*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Energy Efficient, Secure and Noise Robust Deep Learning for the Internet of Things
Committee:
Dr. Saibal Mukhopdhyay, ECE, Chair , Advisor
Dr. Sudhakar Yalamanchili, ECE
Dr. Tushar Krishna, ECE
Dr. Doug Burger, Microsoft
Dr. Santosh Pande, CoC
Abstract:
The objective of this research is to design an energy efficient, secure and noise robust deep learning system for the Internet of Things (IoTs). The research particularly focuses on energy efficient training of deep learning, adversarial machine learning, and noise robust deep learning for edge devices. To enable energy efficient training of deep learning, the research studies impact of a limited precision training of various types of neural networks like convolutional neural networks (CNNs) and recurrent neural networks (RNNs). For CNNs, the work proposes dynamic precision scaling algorithm, and precision flexible computing unit to accelerate CNNs training. For RNNs, the work studies impact of various hyper-parameters to enable low precision training of RNNs and proposes low precision computing unit with stochastic rounding. To enhance the security of deep learning, the research proposes cascade adversarial machine learning and additional regularization using a unified embedding for image classification and low level (pixel level) similarity learning. Noise robust and resolution-invariant image classification is also achieved by adding this low level similarity learning. Mixture of pre-processing experts model is proposed for noise robust object detection network without sacrificing accuracy for the clean images.
