Moving AI and ML from research into production

In this interview from O’Reilly Foo Camp 2019, Dean Wampler, head of evangelism at Anyscale.io, talks about moving AI and machine learning into real-time production environments.

Highlights from the interview include:

Facilitating the transition from research to production in a robust way introduces a number of complications, Wampler says, including governance, GDPR, and traceability rules. Noting the importance of traceability, he offers an example: “If I deploy a model that’s making credit card authorizations, and I keep rejecting someone’s card, and they come on and say, ‘I’m a member of a minority group, and you keep turning down my charges. Are you prejudiced against me?’ or something like this, I need to know exactly what model was used and how it was trained. There are all kinds of logistical issues that have to be addressed in a real-world production environment.” (01:15)

In some cases, AI and machine learning technologies are being used to improve existing processes, rather than solving new problems. Wampler used car loan approvals as an example: “It used to take a day or so to get an auto loan, and that worked. You could just come back to the dealer the next day and dream about your beautiful car that night but not actually have it. Companies like Capital One have gotten that [loan approval process] down to seconds. You can get on the app and get an approval for a loan immediately. So, it’s not something that had to be done in a real-time context, but it changed the world, changed their business being able to do that. There’s a lot of these sort of pragmatic examples.” (02:22)

Wampler also discussed his personal interest in climate change and how individuals and businesses can use AI and machine learning tools to have a more significant influence than one might think. “What I’ve learned is there are a lot of little ways and big ways that add up when we’re working on stuff like this. One of the promises of tools like artificial intelligence is that it can automate human-level activity in a way that would not be feasible with actual humans doing it. More specifically, organizations like Google are already using sophisticated analytics to reduce the amount of energy they use and more efficiently utilize their machines. Individually, things like that are not going to solve climate change, but they add up. Every ton of carbon that you didn’t burn is one step in a solution toward the problem of climate change. For all of us, it really comes down to a whole spectrum of little things we can do that add up, from personal things like how we use energy, heat our homes, cook our food, and so forth, to thinking carefully about how we do our jobs and how we can be efficient in operationalizing these things, thinking about how we can help our customers achieve that, and then figuring out ways that we can have more direct influences.” (04:20)