Machine Learning Bias Testing Using Controlled Visual Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems used in video surveillance and image enhancement exhibit bias in recognizing certain attributes, such as skin-tone, age, and object colors, which can lead to inaccurate results and unfair treatment.
Innovation Solution
A system and method for testing bias in machine learning systems by generating synthetic test datasets using generative AI models or game engines to vary specific visual attributes, allowing for the assessment of bias and providing performance scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning systems are trained on real-world image data, then they achieve practical recognition accuracy, but they exhibit bias toward certain attributes (skin-tone, age, object colors)
Solution Approach 1:
The patent creates synthetic copies of real-world images through generative AI models and game engines. These synthetic images replicate the complexity and diversity of real-world scenarios while allowing precise control over visual attributes. By copying and augmenting real image characteristics in a controlled synthetic environment, the system can generate balanced training data that maintains practical recognition accuracy while eliminating real-world biases.
Solution Approach 2:
The patent systematically varies visual parameters (skin-tone, age, object colors, lighting conditions) in synthetic image generation. By changing these parameters across the full range of possible values and combinations, the training dataset achieves comprehensive coverage of attribute variations. This parameter exploration ensures the machine learning system learns to recognize objects and activities accurately across all attribute categories without developing bias toward any specific range.
2Adaptability or versatility
If diverse training data is collected from real-world sources, then coverage of attribute variations improves, but control over specific attribute distributions is lost
Solution Approach 1:
The patent introduces synthetic image generation systems (generative AI models and game engines) as an intermediary between real-world data collection and machine learning training. This intermediary layer allows the system to inherit the diversity and realism of actual images while gaining precise programmable control over attribute distributions. The synthetic generation process acts as a mediator that translates real-world complexity into controllable parameter spaces.
Solution Approach 2:
The patent performs preliminary attribute balancing during the synthetic image generation phase, before the machine learning training begins. By pre-generating datasets with controlled attribute distributions that ensure equal representation across all visual attributes, the system eliminates the need for post-hoc bias correction. This preliminary action of creating balanced training data upfront simplifies the overall training process and ensures fairness from the start.
Data Source
AI summary
A system for testing bias of a machine learning system for enhancement or identification of objects and/or activities in image data includes a test image generator that receives an input from a source and processes the input to generate a plurality of test images in which a visible attribute of a subject is different in each of the test images. The system also includes a testing module that inputs each of the test images to the machine learning system and outputs a result for each test image, and a bias analysis module that compares the result with an expected result for each test image and generates performance scores indicating the performance of the machine learning system in different categories of the visible attribute.