Hg. Schulze et al., CHARACTERISTICS OF BACKPROPAGATION NEURAL NETWORKS EMPLOYED IN THE IDENTIFICATION OF NEUROTRANSMITTER RAMAN-SPECTRA, Applied spectroscopy, 48(1), 1994, pp. 50-57
We have shown that neural networks are capable of accurately identifyi
ng the Raman spectra of aqueous solutions of small-molecule neurotrans
mitters. It was found that the networks performed optimally when the r
atio of the number of hidden nodes to the number of input nodes was 0.
16, that network accuracy increased with the number of input layer nod
es, and that input features influenced the abilities of networks to di
scriminate or generalize between spectra. Furthermore, networks employ
ing sine transfer functions for their hidden layers trained faster and
were better at discriminating between closely related spectra, but th
ey were less tolerant of spectral distortions than the networks using
sigmoid transfer functions. The latter type of network produced superi
or results where generalization between spectra was required.