Zhang implementation

[tmke8]

Oliver:

I feel as though implementing zhang’s mitigating unwanted bias by Adversarial learning algorithm is possible to implement in a way that doesn’t suck. In his original paper he hand writes a den layer network and in aif 360 they do the same. If there’s not a way in tensorflow or pytorch to easily the project the gradients with respect to one loss onto the gradients with respect to another loss perhaps we’ll have to go a bit lower level. It might be a good opportunity to play around with JAX https://colab.research.google.com/github/google/jax/blob/master/notebooks/neural_network_and_data_loading.ipynb