NOTE: this reports on an AFOG relevant ISchool final project for the Master of Information Management and Systems (MIMS) program. The project was developed by a student team composed of Samuel Meyer, Shrestha Mohanty, Sung Joo Son, and Monicah Wambugu. Students from the team participated in the AFOG lunch working group meetings.
by Samuel Meyer
Opaque: hard to understand; not clear or lucid; obscure. All algorithms are opaque to those who do not understand how they work, but machine learning can be opaque even to those who built it. As a result, much academic work on explainable machine learning focuses on just trying to explain how the model works to those who built it, let alone opening it up to the general public. Unfortunately, many new machine-learning products sold to governments and large companies are only understandable by their developers. If only the developers understand the systems, how can the public make sure that government systems make fair decisions?
Not all is lost, however. Contrary to the narrative that all machine learning is completely a black box, talking to machine learning practitioners will reveal that some machine learning algorithms are generally considered to be interpretable (such as logistic regression and decision trees). They may not always be as accurate as neural nets or other advanced machine learning, but these algorithms are often used in real-world applications because they are easy to implement and easier for experts to interpret.
As a first step toward making general systems that will let the public look at machine learning, we designed and built a web application that allows arbitrary csv-based datasets be fit with logistic regression and viewed by non-experts. It allows machine-learning developers to share what a model is doing with outsiders. If the non-experts want to add comments about factors or weights in the model, they can.
Also, the web app includes an implementation of equal opportunity, a mathematical definition of fairness created by Moritz Hardt, a member of AFOG. This allows users to see what effect the fairness requirement would have on their dataset.
Definition of opaque from dictionary.com.