EKConv: Compressing Convolutional Neural Networks with Evolutionary Kernel Convolution

Abstract Convolutional neural networks (CNNs) have achieved tremendous success in visual recognition tasks but mainly rely on massive learnable parameters. To solve this problem, many effective and efficient convolution operators been proposed, such as group-wise convolution, point-wise depth-wise convolution. However, the above operations model optimize weight relationship within same convolutional layer. reduce network parameters, we explicitly construct between kernels of adjacent layers. Specifically, propose an evolutionary kernel namely EKConv, to generate parameters by efficiently. In particular, EKConv makes current layer inherit from its preceding kernel, which promotes information exchange kernels. More importantly, is a novel plug-and-play module that can be easily embedded into mainstream CNNs. Extensive experimental results show compress CNNs large margin yet barely sacrifice image classification performance.