Compact SVD Based Representation of CNN Kernels for Classification and Time Complexity Analysis

Compact Singular Value Decomposition based low-rank representation of Convolutional Neural Network(CNN) kernels is explored in this work. This representation allows us to reduce the number of computations at the cost of very less change in the accuracy of deep learning models. This work attempts to analyze the extent of benefit achieved by representing kernels using Compact SVD and quantify the change in output accuracy. This work utilized a standard MNIST dataset for analyzing the test accuracy and computation requirements. The benefits of such representation can be extended for various CNN based deep learning models as future works.