Simple encoding of infrared spectra for pattern recognition Part 2. Neural network approach using back-propagation and associative Hopfield memory

By extending an adaptive momentum back-propagation two-layer network with a final associative Hopfield memory the networkʹs total error convergence could be improved remarkably. This design enables simultaneous calculations of the networkʹs weights and biases (batch calculating network). Using only energy-orientated inputs of the mid-infrared spectra of 104 multi-functional carbonyl compounds, the networks were trained by 25 structutal 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 networkʹs response to the increase of artificially produced noise to the inputs. Some of the chosen structural features to train the network remain reliable by increasing the disturbance of the input data and can be related to special regions of the original infrared spectra. Therefore the resulting network design could be suitable to verify the reliability of further structural features for classes of organic compounds other than carbonyl compounds.