An investigation of adaptive learning implemented in an optically controlled neural network

The authors used a synaptic array of amorphous silicon photoconductors to build a feedforward adaptive neural network. Using backpropagation learning, this network can be taught to perform simple tasks of analog computation. The performance of the network compares well with that of an idealized model, despite significant component variation and externally imposed constraints not accounted for in the learning algorithm. Attempts to calculate the synaptic weight coefficients in simulation and to subsequently implement them on the hardware meet with only limited success, demonstrating the importance of performing the adaptive procedures on the hardware itself. These results, which have much wider applicability to other fabrication technologies such as VLSI show the advantages of adaptation in the design and operation of complex analog systems.<>