Less Is More: Matched Wavelet Pooling-Based Light-Weight CNNs With Application to Image Classification

We are concerned with the development of effective training strategies for light-weight deep networks, which facilitates harnessing the power of deep learning on hand-held devices. A vast literature has focused on enhancing the performance of deep convolution models (CNNs) by introducing changes in the model architecture and/or developing new training procedures. Recent literature has witnessed a growing focus on pooling layers, one example for this is wavelet pooling. Pooling is often adopted by CNNs to enlarge the receptive field, yet traditional pooling methods cause information loss which undermines further operations such as feature extraction. By capitalizing on spectral information, wavelet pooling (WP) has addressed this drawback in neighborhood pooling, and we have demonstrated its positive impact on light-weight CNN architectures such as MobileNets in previous research. In this paper, we report the following significant observation: Including all sub-bands of the wavelet decomposition does not necessarily yield higher accuracies than using certain sub- set of sub-bands, i.e., less is more. By identifying which sub-bands to be included in the pooling process for every image during training and testing, we propose a new training and inference procedures, namely, Matched Wavelet Pooling (MWP), for light-weight CNN architectures, and in particular, MobileNets. As an example, the image classification application is considered. On two widely adopted datasets, the proposed MWP algorithm consistently outperform previous pooling methods and attained an average of 20% accuracy increase compared to the recent literature.