Deep PLS: A Lightweight Deep Learning Model for Interpretable and Efficient Data Analytics.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Deep learning models, despite successes, suffer from interpretability issues, large data needs, complex tuning, and high computational costs.
• These limitations hinder the practical application of advanced neural network-based deep models.

Purpose of the Study:

• To propose a new concept, the lightweight deep model (LDM), designed to mitigate the deficiencies of traditional deep learning.
• To introduce and validate a deep partial least squares (DPLS) model as an instance of LDM.
• To generalize DPLS into a more flexible form (GDPLS) with nonlinear mapping.

Main Methods:

• Constructed a deep partial least squares (DPLS) model, exploring LDM from a new perspective.
• Theoretically proved the feasibility and advantages of DPLS.
• Generalized DPLS to GDPLS by incorporating a nonlinear mapping layer between cascaded PLS layers.

Main Results:

• DPLS and GDPLS demonstrated competitive performance against existing deep models in regression and classification tasks.
• The proposed models offer enhanced interpretability and efficiency compared to traditional deep learning methods.
• The study provides clear insights into how the models improve performance and generate accurate results.

Conclusions:

• LDM, exemplified by DPLS and GDPLS, offers a viable, interpretable, and efficient alternative to fully connected neural networks.
• These models address key limitations of deep learning, paving the way for broader practical applications.
• While not replacing advanced deep vision or language models, LDM provides a valuable alternative for specific use cases.