US20250278669
2025-09-04
Physics
G06N20/00
The patent application introduces a system for generating counterfactual explanations in artificial intelligence (AI) models. It focuses on enhancing the explainability of complex machine learning models, such as deep neural networks (DNNs), by providing actionable insights. Counterfactual explanations help identify minimal changes needed in input data to alter the model's output, aiming for more desirable results.
A key component of the system is the Counterfactual Engine, which detects features of a machine learning model and computes objectives based on modifying weights and perturbations of these features. This process involves transforming counterfactuals to achieve consistent and diverse explanations, enabling better understanding and manipulation of AI model outcomes.
The system can be implemented in various forms, including a computer program product or a computer system with specific hardware components. The Counterfactual Engine uses weights that penalize feature perturbations, ensuring the changes are minimal yet effective. The engine iteratively updates counterfactuals and weighting vectors to refine explanations.
Illustrative embodiments highlight the use of invariant maximum values for features and distance algorithms to minimize modifications. The system supports multiple configurations, allowing for alterations and adaptations suited to different data processing environments. It is designed to work with diverse data types and storage devices, offering flexibility in deployment.
The patent emphasizes that the described embodiments are examples and not limitations, suggesting that many variations are possible within its scope. The system can be integrated with various existing technologies, architectures, and applications, enhancing its applicability across different AI-driven fields. This adaptability ensures broad utility in improving AI model transparency.