Dynamic erasable dye-polymer for all-optical neural networks

Abstract

Thus far, neural network models have mostly been implemented on computers and electronic hardware. Although there is a potential for optics to offer high density parallel interconnects at high speeds, optical neural network implementation has been limited to single layer machines confined to solving only a very narrow range of pattern recognition problems. Multilayer neural networks are much more powerful than the single layer machines, because they represent a class of universal approximators; however, no practical implementation has yet been reported. The primary reason for this slow progress in the optical implementation of neural networks is the lack of suitable materials that can provide dense, modifiable synaptic interconnections. As a result of broad material research efforts in organic/polymer media, we have developed a new dynamic holographic erasable dye-polymer material. Using this material, high efficiency polarization holograms can be recorded, selectively enhanced or erased in real time, so that dense, modifiable neural interconnects can be implemented.