Abstract
Learned per-wavelength weights are widely used to make chemometric neural networks interpretable. A common design passeseach weight through a sigmoid and initialises it from domain knowledge, expecting training to refine the prior against the spectra.We ask which weights the data can inform. The gradient reaching a weight is attenuated by s(1- s), so weights initialised near eitherend of the unit interval are effectively insulated from the data over realistic training budgets; where the gradient does arrive, spectralcollinearity constrains combinations of neighbouring weights rather than individual ones. We analyse a prior-guided spectral gatingmodel on near-infrared and Raman data, with synthetic verification against known ground truth. Over the bands given the maximumprior weight the data supply no measurable curvature — a contraction ratio of unity across the approximations, reference scales,optimisers and fitted solutions examined — while the curvature eigenvectors show the data determining regional combinationsof weights and leaving trades between neighbours free. Neither bound appears in predictive performance: with realistic spectralcorrelation, a prior contradicting the true relevance matches the correct one to within a pre-specified margin of 0.02 in held-out R2while placing 2% of the gate weight on the relevant regions against 72%; region-level permutation shows the network relying onthose regions regardless. Under the training conditions examined, prior specification therefore bounds what spectral attention canlearn, and predictive accuracy cannot validate the explanation. Normalising a prior by its maximum turns a flat top into a block ofinert parameters. Keeping the prior’s weights away from the extremes raises the agreement between the gate and an operationalreference for band relevance — the absolute coefficients of a calibrated partial least squares model, not a chemical ground truth —in every one of five partitions grouped by harvest population; the accompanying advantage in held-out accuracy does not survivethe same grouping, and we report it as specific to the partition on which it was measured.

This publication has 16 references indexed in Scilit: