Cascaded Subpatch Networks for Effective CNNs

TL;DR

This paper introduces Cascaded Subpatch Networks (CSNet), a novel CNN architecture that uses smaller subpatch filters in a cascaded manner to enhance feature representation and achieve state-of-the-art results on CIFAR10.

Contribution

The paper proposes a new subpatch filter design and a cascaded network structure that improves feature extraction and reduces parameters compared to traditional CNNs.

Findings

CSNet achieves 5.68% error on CIFAR10, the best to date.

The proposed method demonstrates superior effectiveness and compactness.

Experimental results on four datasets validate the approach.

Abstract

Conventional Convolutional Neural Networks (CNNs) use either a linear or non-linear filter to extract features from an image patch (region) of spatial size $ H\times W $ (Typically, $ H $ is small and is equal to $ W$, e.g., $ H $ is 5 or 7). Generally, the size of the filter is equal to the size $ H\times W $ of the input patch. We argue that the representation ability of equal-size strategy is not strong enough. To overcome the drawback, we propose to use subpatch filter whose spatial size $ h\times w $ is smaller than $ H\times W $. The proposed subpatch filter consists of two subsequent filters. The first one is a linear filter of spatial size $ h\times w $ and is aimed at extracting features from spatial domain. The second one is of spatial size $ 1\times 1 $ and is used for strengthening the connection between different input feature channels and for reducing the number of parameters. The subpatch filter convolves with the input patch and the resulting network is called a subpatch network. Taking the output of one subpatch network as input, we further repeat constructing subpatch networks until the output contains only one neuron in spatial domain. These subpatch networks form a new network called Cascaded Subpatch Network (CSNet). The feature layer generated by CSNet is called csconv layer. For the whole input image, we construct a deep neural network by stacking a sequence of csconv layers. Experimental results on four benchmark datasets demonstrate the effectiveness and compactness of the proposed CSNet. For example, our CSNet reaches a test error of $ 5.68\% $ on the CIFAR10 dataset without model averaging. To the best of our knowledge, this is the best result ever obtained on the CIFAR10 dataset.