Evaluating the Potential of Wavelet Pooling on Improving the Data Efficiency of Light-Weight CNNs

Wavelet pooling (WP) in neural network architectures has recently demonstrated more discriminative power than traditional methods. This is mainly because the latter suffer from spatial information loss while wavelet harnesses of spectral information. However, potential WP increasing data efficiency and extent this have not been investigated yet. Data refers to volume training required attain a certain performance level during inference, e.g., recognition accuracy. In research, we are concerned with evaluating light-weight architectures–MobileNets. Across wide variety seven datasets/applications including object (CIFAR-10, STL-10, CINIC-10, Intel Image Classification datasets) diagnostic imaging (colon diseases, brain tumors, malaria cell images datasets), considering classification accuracy as metric, show that achieves an average saving exceeds 30% compared techniques. For other measures, namely, precision, recall, F1-score, report for datasets 22% datasets. By focusing on architecture, research further emphasizes significance testing resources-challenged settings such applications edge computing green deep learning.