Incorporating Horizontal Connections in Convolution by Spatial Shuffling

Ikki Kishida, Hideki Nakayama

TL;DR: We propose spatially shuffled convolution that the regular convolution incorporates the information from outside of its receptive field.

Abstract:

Convolutional Neural Networks (CNNs) are composed of multiple convolution layers and show elegant performance in vision tasks.
The design of the regular convolution is based on the Receptive Field (RF) where the information within a specific region is processed.
In the view of the regular convolution's RF, the outputs of neurons in lower layers with smaller RF are bundled to create neurons in higher layers with larger RF.
As a result, the neurons in high layers are able to capture the global context even though the neurons in low layers only see the local information.
However, in lower layers of the biological brain, the information outside of the RF changes the properties of neurons.
In this work, we extend the regular convolution and propose spatially shuffled convolution (ss convolution).
In ss convolution, the regular convolution is able to use the information outside of its RF by spatial shuffling which is a simple and lightweight operation.
We perform experiments on CIFAR-10 and ImageNet-1k dataset, and show that ss convolution improves the classification performance across various CNNs.

Code: https://github.com/AccountForSubmission/ICLR2020_Spatially_Shuffled_Convolution

Keywords: shuffle, convolution, receptive field, classification, horizontal connections