Uninformative Feature Baseline for ML Attribution Uncertainty
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing explainable artificial intelligence (XAI) systems lack a baseline for evaluating the trustworthiness of feature attributions in machine learning models, as they do not provide a frame of reference to determine the meaningfulness of feature contributions, leading to uncertainty in model output, especially in autonomous systems like autonomous vehicles.
Innovation Solution
The introduction of uninformative features that do not correlate with the model's prediction task, allowing for comparison with real features to assess attribution reliability and uncertainty in model outputs by training models on combined datasets of real and uninformative features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If feature attributions are generated to explain model output, then model interpretability is improved, but trustworthiness of explanations deteriorates due to lack of baseline for comparison
Solution Approach 1:
The patent introduces uninformative features as an intermediary baseline that mediates between model predictions and explanation evaluation. These artificial features serve as a reference point to calibrate and evaluate the attributions of real features, enabling trustworthiness assessment by comparing whether real features receive higher or lower attribution values than the uninformative baseline.
2Measurement precision
If uninformative features are added to the dataset, then baseline for evaluation is improved, but device complexity increases
Solution Approach 1:
The patent applies local quality by making uninformative features locally distinct from real features through specific design characteristics (e.g., constant values, random noise, or features explicitly constructed to have no predictive power). This allows the baseline to be embedded within the dataset without requiring fundamental changes to the overall data structure or processing pipeline.
3Reliability
If real features are removed from explanation when uninformative features rank higher, then trustworthiness is improved, but loss of information increases
Solution Approach 1:
The patent applies partial action by selectively removing only those real features from explanations that rank lower than uninformative baseline features. Features that rank higher than the baseline are retained in the explanation. This partial filtering approach maintains trustworthiness by excluding potentially spurious attributions while preserving informative real features, thus avoiding excessive information loss.
Data Source
AI summary
In some aspects, a computing system may generate uninformative features that may be added to a dataset of real features to use as a baseline for determining the quality of an explanation of model output. The uninformative features may be features that do not correlate with what a model is tasked with predicting (e.g., the uninformative features may be random values), and the real features may be informative and correlate with what the model is tasked with predicting (e.g., variables of a dataset sample). A machine learning model may be trained on a dataset that includes both the real features and the uninformative features. The computing system may generate feature attributions for model output, which may include feature attributions for the uninformative features and the real features in the dataset.