Shaped-Charge Learning Architecture for the Human-Machine Teams.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Human-Computer Interaction

Background:

• Deep learning (DNNs) and transformers show limitations in human-machine teams.
• Current DNNs lack explainability, clear generalization insights, and robust integration with reasoning techniques.
• DNNs exhibit vulnerabilities to adversarial attacks, hindering their use in collaborative teams.

Purpose of the Study:

• To propose a novel Meta-learning/DNN → kNN architecture.
• To overcome the limitations of stand-alone DNNs in supporting human-machine teams.
• To enhance explainability, interpretability, and robustness against adversarial attacks.

Main Methods:

• Integration of deep learning with explainable kNN (k-Nearest Neighbors) at the object level.
• Implementation of a deductive reasoning-based meta-level control learning process.
• Validation and correction of predictions for enhanced interpretability.

Main Results:

• The proposed architecture provides greater explainability compared to traditional DNNs.
• Improved machinery for integrating with various reasoning techniques.
• Enhanced defense against adversarial attacks in team settings.

Conclusions:

• The Meta-learning/DNN → kNN architecture offers a significant advancement for human-machine teams.
• This approach addresses key limitations of DNNs, promoting more reliable and interpretable collaboration.
• The architecture is analyzed from structural and maximum entropy production viewpoints.