of tasks, especially in a supervised learning environment. They have surpassed, or as good as, human in Go, playing video games, accurately identifying objects in imagesand videos, diagnosing certain diseases from medical images, etc.
In this talk, I will start with a brief review of classical (pre deep learning) image restoration methods, followed by some recent applications of deep learning in image restoration and image analysis. I will present my personal understanding of deep learning in image restoration from the perspective of applied mathematics, which inspired two of our recent work on combining numerical differential equation and deep convolutional architecture design. One is to design transparent deep feed-forward convolutional networks to accurately predict dynamics of complex systems and to uncover the underlying hidden PDE models. The other one is to interpret some of the popular deep CNNs, such as ResNet, FractalNet, PolyNet, RevNet, etc., in terms of numerical differential equations; and to propose new deep architectures that can further improve the prediction accuracy of the existing networks in image classification.
|
