Deep multiblock predictive modelling using parallel input convolutional neural networks | JoVE Visualize

Area of Science:

Chemometrics
Machine Learning
Spectroscopy

Background:

Multiblock data analysis is crucial for integrating information from diverse sources in chemometrics.
Traditional latent space modeling methods like Partial Least-Square (PLS) regression are common, but deep learning (DL) approaches, particularly Convolutional Neural Networks (CNNs), show superior performance in single-block analysis.
The application of DL for simultaneous multiblock data analysis remained unexplored prior to this study.

Purpose of the Study:

To introduce and evaluate a novel deep learning approach for multiblock predictive chemometric analysis using parallel input CNNs.
To compare the performance of the proposed method against single-block CNNs and a conventional multiblock method (SO-PLS).
To assess the efficacy of deep multiblock analysis for data fusion applications.

Main Methods:

Developed a parallel input CNN model where each data block is processed by individual convolutional layers before feature fusion.
Applied the model to a large visible and near-infrared (Vis-NIR) spectroscopic dataset for dry matter prediction in mangoes, treating Vis and NIR data as separate blocks.
Compared the parallel input CNN model against single-block CNNs and sequentially orthogonalized Partial Least-Square (SO-PLS) regression.

Main Results:

The proposed parallel input CNNs based deep multiblock analysis significantly outperformed both single-block CNNs and SO-PLS regression.
Achieved a root mean squared error of prediction (RMSEP) of 0.818%, which was 4% lower than single-block CNNs and 20% lower than SO-PLS.
The deep multiblock approach demonstrated a further ~3% reduction in RMSE compared to the previously best-known results on this specific mango dataset.

Conclusions:

Parallel input CNNs offer a powerful and effective approach for deep multiblock chemometric analysis.
This method demonstrates superior performance in data fusion tasks, outperforming existing single-block and traditional multiblock techniques.
The developed deep multiblock analysis framework is a valuable tool for leveraging multiple data sources in chemometric predictions.