Mining Global and Local Semantics From Unlabeled Spectra for Spectral Classification.

Area of Science:

• Analytical Chemistry
• Medical Diagnostics
• Spectroscopy
• Deep Learning

Background:

• Non-destructive detection using molecular vibrational spectroscopy is crucial for analytical chemistry and medical diagnostics.
• Deep learning integration has improved spectral recognition but requires extensive annotated data.
• Current methods' reliance on large labeled datasets limits their broad applicability.

Purpose of the Study:

• To propose a novel self-supervised learning approach, Global and Local Semantics Mining (GLSM), for analyzing unlabeled spectra.
• To overcome the limitations of data-hungry deep learning methods in vibrational spectroscopy.
• To enable effective spectral recognition with minimal annotated data.

Main Methods:

• Developed GLSM, a self-supervised learning framework to capture global and local semantic information from unlabeled spectra.
• Introduced two proxy tasks: global semantic mining (mutual transformation of spectral views) and local semantic mining (noisy spectrum reconstruction).
• Pretrained the model on unlabeled data, enabling fine-tuning with limited labeled data for semi-supervised and transfer learning.

Main Results:

• GLSM effectively captures both global and local semantic information within spectral data.
• The method demonstrates robustness to variations in peak positions and enhances extraction of fine-grained spectral details.
• Experiments on three datasets confirmed GLSM's effectiveness in semi-supervised and transfer learning spectral recognition tasks.

Conclusions:

• GLSM significantly reduces the dependency on large annotated spectral datasets.
• The approach enhances spectral recognition accuracy and robustness.
• GLSM shows substantial potential for real-world applications in spectral analysis.