Design and analysis of fixed planar holographic interconnects for optical neural networks

Abstract

A neural-network architecture of multifaceted planar interconnection holograms and optoelectronic neurons is analyzed. Various computer-generated hologram techniques are analyzed and tested for their ability to produce an interconnection hologram with high-accuracy interconnects and high diffraction efficiency. A new technique is developed by using the Gerchberg-Saxton algorithm, followed by a random-search error minimization that produces the highest interconnect accuracy and the highest diffraction efficiency of the techniques tested. Analysis of the system shows that the hologram has the capacity to connect 5000 neuron outputs to 5000 neuron inputs with bipolar synapses and that the encoded synaptic weights have an accuracy of ~5 bits. A simple feedback system is constructed and demonstrated.