A dimensionality reduction approach for convolutional neural networks

Abstract The focus of this work is on the application classical Model Order Reduction techniques, such as Active Subspaces and Proper Orthogonal Decomposition, to Deep Neural Networks. We propose a generic methodology reduce number layers in pre-trained network by combining aforementioned techniques for dimensionality reduction with input-output mappings, Polynomial Chaos Expansion Feedforward motivation behind compressing architecture an existing Convolutional Network arises from its usage embedded systems specific storage constraints. conducted numerical tests demonstrate that resulting reduced networks can achieve level accuracy comparable original being examined, while also saving memory allocation. Our primary emphasis lies field image recognition, where we tested our using VGG-16 ResNet-110 architectures against three different datasets: CIFAR-10, CIFAR-100, custom dataset.