Non-Negative Matrix Factorization Improves Residual Neural Networks

Residual neural networks enable the use of very deep architectures. These architectures benefit by passing the identity information from a layer directly to subsequent layers. Extensive research has been conducted to improve the performance of residual neural networks. In this paper, we propose a method to improve performance by providing a more informative input using non-negative matrix factorization. The method combines the invariant features learned from the training data with the extracted, fine-tuned features at the end of the residual block. Our experimental results confirm that the proposed architecture improves performance on the CIFAR-10 and CIFAR-l00 image classification datasets in terms of test error