ABSTRACT
Efficient optoelectronic implementations of neural networks can vary widely in complexity, from single-layer linear correlation–cancelation loops to multilayer nonlinear time-delay neural networks. However, they have in common the use of optical processing for the high-dimensional linear operations and electronic processing for the low-dimensional nonlinear operations. Of special importance is the ability to map the highest possible number of linear operations making use of existing optoelectronic devices, thus achieving the highest possible operational throughput before reaching any processing bottleneck. The use of three-dimensional volume holograms is instrumental in order to achieve this goal. Similarly, the use of four-dimensional spectral hole burning is also promising in further expanding the dimensionality (and computational throughput) of future optoelectronic neural processors.
