AI Explanation Using Prototypes and Criticisms for Stable Rationales
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models lack explainability, making it difficult to understand the rationale behind their outputs, which affects fairness, reliability, and accuracy, and current methods for explaining AI thinking are unstable, limited to specific model types, or difficult to apply to high-dimensional data.
Innovation Solution
A method involving configuring prototypes and criticisms to represent dataset distributions, calculating feature-based rationales by considering both local and global data instances, and generating output screens with rationale content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing AI thinking identification methods (perturbation-based, backward propagation-based, example-based) are used, then some explanation capability is provided, but the methods suffer from high instability, limitations to specific model types, and difficulty in deriving model interpretations
Solution Approach 1:
The method segments the dataset into two distinct sets: prototypes representing typical instances and criticisms representing atypical instances. This segmentation allows the explanation system to separately analyze what the model learns from typical patterns versus what it learns from exceptions, providing more stable and interpretable explanations across different model types.
Solution Approach 2:
The patent introduces prototypes and criticisms as intermediary concepts between the raw data and the model predictions. These intermediaries serve as representative samples that mediate the explanation process, making the model's decision-making process more transparent and stable without being constrained to specific model architectures.
2Measurement precision
If perturbation-based methods (e.g., LIME, Influence Functions) are used to explain AI behavior, then local explanation capability is provided, but the methods exhibit high instability due to sampling variations and cannot effectively handle high-dimensional data
Solution Approach 1:
Instead of using random perturbation sampling that leads to instability, the method creates copies of representative data instances (prototypes and criticisms) that capture the essential characteristics of the data distribution. These copies serve as stable reference points for explanation, eliminating the randomness and instability inherent in perturbation-based approaches while maintaining local explanation accuracy.
3Loss of information
If backward propagation-based methods (e.g., LRP, SmoothGrad, DeepLIFT, Integrated Gradients) are used, then feature importance can be calculated through error back-propagation, but these methods are limited to deep learning models and require pre-defined baselines
Solution Approach 1:
The patent develops a universal explanation framework based on prototypes and criticisms that can be applied to any machine learning model type, not just deep learning models. This approach eliminates the need for model-specific algorithms and pre-defined baselines, making the explanation method universally applicable across different model architectures while still effectively capturing feature importance information.
4Productivity
If existing methods calculate variable influence in models like min(x1, x2), then the calculation is completed, but variables with large values are incorrectly calculated as having zero influence, leading to misleading explanations
Solution Approach 1:
The method applies local quality analysis by examining how model predictions change when specific features are modified in the context of prototypes and criticisms. Instead of relying on global sensitivity measures that fail for non-linear models like min(x1, x2), the approach locally evaluates feature influence by comparing predictions on prototype-criticism pairs, accurately capturing the true influence of each feature even when values are large.
Data Source
AI summary
A method performed on a computing device for providing explainable artificial intelligence (XAI) according to an embodiment of the present disclosure includes configuring a first set of prototypes representing each data instance of an entire dataset input to a machine-learned model, configuring a second set of criticisms that are samples of data instances not expressed by the prototypes among each data instance of the entire dataset, and calculating a first feature based rationale of output data for a first data instance of the model by considering all prototypes included in the first set and all criticisms included in the second set.