Prior Specification Bounds What Spectral Attention Can Learn
Authors:
- Guilherme CoelhoGCGuilherme Coelho
- Clarimar José CoelhoCCClarimar José Coelho
Preprint
- 23 October 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
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.Keywords
This publication has 16 references indexed in Scilit:
- Attention-enhanced residual autoencoder for NIR spectral feature extraction and classification of grain varietiesScientific Reports, 2025
- Deep learning for Raman spectroscopy: Benchmarking models for upstream bioprocess monitoringMeasurement, 2025
- Stop explaining black box machine learning models for high stakes decisions and use interpretable models insteadNature Machine Intelligence, 2019
- Modern practical convolutional neural networks for multivariate regression: Applications to NIR calibrationChemometrics and Intelligent Laboratory Systems, 2018
- Convolutional neural networks for vibrational spectroscopic data analysisAnalytica Chimica Acta, 2017
- Conditional variable importance for random forestsBMC Bioinformatics, 2008
- Performance of some variable selection methods when multicollinearity is presentChemometrics and Intelligent Laboratory Systems, 2005
- Interval Partial Least-Squares Regression ( i PLS): A Comparative Chemometric Study with an Example from Near-Infrared SpectroscopyApplied Spectroscopy, 2000
- A Practical Bayesian Framework for Backpropagation NetworksNeural Computation, 1992
- Standard Normal Variate Transformation and De-Trending of Near-Infrared Diffuse Reflectance SpectraApplied Spectroscopy, 1989