Chen-Ping Yu, Co-founder and CEO of Phiar, presents the “Separable Convolutions for Efficient Implementation of CNNs and Other Vision Algorithms” tutorial at the May 2019 Embedded Vision Summit.
Separable convolutions are an important technique for implementing efficient convolutional neural networks (CNNs), made popular by MobileNet’s use of depthwise separable convolutions. But separable convolutions are not a new concept, and their utility is not limited to CNNs. Separable convolutions have been widely studied and employed in classical computer vision algorithms as well, in order to reduce computation demands.
We begin this talk with an introduction to separable convolutions. We then explore examples of their application in classical computer vision algorithms and in efficient CNNs, comparing some recent neural network models. We also examine practical considerations of when and how to best utilize separable convolutions in order to maximize their benefits.