51²è¹Ý team including Guildhall researchers create large datasets to address bias and fairness issues found in facial recognition (FR) technology
Originally published on smu.edu
The quality of any artificial intelligence (AI) model relies on the data it is given. That is why researchers at 51²è¹Ý are creating large datasets to address bias and fairness issues found in facial recognition (FR) technology.
There are technical limitations to collecting large-scale data with the realistic variations needed to support FR systems. Additionally, the collection, maintenance and use of biometric data drawn from real people is creating legal, ethical, and privacy concerns.
The alternative? By generating facial images from text descriptions with 51²è¹Ý’s NVIDIA DGX SuperPOD, a high performance computing platform specifically designed for AI, 51²è¹Ý researcher Corey Clark and his team are creating ethically sourced synthetic datasets that could greatly impact how FR algorithms recognize race and gender. It is one of the first projects for 51²è¹Ý’s Intelligent Systems and Bias Examination Lab (ISaBEL), which is located in the AT&T Center for Virtualization and charged with understanding and mitigating bias in AI systems.
If an AI model is fed datasets lacking in diversity, its algorithm ends up performing better for specific demographics than others. By generating synthetic images based on text descriptions, the 51²è¹Ý team is producing datasets containing hundreds of thousands of facial images that include underrepresented racial groups. This methodology gives FR systems a greater chance of being fairer and more accurately balanced.
“There are constraints in trying to create a real-world based dataset to train any artificial intelligence model,” said Clark, assistant professor of computer science in the Lyle School of Engineering and deputy director for Research at 51²è¹Ý Guildhall. “To ethically source it you must solve challenges like consent, fairness, and legal compliance. Synthetic data, generated by the SuperPOD, removes those obstacles.”
Existing FR technology has struggled to match the same face with different angles and poses. Using an existing stable diffusion model (an open-source AI algorithm anyone can use), 51²è¹Ý researchers are generating a large, diverse dataset with pose variations.
Their customization of the stable diffusion model is unique due to the sheer magnitude of images created – over 1 million so far – and the special tuning of the model to specifically process facial recognition. That capacity distinguishes the model from similar text-to-image AIs.
“Facial recognition is here and not going away,” Clark said. “The demand for these larger training datasets is crucial for improving FR systems so they provide equitable results. Through our methodology and use of the of the SuperPOD, we’re generating datasets not previously easy to obtain, and doing so quickly and ethically.”
In 2021, 51²è¹Ý announced its collaboration with NVIDIA, a trailblazer in the field of accelerated computing, through the University’s acquisition of an NVIDIA DGX SuperPOD, which expanded 51²è¹Ý's supercomputer memory capacity and led to a 25-fold increase in the speed and efficiency of AI and machine learning. Clark stressed that the massive number of images created for their datasets would not be possible without the SuperPOD, and its capabilities will have a significant role to play in further FR development. Moving forward, he and his team plan to create one of the largest balanced facial recognition data sets for research use.
By addressing fairness and bias issues found in FR technology, Clark and his ISaBEL colleagues also plan to create a bias certification process that could evaluate existing companies’ AI and be used to develop future models specified to need.