Image Sensor Bias Mitigation Using XAI Contribution Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies fail to address the challenge of bias in datasets, which are not addressed by existing technologies have not effectively mitigated bias in AI training data, leading to unfair outcomes.
Innovation Solution
Utilizing explainable AI (XAI) to predict sensitive attributes from non-sensitive attributes, calculate contribution degrees, and modify correct answer labels to mitigate bias, and retrain models based on influence functions to reduce model bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a complete dataset ensuring fairness with respect to race and sex is collected, then fairness of AI determination is improved, but it is practically difficult to collect such data because random collection causes biased data distribution
Solution Approach 1:
The patent applies preliminary action by calculating influence functions before model training to identify biased data in advance. The system computes the influence function for each training data point regarding sensitive attributes (race, sex) before the model is trained, then uses this information to selectively remove or adjust biased data points, thereby ensuring fairness before the learning process begins rather than attempting to collect perfectly fair data from scratch
Solution Approach 2:
The patent replaces the mechanical approach of manually curating fair datasets with an automated computational approach using influence function calculations. Instead of relying on manual data selection processes, the system uses mathematical calculations to automatically identify and remove biased data points based on their influence on predictions regarding sensitive attributes, substituting computational analysis for manual data curation
2Reliability
If bias mitigation processing is added to the AI training process, then fairness of AI outcomes is improved, but the complexity of the processing system increases
Solution Approach 1:
The patent extracts the bias mitigation function as a separate, independent processing step from the main model training process. The influence function calculation and data selection logic are implemented as distinct modules that operate before training begins, allowing the core training algorithm to remain simple while adding fairness guarantees through a separate preprocessing stage that identifies and removes biased data points
Solution Approach 2:
The patent performs bias mitigation actions before model training begins by calculating influence functions and identifying biased data points in advance. This preliminary analysis allows the system to prepare a fair training dataset without adding complexity to the actual training process, as the bias removal logic is executed once before learning starts rather than during each training iteration
Data Source
Figure 1
Figure 2
Figure 3
AI summary
To visualize and remove bias. An information processing device includes: an acquisition unit that acquires information regarding bias caused by a sensitive attribute included in data or a model by using explainable artificial intelligence (XAI); and a processing unit that executes processing of mitigating the bias of the data or the model on a basis of the acquired information. The acquisition unit includes a prediction model that predicts the sensitive attribute from a data item of a non-sensitive attribute in input data including a plurality of data items, and a contribution degree calculation unit that calculates the degree of contribution of each data item of the non-sensitive attribute on the basis of a determination basis of the prediction model obtained by using XAI.