QUANTITATIVE COMPARISON OF BIDIRECTIONAL AND OPTIMAL ASSOCIATIVE MEMORIES FOR BACKGROUND PREDICTION OF SPECTRA

Quantitative comparisons of a bidirectional associative memory (BAM), a modified BAM and an optimal associative memory (OAM) neural network are presented for background prediction of infrared (IR) spectra. Thes e memories were evaluated using 2 cm(-1) resolution IR spectra. The ef ficacies of these methods were quantitatively evaluated using root mea n square prediction errors of 100% transmittance lines. In all cases, the OAM performed superiorly to the BAMs. The OAM has no retrieval err or, because it stores patterns that are orthogonal. Binary encoding of spectra is advocated for BAMs, because the stored patterns are approx imately orthogonal. Once the number of grids is large enough to differ entiate stored spectra, the dependence on the number of resolution ele ments disappears. The OAM is a technique that can be applied to any ty pe of data as long as two conditions are satisfied: the background spe ctra and the sample spectra must have points of intersection and the s ignal variations in the sample need to be different from the backgroun d variations.