1D Gradient-Weighted Class Activation Mapping, Visualizing Decision Process of Convolutional Neural Network-Based...

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Deep learning models offer high accuracy in 1D spectroscopy but suffer from low interpretability due to their "black-box" nature.
• Existing visualization methods like CAM and Grad-CAM are designed for 2D data and fail to accurately represent spectral data importance.

Purpose of the Study:

• To develop a novel visualization algorithm, 1D Grad-CAM, for Convolutional Neural Network (CNN)-based models in 1D spectroscopy.
• To improve the interpretability of deep learning models for qualitative and quantitative spectral analysis.

Main Methods:

• Developed 1D Grad-CAM by modifying classical Grad-CAM, removing gradient averaging (GAP) and ReLU operations.
• Introduced "difference" (purity/linearity) and "feature contribute" metrics for evaluating model performance.
• Applied the algorithm to analyze vegetable oil adulteration using Raman spectroscopy and ResNet.

Main Results:

• 1D Grad-CAM demonstrated a stronger correlation between gradients and spectral locations, capturing spectral features more comprehensively.
• The algorithm enabled reliable evaluation of qualitative accuracy and quantitative precision of CNN models.
• Visualization of ResNet for vegetable oil analysis confirmed the method's effectiveness in reflecting high accuracy and precision.

Conclusions:

• 1D Grad-CAM provides clear insights into CNN decision-making processes for 1D spectral data.
• The developed algorithm enhances the interpretability and reliability of deep learning models in spectroscopy.
• 1D Grad-CAM facilitates broader applications of CNNs in the 1D spectroscopy field.