Advantages of a hierarchical system of neural-networks for the interpretation of infrared spectra in structure determination

A hierarchical system of small feed forward neural-networks is used to extract structural information from infrared spectra. The top-level network gives a rough classification in five non-exclusive classes: compounds containing carbonyl, hydroxyl, amino groups, aromatic compounds and ethylenic compounds. For each class, a dedicated network is designed to identify more specific structural features. Depending upon those structural features, the hierarchy might be extended to deeper levels. Specialised networks are activated in a cascade-like effect by the output of the upper-level networks. The training of each specialist network is performed using learning and test sets made of compounds identified by the upper level networks as belonging to this class. Thanks to this approach and to the optimisation of decision thresholds, the quality of the responses is excellent, and compounds wrongly classified by one network do not lead automatically to other errors. One major advantage of this approach is the limited size of each network involved. Networks with few outputs are easier to optimise, and their performance is better than that of larger networks. Moreover linking the response sets from the different refinement levels allows improvement of response quality and in some cases inference of other structural features by combination of responses. Hierarchical neural-network systems are well suited for the interpretation of infrared spectra. They perform very well, and the different refinement levels of information permit great flexibility in the ways they may be used. The modular organisation allows modification of some parts of the system without damaging the whole hierarchy.