Stochastic Orthogonalization and Its Application to Machine Learning

Orthogonal transformations have driven many great achievements in signal processing. They simplify computation and stabilize convergence during parameter training. Researchers have introduced orthogonality to machine learning recently and have obtained some encouraging results. In this thesis, three new orthogonal constraint algorithms based on a stochastic version of an SVD-based cost are proposed, which are suited to training large-scale matrices in convolutional neural networks. We have observed better performance in comparison with other orthogonal algorithms for convolutional neural networks