DEGREE OF COMPLEXITY - A MAXIMUM-ENTROPY BASED ERROR-MEASURE FOR LEARNING ENDEAVORS IN NEURAL NETWORKS

Neural complex (real or artificial) which is an embodiment of massivel y connected set of neurons represents a cellular automaton ''trained t o learn'' and predict via endeavours managed by a set of protocols inv olving collection, conversion, transmission, storage and retrieval of information. The training or the learning effort is to recognize and c ounter-balance the effects of the cellular disturbances (noise) presen t in the neural system which may tend to disorganize the system's conv ergence towards an objective function (mediated through learning proto cols). The extent of disorganization caused by such disturbances can b e specified by a disorderliness parameter set by the maximum entropy c onsiderations. Such an entropy functional depicts implicitly the degre e of complexity of the system (in spatiotemporal domains) as well. The refore, the disorderliness in the neural complex can be specified by a complexity metric. Using this metric-parameter as an error-measure (o r cost-function), a control strategy (such as the backpropagation base d gradient-descent method) can be developed to train a multilayered pe rceptron. Present study offers relevant algorithmic considerations and simulation results.