What we did
Immadata is a robust, efficient, and innovative concept that we propose specifically to combat facial recognition bias. Immadata does this by generating fake hyper-realistic images of faces. This then feeds into our facial recognition program which is trained using those same photos. This aims to solve the problem of lack of datasets with people of color, females, and other misrepresented groups. This project does not only stop at this use case. It can be used to generate any kind of photo, with the right training, and potentially other types of data can be looked at as well.
We have never tackled machine learning not facial recognition before. Our biggest challenge was manipulating StyleGAN to produce only the photos we wanted (mostly minorities). This was extremely hard since although StyleGAN is easier to manipulate than other GANs, it requires immense knowledge in computer science topics that we had no clue about. Nonetheless we learned a lot. Especially about GANs. We needed up using a technique called style mixing to get the result we want.