Use of symmetric kernels for convolutional neural networks
This work addresses the need for more robust and generalizable CNNs in computer vision, but it is incremental as it builds on existing kernel-based methods.
The authors tackled the problem of improving generalization in convolutional neural networks by introducing symmetric kernels that enforce invariance to horizontal flips, vertical flips, and approximate rotations, resulting in better generalization as a regularizer, though with a more complicated training process.
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.