Abstract
Adder neural networks is a new kind of deep learning model in which the
original massive multiplications in convolutions are replaced by
additions. The overall energy consumption using adder networks is
reduced significantly. However, there is an accuracy drop in image
classification task. In this letter, we present a add-wavelet transform
block (AWT) instead of the existing down-sampling operations. Based on
the AWT, we propose a novel adder neural networks (AddWaveNets) to
improve classification accuracy. Experimental results on CIFAR datasets
show that our proposed AddWaveNets achieves significant improvements in
classification accuracy and a powerful ability of feature learning
compared to state-of-the-art quantization networks.