Practical realization of any computerized Optical Multichannel Spectra Analyzer (OMA) implies that rather sophisticated software must be developed to process the data obtained from AD converter, connected to the optical detector (usually based on linear CCD or an array of photo diodes). Neural network (NN) may be regarded as a modern and beneficial alternative to that common approach, because NN is traditionally considered as a good instrument to solve the problems of Pattern Recognition and Signal Analysis. Unfortunately all the necessary parameters for neural network setup could not be easily calculated in a formal way. Besides, most of the OMA-type units are supplied with the detectors containing many hundreds of diode cells to gain high spectral (or spatial) resolution. It means that a NN under consideration must be able to process a great deal of inputs. The problem ofNN design is additionally complicated by the necessity to keep its functionality under condition of various time dependent discrepancies, drifts, noise etc. Some approaches to the practical realization ofNN-based OMA and results achieved are presented.