Reinforcement of extrapolation of multi-layer neural networks

Multi-layer neural networks have interpolating function that is used for various application fields, i.e. estimations for relationships between chemical compounds and physiological activities. The networks have been a practical tool on the fields, therefore reinforcement of the function is required continuously, and recently extrapolation is also required. For the objectives, we considered some defects on the learning of neural networks, and eliminated them as for neuron functions, symmetry of output from networks, and scaling on the learning data. We tested the effects on typical model calculations. As experiences from the tests, we got useful functions for extrapolation, improvement of interpolations, and detecting a vertex. The introduced techniques are practical, and give high performance calculations for quantitative structure-activity relationships