Higher Order Adaptive Networks - Some Aspects of Multi-Class and Feed-Back Systems

Single layer networks of adaptive logic functions have been used extensively for pattern recognition. They do, however, have limitations in as far as the output at a given time 't' is a function of the input at that time only. The system is equivalent to a conventional combinational logic circuit. If some form of feed-back is introduced, the output becomes dependent not only on the current input pattern, but also on all previous patterns input to the system. The network then has sequential properties. Two forms of higher order system are presented in this paper. The histogram of the output responses is recognised by a second layer network. This achieves an effective increase in discrimination and can correct substitution errors on the first layer. The second system involves feed-back of a pattern generated at the ouput of the net. These output patterns are prototypes which the system has been trained to recognise. The feed-back leads to convergence in pattern space and a consequent reduction in the variability within a given pattern class. This behaviour has implications in the area of pattern recognition, associative memory and image recall systems.