Responsible Machine Learning

Preface

Machine learning (ML) sits at the cross section of business applications, statistics, and computer science. It’s seen several waves of hype and disappointment in its roughly 60-year history. It’s a big, technical subject, and it’s also emerging as a powerful commercial technology. Yet, like other powerful technologies, it presents both opportunities and challenges. It can be a tool or a weapon. It can generate revenue, and, in some instances, be transformational for organizations. But it can also fail, be attacked, and cause significant incidents. From where we sit, most of the ML community appears too focused on ML’s hype and upside. When we focus only on the upside of technology, we turn our backs on obvious realities that must be addressed to enable wider adoption. Perhaps this is why we’ve found it a little strange to write a report called Responsible Machine Learning. After all, we don’t often hear about “responsible” aviation or “responsible” nuclear power. Responsibility and risk mitigation are usually baked into our notions of these technologies, and this apparently hasn’t happened yet for ML. This could be because we know what happens when commercial jetliners or nuclear reactors fail or are attacked, but as a community, we’re not yet so sure about the consequences of lousy ML.

If this is the case, it betrays a lack of effort by the practitioner community. While the consequences of ML failures and attacks are just beginning to filter into the news, if you know ...