Adaptive convolution kernel for artificial neural networks

Many deep neural networks are built by using stacked convolutional layers of fixed and single size (often 3 × 3) kernels. This paper describes a method for learning the kernels to provide varying in layer. The utilizes differentiable, therefore backpropagation-trainable Gaussian envelope which can grow or shrink base grid. Our experiments compared proposed adaptive ordinary convolution simple two-layer network, deeper residual U-Net architecture. results popular image classification datasets such as MNIST, MNIST-CLUTTERED, CIFAR-10, Fashion, “Faces Wild” showed that statistically significant improvements on A segmentation experiment Oxford-Pets dataset demonstrated replacing U-shaped network with 7 improve its performance ability generalize.