Tensor based machine learning with human computer interaction applications

In this research, first we will using tensor network techniques speedup the processing in neural network since computational cost is a major bottleneck in neural network based deep learning. Note that the weight matrix in each layer of neural network is with huge size, because of millions of parameters. This may cause large time complexity to calculate all the parameters. To overcome the large time complexity, we could use tensor decomposition to calculate the low-rank weight matrix, to reduce a large amount of parameters, therefore to reduce the time complexity. The tensor networks include Hierarchical Tucker (HT), Tensor Train (TT), CP, and Tucker decomposition etc. Second, we propose to develop tensor based algorithms to solve spatiotemporal corrupted problem, complex background problem in human action recognition (HAR) and denoising problems in medical images. The human action videos are represented as a high-dimensional tensor, and we can solve the spatiotemporal corrupted problem by tensor completion, and recurrent neural network to deal with the missing frames problem; also we could denoise the high-dimensional medical images by tensor completion.

Faculty Supervisor:

Xiao-Ping Zhang

Student:

Chengcheng Jia

Partner:

Huawei Canada

Discipline:

Engineering - biomedical

Sector:

Professional, scientific and technical services

University:

Ryerson University