Debiasing-Facial-Detection

Mitigating algorithmic bias

https://www.youtube.com/watch?v=59bMh59JQDo&feature=emb_logo

Watch the video to get insights about Human Bias and how it effects our Machine learning model.

Variational autoencoder (VAE)

We can reduce the algorithmic Bias using VAE.It is an autoencoder whose encodings distribution is regularised during the training in order to ensure that its latent space has good properties allowing us to generate some new data.

Debiasing variational autoencoder architecture (DB-VAE)

Comparison between CNN and DB-VAE

As we can see that the our model is less biased in DB-VAE.

Resources

Dataset

positive image http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Negative image http://www.image-net.org/

Refrences

http://introtodeeplearning.com/AAAI_MitigatingAlgorithmicBias.pdf

https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73

https://www.brookings.edu/research/algorithmic-bias-detection-and-mitigation-best-practices-and-policies-to-reduce-consumer-harms/

https://www.tensorflow.org/api_docs/python

© MIT 6.S191: Introduction to Deep Learning http://introtodeeplearning.com