AI put in charge of creating fairness training of AI

Researchers say they have developed a way to create data “pipelines” to supply training models with high-quality software-fabricated faces. Their results, however, show that, at least in this case, artificial intelligence does not perform better in training AI than debiasing models.

The faces can be used to automatically generate and supplement training datasets, according to a recent paper. In either case, the generated faces can address attributes underrepresented in a database without a human designing the faces.

The work is designed to make AI more accurate at identifying non-white, middle-aged males. It was never the case that people of color and females routinely were represented in the proportion they occupy in reality, making AI training incomplete.

And, as biometric privacy rights become more of a hurdle, it has only gotten more difficult to find large, representative datasets.

The paper, written by two University of Southern California scientists, has not yet been peer-reviewed. Their research was partly funded by the Intelligence Advanced Research Projects Activity unit within the Office of the Director of National Intelligence.

The researchers say their technique allows face-generating algorithms to create larger or smaller attribute classes, making more “chubby” (their description) females with baggy eyes, wavy blond bangs and heavy makeup, for example.

More work is needed, however.

According to the report, “although the network trained with the proposed synthetic data does not outperform the debiasing models in recognition performance, it is impressive that synthetic images can achieve consistent performance with real data and further yields better fairness.”

Article Topics

AI  |  biometric bias  |  biometrics  |  face biometrics  |  research and development  |  synthetic data