SIMPLE ENCODING OF INFRARED-SPECTRA FOR PATTERN-RECOGNITION .2. NEURAL-NETWORK APPROACH USING BACKPROPAGATION AND ASSOCIATIVE HOPFIELD MEMORY

By extending an adaptive momentum back-propagation two-layer network w ith a final associative Hopfield memory the network's total error conv ergence could be improved remarkably. This design enables simultaneous calculations of the network's weights and biases (batch calculating n etwork). Using only energy-orientated inputs of the mid-infrared spect ra of 104 multi-functional carbonyl compounds, the networks were train ed by 25 structural features 104-fold for each of three input sets (19 , 27 and 38 inputs). In a comprehensive statistical investigation the behavior was studied of the network's response to the increase of arti ficially produced noise to the inputs. Some of the chosen structural f eatures to train the network remain reliable by increasing the disturb ance of the input data and can be related to special regions of the or iginal infrared spectra. Therefore the resulting network design could be suitable to verify the reliability of further structural features f or classes of organic compounds other than carbonyl compounds.