Adaptive Kernel by Layer-wise Convolution
Sahan Ahmad, Aminul Islam · 2023
Convolutional Neural Networks (CNN) have demonstrated outstanding performance in computer vision tasks such as image classification, detection, segmentation, and medical image analysis. Adaptive kernels have been successfully used to gain better performance in CNN-based variants. In general, the optimizer updates the weights of the kernel. Recent research suggests that applying a deterministic adaptive kernel after a convolution operation can help a CNN to gain better generalization. However, how to generate the weights of the deterministic adaptive kernel is still a field of active research. To answer this question, we propose an adaptive kernel method that utilizes convoluted data of a layer to construct a set of dynamic Gaussian kernels. Then we perform the convolution operation on the convoluted data and a randomly chosen Gaussian kernel from the set. To validate the proposed adaptive kernel method, we use eighteen different CNN variants and three different datasets (CIFAR10, CIFAR100, and SVHN) on the image classification task. Our method can be applied to any CNN variant as a plug-and-play, and it does not introduce any trainable parameter to the network. We also provide a detailed analysis of the impact of our proposed adaptive kernel method.