Responsible Automation: Towards Interpretable & Fair AutoML

Automatic Machine Learning (AutoML) is a subfield of machine learning which aims to automate the training & tuning of machine learning models. One of the main goals of an AutoML tool is to train the “best” model possible in the least amount of computation time, with zero/minimal configuration by the user. AutoML tools reduce the expertise required for practitioners to train powerful machine learning models, which has expanded and accelerated the application of machine learning to problems in both academic research and industry. AutoML greatly speeds up the workflow and efficiency of even the most experienced data scientist. As automation and use of machine learning increases, in particular with the proliferation of open source AutoML tools, there’s an increased risk in misuse of, or harm by, machine learning models used in real world applications. In order to reduce the risk of harmful models being deployed, machine learning tools, and especially AutoML tools, can offer easy-to-use or automated interpretability and algorithmic fairness methods that can be used to evaluate and probe machine learning models. Interpretability and fairness methods should always be applied to machine learning models before they are deployed into production where they can make or influence important decisions affecting people’s lives. In this session, you will learn about: - Automated Machine Learning and open source H2O AutoML - Interpretability methods for H2O models - Algorithmic fairness (disparate impact) for H2O models - Demo using U.S. Home Mortgage Disclosure Act (HMDA) data