New paper: Reevaluating CNNs for Spectral Analysis – A Focus on Raman Spectroscopy

New paper: Are CNNs truly the best choice for Raman spectroscopy? 🧐

✅ We revisit the inductive bias of CNNs—specifically translational invariance—and its impact on spectral analysis.
✅ We show that pooling parametrization is key to tuning the model’s translational invariance to match instrument stability and spectral variability, giving you a practical trade-off between robustness and resolution.
✅ We show how semi-supervised methods and contrastive pretraining can boost accuracy even with very limited labeled data, a crucial advantage for autonomous deployments.

Findings summarized in our recent ACS Earth and Space Chemistry paper [1].

References

[1]

Reevaluating Convolutional Neural Networks for Spectral Analysis: A Focus on Raman Spectroscopy.

Deniz Soysal, Xabier García-Andrade, Laura E. Rodriguez, Pablo Sobron, Laura M. Barge, and Renaud Detry.

ACS Earth and Space Chemistry, 2025.

DOI

Bib

PDF

@article{soysal2025a,
  author = {Soysal, Deniz and García-Andrade, Xabier and Rodriguez, Laura E. and Sobron, Pablo and Barge, Laura M. and Detry, Renaud},
  title = {Reevaluating Convolutional Neural Networks for Spectral Analysis: A Focus on Raman Spectroscopy},
  year = {2025},
  annote = {doi: 10.1021/acsearthspacechem.5c00136},
  doi = {10.1021/acsearthspacechem.5c00136},
  journal = {ACS Earth and Space Chemistry},
  publisher = {American Chemical Society},
}