Should Kernels Be Trained in CNN?- a Paradigm of AG-Net

Abstract

Listen

Translate article

The Convolutional Neural Network constantly updates the weights of kernels to learn the feature representation, which makes the computational cost unaffordable. This work first proposed a Randomized Convolution Kernel with a kernel group to extract the multidimensional feature of each pixel. An AG-Net is then constructed, which can generate a layer containing multiple Gaussian Mixture Models to replace the convolutional layer. There are several Randomized Convolution Kernels in AG-Net to generate several multidimensional feature sets according to different multidimensional features. And each multidimensional feature set gets a Gaussian Mixture Model with Adaptive Resonance Theory. In training, each input is mapped by the Gaussian Mixture Models and the kernel sets. Then a fully-connected layer is used for high-level reasoning. Experiments show that the weights of kernels can be random, and the feature maps based on the similarity of pixels in multidimensional features can be well used in image processing.