Project 44: Optimization for the Multi-layer Convolutional Sparse Coding
Contact Information
Assoc. Prof. Wenrui Dai
Email: daiwenrui@sjtu.edu.cn
Project Description and Objectives
Convolutional sparse coding (CSC) has been demonstrated to facilitate the representation of high-dimensional visual signals in the tasks of image classification, visual recognition, image reconstruction, and feature extraction. This unsupervised method improves the efficiency of sparse representation by posing a global model with localized dictionaries. Recently, online learning and consensus optimization have been studied to enable scalability in high-dimensional feature learning, but still suffer from a degraded reconstruction performance for multi-layer cases. This project aims to develop a multi-layer dictionary learning method for multi-layer CSC. This optimization method is supposed to guarantee convergence with a bounded approximation error under the varying sparsity requirement for multiple layers. Furthermore, this project also plans to study the connection between multi-layer dictionary learning and deep convolutional neural networks.
Eligibility Requirements
Basic knowledge of signals and systems, digital signal processing, digital image processing, matrix analysis, and optimization theory.
Mastery of more than one programming language, C/C++ and MATLAB preferred.
Main Tasks
Develop a multi-layer dictionary learning method for multi-layer convolutional sparse coding.
Analyze the convergence condition and approximation error of the proposed method.
Establish its connection with deep convolutional neural networks for interpretability.
Website