Algorithmic aspects of a neuron for coherent wave synapse realizations

We study a neuron functional form (which we call the absolute value neuron) that arises naturally in analog devices in which the neuron synapses are realized using coherent oscillatory wave signals. We discuss algorithmic aspects of this neuron at both the single neuron and network level, including computational capabilities, generalization, and training. A numerical study of absolute value neural networks on two data sets is presented, which demonstrates performance competitive with standard neural networks.