Adaptive Convolution Kernel for Artificial Neural Networks

TL;DR

This paper introduces a differentiable method for training convolutional kernel sizes within neural networks, enabling adaptive kernel growth or shrinkage to improve performance across various image tasks.

Contribution

It presents a novel Gaussian envelope-based approach for adaptive convolution kernels that can be trained via backpropagation, enhancing neural network flexibility and accuracy.

Findings

Adaptive kernels outperform fixed-size kernels in image classification tasks.

Replacing standard convolution with adaptive layers improves segmentation performance.

Statistically significant gains observed across multiple datasets.

Abstract

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