Suppression helps: Lateral Inhibition-inspired Convolutional Neural Network for Image Classification

Chengyuan Zhuang, Xiaohui Yuan, XUAN GUO

Keywords: Lateral Inhibition, Convolutional Neural Networks

TL;DR: Improving feature learning with lateral inhibition

Abstract: Convolutional neural networks (CNNs) have become powerful and popular tools since deep learning emerged for image classification in the computer vision field. For better recognition, the dimensions of depth and width have been explored, leading to convolutional neural networks with more layers and more channels. In addition to these factors, neurobiology also suggests the widely existing lateral inhibition (e.g., Mach band effect), which increases the contrast of nearby neuron excitation in the lateral direction, to help recognition. However, such an important mechanism has not been well explored in modern convolutional neural networks. In this paper, we explicitly explore the filter dimension in the lateral direction and propose our lateral inhibition-inspired (LI) design. Our naive design incorporates the low-pass filter, while eliminating the central weight to mimic the inhibition strength decay. The inhibition value is computed from the filtering result of the input, with a simple learnable weight parameter per channel for multiplication to decide the strength. Then the inhibition value is subtracted from the input as suppression, which could increase the contrast to help recognition. We also suggest an alternative using depthwise convolution, as a general form. Our design could work on both the plain convolution and the convolutional block with residual connection, while being compatible with existing modules. Without any channel attention along the channel dimension, the preliminary results demonstrate an absolute improvement of 3.68\% and 0.69\% over AlexNet and ResNet-18, respectively, in the ImageNet data set, with little increase in parameters, indicating the merits of our design to help feature learning for image classification.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning