A Convolutional Neural Network with a Wave-Based Convolver

In this paper, we demonstrate that physical waves can be used to perform convolutions as part of a state-of-the-art neural network architecture. In particular, we show that the damping of waves, which is unavoidable in a physical implementation, does not diminish their usefulness in performing the convolution operations required in a convolutional neural network (CNN), and the damping only slightly decreases the classification accuracy of the network. These results open the door for wave-based hardware accelerators for CNNs.