Data-Driven Representation, Learning and Applications: From Compressed Sensing to Deep Neural Networks

Compressed Sensing;Deep Neural Networks;Representation;Learning;Data-Driven;Electrical Engineering

Xiong, TaoChin, Sang P. ;

Johns Hopkins University

关键词: Compressed Sensing;    Deep Neural Networks;    Representation;    Learning;    Data-Driven;    Electrical Engineering;   

Others  :  https://jscholarship.library.jhu.edu/bitstream/handle/1774.2/58674/XIONG-DISSERTATION-2017.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: JOHNS HOPKINS DSpace Repository

【 摘 要 】

As the development of high-density sensors, the compressed sensing (CS) and sparse representation have been successfully implemented for building power-efficient systems by reducing sampling, transmission bandwidth and storage capacity. CS-based systems essentially compress the original signals (e.g. neural spikes, images and videos) into a few measurements, and wirelessly transmit them to the receiver for reconstruction and analysis. Therefore, the hidden simplified sparse structureof the original signals from these measurements becomes the key indicator of the performance of such CS-based systems. On the other hand, recent advances in deep learning have dramatically improved the state-of-the-art and enabled significant progress in solving problems such as speech and visual object recognition. Taking advantage of its exceptional performance in solving computer vision tasks, we also exploit and incorporate deep neural networks to extract the inherent and hierarchy representation behind the large scale database. In future work, approaches to bridge these two techniques and more investigations of deep neural network architectures will be studied.In this thesis, we first briefly review the background of compressed sensing, sparse representation and dictionary learning. Based on these theories, we describe a CS-based multi-channel system for neural recordings and spike sorting an a spatiotemporal CS pixel-wise control imaging system. Both systems are evaluated to demonstrate the better performance in terms of reconstruction quality, compression ratio and power efficiency compared to conventional and other state-of-the-art works. Hardware live demonstration for both systems are presented as well. After discussion on compressed sensing and sparse representation, we start to exploit the deep neural networks for human vertebrae localization and identification. From the perspective of deep learning, we propose an automatic and accurate deep image-to-image network with message passing schemes and shape basis learning for human vertebrae detection. Both convolutional and recurrent neural networks are investigated in this work. The proposed network has outperformed the state-of-the-art works on public challenging database. In addition, we also experimentally show the proposed network can thrive on large-scale databases and its extension to other landmark detection applications.