Integrating Large Circular Kernels into CNNs through Neural Architecture Search

TL;DR

This paper introduces large circular kernels into CNNs, demonstrating their advantages and integrating them into neural architecture search to improve performance and robustness, inspired by biological visual systems.

Contribution

The paper proposes an efficient implementation of circular kernels, expands NAS search space to include them, and shows their benefits over traditional square kernels.

Findings

Large circular kernels outperform square kernels in CNNs.

Neural architectures with circular kernels are more rotation invariant.

Models with circular kernels are more robust to image transformations.

Abstract

The square kernel is a standard unit for contemporary CNNs, as it fits well on the tensor computation for convolution operation. However, the retinal ganglion cells in the biological visual system have approximately concentric receptive fields. Motivated by this observation, we propose to use circular kernel with a concentric and isotropic receptive field as an option for the convolution operation. We first propose a simple yet efficient implementation of the convolution using circular kernels, and empirically show the significant advantages of large circular kernels over the counterpart square kernels. We then expand the operation space of several typical Neural Architecture Search (NAS) methods with the convolutions of large circular kernels. The searched new neural architectures do contain large circular kernels and outperform the original searched models considerably. Our additional analysis also reveals that large circular kernels could help the model to be more robust to the rotated or sheared images due to their better rotation invariance. Our work shows the potential of designing new convolutional kernels for CNNs, bringing up the prospect of expanding the search space of NAS with new variants of convolutions.