debiasA
About
This personal project is inspired by some of the exploratory work I performed regarding implicit bias and discrimination in deep learning models. That can be found here: github.com/bg459/debias.
debiasA seeks to automate the process of reducing bias that is often inherent in a typical deep learning model. The goal is to simplify the steps needed to create fair and objective models. Given as input a neural network pipeline, viz. the model structure, processed train/test data, hyperparameters, optimizer, debiasA will create a adversarial network pipeline that influences training of the given model.
Problem Statement
It is well known that the classical deep learning model is subject to discrimination and implicit biases, e.g. in contexts such as loan defaults, criminal recidivism, etc. Active steps must be taken to curtail this tendency and metrics should be created to measure the extent that this treatment is occurring.
Proposed Solution
The proposed solution is to modify the given neural network by appending an adversarial network which targets the class of interest.
For example, consider a network which uses several features about an individual to predict income. This might show up in a loan default prediction task. If one wishes to diagnose possible bias in the "race" feature in this prediction, the adversarial network will seek to predict race based on the predicted income. If the predicted income is a good predictor of race, then the network sends an adversarial signal back to the network, since income is not supposed to be a good predictor of race, in theory (especially if race is not even a feature to begin with!).
In this way, the loss function of the entire network
because the function wishes to punish accurate prediction of race given income, which would mean a low
Usage
The program requires the following files in its directory. I have included samples of the files in the repository. The names of these files can be anything, so long as their respect names are passed in on command line. Otherwise, this will be the default file locations.
- base_model.h5: This is the network structure of the original network. Future adversarial layers will be built atop this network.
- processed_train.pkl: This is a pickle of the training dataset right before it goes into the original network for training. This way there are no discrepancies with data processing.
- processed_test.pkl: This is a pickle of the test dataset right before used as testing. In theory this should be similar to the train in terms of dimension, data types, etc.
The program, at run, also takes a set of hyperparameters at command line. It is run through a main Python shell, through which one can also find the exhaustive list of options to give the program.
python main.py -help
Here are some of the more important/relevant options to feed.