Improving Axial-Attention Network via Cross-Channel Weight Sharing

Resumo

In recent years,
Hypercomplex-inspired neural networks improved deep CNN architectures due to
their ability to share weights across input channels and thus improve cohesiveness
of representations within the layers.
The work described herein studies the effect of replacing existing layers in
an Axial Attention ResNet with their quaternion variants that use cross-channel weight sharing
to assess the effect on image classification.
We expect the quaternion enhancements to produce improved feature maps with more interlinked representations.
We experiment with the stem of the network, the bottleneck layer, and the
fully connected backend, by replacing them with quaternion
versions.
These modifications lead to novel architectures which yield improved accuracy
performance on the ImageNet300k classification dataset.
Our baseline networks for comparison were the original real-valued ResNet, the original quaternion-valued
ResNet, and the Axial Attention ResNet.
Since improvement was observed regardless of which part of the network was modified,
there is a promise that this technique may be generally useful in improving
classification accuracy for a large class
of networks.