Information Theoretic Learning-Enhanced Dual-Generative Adversarial Networks With Causal Representation for Robust...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning
• Causal Inference

Background:

• Machine and deep learning show promise for intelligent systems but struggle with out-of-distribution (OOD) generalization in applications like smart manufacturing and intelligent transportation systems (ITSs).
• Existing models face limitations in training robustness, particularly when encountering data outside their training distribution.

Purpose of the Study:

• To design and introduce a deep generative model framework that enhances robust OOD generalization.
• To integrate information theoretic learning (ITL) and causal representation learning (CRL) within a dual-generative adversarial network (Dual-GAN) architecture.

Main Methods:

• Developed an ITL- and CRL-enhanced Dual-GAN (ITCRL-DGAN) model.
• Incorporated an autoencoder with CRL (AE-CRL) for causality-inspired feature representations and dual-adversarial training.
• Utilized a feature separation strategy and information theory to build and refine a causal graph, enhancing feature representation with counterfactuals.

Main Results:

• The ITCRL-DGAN model demonstrated superior learning efficiency and classification performance.
• Experimental results on an open-source dataset confirmed outstanding robust OOD generalization capabilities.
• The proposed model outperformed three baseline methods in handling OOD scenarios.

Conclusions:

• The ITCRL-DGAN framework effectively enhances robust OOD generalization in machine learning paradigms.
• The integration of ITL and CRL within a Dual-GAN architecture provides a powerful approach for improving model performance in complex, real-world applications.
• The study highlights the potential of causal inference and information theory for advancing AI in intelligent systems.