Defensible AI Expert Networks With Gradient Descent Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI systems lack transparency in their decision-making processes, leading to potential catastrophic failures, discriminatory outcomes, and legal liabilities due to learning invalid or biased correlations from skewed data, which are not easily understood or reviewed by humans.
Innovation Solution
Implementing a defensible artificial intelligence (DAI) system using gradient descent training and expert networks with meaning-assigned fact nodes, allowing for human-readable explanations and pruning low-value associations to ensure valid and optimal decision-making.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional AI systems are used to learn from data, then decision-making capability is improved, but transparency and human understandability deteriorate
Solution Approach 1:
The patent introduces an intermediary layer between the neural network's internal representations and the external world. This layer includes natural language generation components that translate the network's learned concepts into human-readable explanations, and visualization tools that display the decision-making process in an interpretable format. This mediator preserves the high decision-making capability of traditional AI while restoring transparency through multiple channels.
2Loss of time
If AI systems learn from limited or skewed training data, then model training speed is improved, but decision validity and reliability deteriorate
Solution Approach 1:
The patent implements preliminary actions in the form of data quality assessment and augmentation procedures before the main training process. The system performs initial analysis of training data to identify skewness and limitations, then applies appropriate transformations, synthetic data generation, or selective sampling to improve data representativeness. This preliminary preparation maintains training efficiency while significantly improving the reliability of decisions made by the model.
3Productivity
If AI systems make autonomous decisions without human review, then productivity is improved, but risk of catastrophic failures increases
Solution Approach 1:
The patent implements a tiered review system where not all decisions require full human review. High-confidence, low-risk decisions are approved autonomously for immediate action, while uncertain or high-stakes decisions trigger partial human review. This partial action approach maintains high productivity for routine matters while applying appropriate human oversight where needed, balancing speed with safety.
4Measurement precision
If AI systems learn complex correlations from data, then decision accuracy is improved, but compliance with anti-discrimination laws deteriorates
Solution Approach 1:
The patent applies local quality by treating different regions of the decision space differently. The system identifies protected attributes and sensitive correlation patterns in the data, then applies localized constraints, transformations, or filtering specifically to those regions. This allows the model to maintain high overall accuracy while eliminating discriminatory patterns in specific local areas where they occur, ensuring compliance with anti-discrimination requirements.
Data Source
AI summary
Defensible AI systems and methods may provide technical solutions for technical problems facing typical AI systems. An expert system may be used to address problems facing AI that systems operate without providing visibility into their internal decision-making processes. The expert system may be developed with meaning-assigned fact nodes. A gradient descent style training process may be used to improve the performance of expert system networks. In an example, a gradient descent training process identifies the contributions of rules and fact values to the outcome fact values, then distributes a portion (e.g., velocity value determined portion) of the error to each rule input weighting based on its proportion of overall contribution. These expert systems may use various approaches to training, such as various selected inputs used to calculate the difference value (e.g., error value), various network designs, various error and augmentation levels, and various different training levels.