Multilayer Sparse Coding Networks for Image Classification

Abstract

In this proposal, we propose a novel multilayer sparse coding network capable of efficiently adapting its own regularization parameters to a given dataset. The network is trained end-to-end with a supervised task-driven learning algorithm via error backpropagation algorithm. Our proposed multilayer sparse coding consists of a number cascaded composite sparse coding modules. Each module consists of a fat dictionary sparse coding that is applied on a receptive field of an input signal followed by a skinny dictionary sparse coding. The fat dictionary sparse coding is applied to generate a high dimensional sparse representation which is very useful for image classification tasks. The skinny dictionary sparse coding is used to produce a much lower dimensional feature embedding for reducing the computational cost for the subsequent sparse coding layers and still preserving all the energy information from the previous sparse coding layer. Integral to computational efficiency, these skinny dictionaries compress the high dimensional sparse codes into lower dimensional structures. During training, the network learns both the dictionaries and the regularization parameters of each sparse coding layer so that the re-constructive dictionaries are smoothly transformed into increasingly discriminative representations. We also propose to incorporate a new weighted sparse coding scheme into our sparse recovery procedure, offering the system more flexibility to adjust sparsity levels. Integral to computational efficiency, these skinny dictionaries compress the high dimensional sparse codes into lower dimensional structures. The adaptivity and discriminability of our multilayer sparse coding network will be demonstrated on four benchmark datasets, namely CIFAR_10, CIFAR-100, SVHN and MNIST, most of which are considered difficult for sparse coding models.

Document Details

Document Type
DoD Grant Award
Publication Date
Jul 24, 2019
Source ID
W911NF1910375

Entities

People

  • Nasser M. Nasrabadi

Organizations

  • Army Contracting Command
  • United States Army
  • West Virginia University

Tags

Fields of Study

  • Computer science

Readers

  • Computer Programming and Software Development.
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks