Content addressable networks for initialization of backpropagation with zero error solutions

Content addressable networks (CANs) are efficient systems trained to perform discrete mappings of arbitrary accuracy. CAN networks learn backpropagation-like associations with an exact, error-free solution, and this solution may be used as a starting point for a backpropagation network by a straightforward conversion. The ability to initialize backpropagation quickly from a CAN solution is provided by the fast convergence rate of CAN networks, the efficiency of the intrinsically discrete CAN network components, the low costs of CAN implementation in optical and VLSI hardware, and the arbitrary accuracy of the zero-error CAN solution. A simulation demonstrating CAN initialization of backpropagation is included