Machine Learning for High-Risk Applications

Chapter 12. How to Succeed in High-Risk Machine Learning

While artificial intelligence and machine learning have been researched for decades, and used in some spaces for almost as long, we are in the early stages of the adoption of ML in the broader economy. ML is an often immature and sometimes high-risk technology. ML is exciting and holds great promise, but it’s not magic, and people who practice ML don’t have magical superpowers. We and our ML technologies can fail. If we want to succeed, we need to proactively address our systems’ risks.

This entire book has put forward technical risk mitigants and some governance approaches. This final chapter aims to leave you with some commonsense advice that should empower you to take on more difficult problems in ML. However, our recommendations are probably not going to be easy. Solving hard problems almost always requires hard work. Solving hard problems with ML is no different. How do we succeed in high-risk technology endeavors? Usually not by moving fast and breaking things. While moving fast and breaking things might work well enough for buggy social apps and simple games, it’s not how we got to the moon, fly around the world safely on jets, power our economy, or fabricate microchips. High-risk ML, like each of these other disciplines, requires serious commitments to safety and quality.

If we are in the early days of ML adoption, we are at the very dawn of ML risk management. Only in 2022 did the National Institute for Standards ...