Radius Adaptive Convolutional Neural Network

TL;DR

This paper introduces a radius-adaptive convolutional neural network (RACNN) that dynamically adjusts kernel sizes based on input content, achieving higher speeds without increasing model complexity.

Contribution

It proposes a novel adaptive convolution method that learns to select kernel radii dynamically, improving speed while maintaining similar parameter counts.

Findings

Higher inference speeds achieved with RACNN.

Maintains comparable accuracy to standard CNNs.

Code implementation is publicly available.

Abstract

Convolutional neural network (CNN) is widely used in computer vision applications. In the networks that deal with images, CNNs are the most time-consuming layer of the networks. Usually, the solution to address the computation cost is to decrease the number of trainable parameters. This solution usually comes with the cost of dropping the accuracy. Another problem with this technique is that usually the cost of memory access is not taken into account which results in insignificant speedup gain. The number of operations and memory access in a standard convolution layer is independent of the input content, which makes it limited for certain accelerations. We propose a simple modification to a standard convolution to bridge this gap. We propose an adaptive convolution that adopts different kernel sizes (or radii) based on the content. The network can learn and select the proper radius based on the input content in a soft decision manner. Our proposed radius-adaptive convolutional neural network (RACNN) has a similar number of weights to a standard one, yet, results show it can reach higher speeds. The code has been made available at: https://github.com/meisamrf/racnn.