MINIMAL NEURAL NETWORKS - CONCERTED OPTIMIZATION OF MULTIPLE DECISION PLANES

Minimal neural networks (MNNs) differ from other neural networks in that they use localized processing. A MNN has been developed that builds classification trees with branches composed of multiple processing units, as opposed to conventional neural works, which require a network composed of layers of processing units to be configured before training. A global entropy minimization may be achieved at a branch by combining the processing logic using principles from fuzzy set theory. Weight vectors are adjusted using an angular coordinate system and gradients of the fuzzy entropy function. The branches are optimal with respect to fuzziness and can accommodate non-linearly separable or ill-conditioned data. Unlike other neural networks, the mechanism of inference may be traced. In addition, the observation scores, the variable loadings of the weight vectors and the structure of the classification tree yield a bounty of qualitative and diagnostic information. Synthetic bivariate data and pyrolysis-tandem mass spectra are used to compare the MNN with a backpropagation neural network.