Experimental demonstration of large-scale holographic optical neural network

The authors present an experimental implementation of the N⁴ holographic optical neural network and demonstrations of the potential capabilities of large-scale and high-speed pattern recognition applications of this architecture. The extremely large space-bandwidth-product of the holographic materials makes it possible to construct large-scale static holographic optical neural networks (HONNs). A 1024-neuron HONN has been experimentally demonstrated for pattern recognition operations. A 32×32 optical neural network was trained to recognize different airplanes and to give the correct names at the output detector array. The computational time for training the four pairs of patterns took only a few minutes on a 386 microcomputer. It is estimated that with proper selection, over 200 pairs of heteroassociative patterns can be stored in the neural network of 1024 neurons for target recognition using the interpattern association (IPA) neural network model (T. Lu et al., 1990). By recording more holograms on a holographic plate, a 128×128 neural network and eventually a 256×256 neural network can be constructed, thus a high-speed real-time recognition of an enemy´s aircraft can be performed