Use of symmetric kernels for convolutional neural networks

TL;DR

This paper introduces symmetric convolutional kernels in CNNs to achieve invariance to flips and rotations, acting as regularizers that enhance generalization but complicate training.

Contribution

It proposes new symmetric kernels for CNNs that provide invariance to flips and rotations, improving regularization and generalization.

Findings

Symmetric kernels induce invariance to flips and rotations.

Using these kernels acts as a regularizer, improving generalization.

Training with symmetric kernels is more complex but yields better performance.

Abstract

At this work we introduce horizontally symmetric convolutional kernels for CNNs which make the network output invariant to horizontal flips of the image. We also study other types of symmetric kernels which lead to vertical flip invariance, and approximate rotational invariance. We show that usage of such kernels acts as regularizer, and improves generalization of the convolutional neural networks at the cost of more complicated training process.