Vitaliy got a PhD in Physics and Mathematics sciences at Karazin Kharkiv National University on the specialty of Mathematical Modeling and Numerical Methods in 2010. From 2006 to 2014 he was engaged in the development of numerical methods in the theory of singular and hypersingular integral equations and modeling the tasks of aerodynamics and radiophysics. Since 2014 he works at Samsung R&D Institute Ukraine on projects for the solution of computer visions and machine learning tasks. The spheres of interests are: convolutional neural networks, integral equations, methods of computer vision.
Topic: Accelerating Convolutional Neural Network Using Low-Rank Regularization
Short Description: Thanks to the redundancy (correlation) among filters in Convolutional Neural Network (CNN), original 4D weight tensors can be approximated by very low-rank basis. Low Rank Approximation methods directly decompose an original large model to a compact model with more lightweight layers. These approaches removes the redundancy in the convolution kernels and accelerate the CNN.