Abstract: 

Building machine learning models that are accurate and bring value to your organization is not an easy task. You need to define the problem, find and clean the data, engineer features, select and tune machine learning models, evaluate it and put it in production. It’s is quite a lot already, but today we are starting to care more and more about explainability and fairness of machine learning models and the impact it has on society at large. This adds yet another level of complexity. How should you approach it and what are the tools that can help you do that?

In this workshop, we will focus on fairness and removing bias from your models. More specifically we will look at how to apply fairness principles in practice. We will talk about model bias, typical examples of it, and the benefits of unbiased models (or more realistically less biased models). We will consider the risks and costs (in reduced accuracy) that you need to incur when implementing fair models in real-world scenarios.

More specifically, we will go over the fairness metrics and figure out which metrics could be important to your use case. We will take a look at the landscape of tools available today that can help you explore model fairness. In particular, we will focus on aif360, a scikit-learn compatible toolkit developed at IBM that can help reduce unwanted bias and build better, more human-focused machine learning models.

Hopefully, after this talk, you will be able to approach your next machine learning project fully equipped to build fair, human-centered machine learning models for the benefit of us all.

Bio: 

Senior Data Scientist at neptune.ml, graduated Physics at University of Silesia in Katowice and Finance at University of Economics in Wroclaw. Worked on various data science projects involving facial recognition, optical character recognition, cancer detection and classification, satellite image segmentation, text mining labor market data and many more. He was a member of the teams that won MICCAI Munich 2015 „Combined Imaging and Digital Pathology Classification Challenge”, won MICCAI Athens 2016 „Pet segmentation challenge using a data management and processing infrastructure” and won crowdAI "Mapping Challenge" competition in 2018.