A code in the nodes
Lynn A. Streeter, Candace Kamm, Yana Kane-Esrig · The Journal of the Acoustical Society of America · 1988
Four two-layer associative networks (one layer of hidden nodes) were trained to classify spectra extracted from naturally produced vowels of one talker, using the hack-propagation learning algorithm [Rumelhart et al., Nature 323, 533–536 (1986)]. The inputs to the networks were either LPC spectra or “perceptual” spectra, obtained by convolving the Bark-transformed spectrum with an asymmetric filter. The networks trained on LPC input classified test vowels more accurately than the networks trained on the perceptually based spectra. The networks' solutions were examined to determine (a) whether it was possible to characterize the features underlying the solutions and (b) whether the input spectral representation affected the nature of the solutions. Principal components analyses of the activation patterns of the hidden units showed that three dimensions accounted for over 80% of the variance for all networks. For networks trained on “perceptual” spectra, the first three dimensions were highly correlated with F2/F1 ratio, F3/F2 ratio, and F3, respectively. For LPC input, the first three dimensions correlated with F1, F3/F2 ratio, and F2/F1 ratio. Thus these networks classified the vowels using features related to the formant space.