Using DB-VAE's in Facial Recognition

Using a DB-VAE, we was able to reduce any bias a classifier would have. This classifier will be able to tell you whether or not you are human using the distribution of the latent space.

Comment

Inspiration

One of the first facial recognition projects I did had a strong bias towards my mom for some reason, so we decided to see if we could fix that because it made both of us feel bad.

What it does

Our program utilizes DB-VAEs in order to classify images without bias in an unsupervised way. This means that latent bias will also be accounted for in addition to decreasing the cost of labor extensively

How we built it

We built it on a research paper about mitigating bias in variational autoencoders.

Challenges we ran into

Tensorflow was not very cooperative and it was very hard to set the project up on the gpu, but after that the only challenges we had was with little issues like the dimensions of an array being wrong.

Accomplishments that we're proud of

Understanding a very complicated research paper and the math behind it. In addition, we were very time crunched, so I think we did a good job considering how much time we spent.

What we learned

About VAEs and a lot more about the reduction of bias in general. It also brought to our attention to how serious the issue of bias in machine learning is.

What's next for Using DB-VAE's in Facial Recognition

a more complex UI system that is able to collect AND show the output in one window. A website would also be beneficial to learn more about machine learning on cloud based systems.

Built With

Share this project:

Updates