GE: Success of AI systems tied to engineering, not algorithms

Building AI systems is less an algorithmic challenge and more a software engineering challenge, according to GE's Jeff Erhardt, vice president of intelligent systems. Like other types of enterprise software, AI systems have to be secure, robust and scalable, especially in the industrial space where inaccurate recommendations could have catastrophic consequences.

For that reason, AI and machine learning aren't thought of as traditional data science endeavors at GE, as Erhardt explains here.

Editor's note: The following was edited for clarity and brevity.

GE is known for building physical objects like wind turbines and jet engines. Your job is to make those physical objects smart. How hard is it to integrate AI and machine learning into pre-existing systems?

Jeff Erhardt: There's a lot of hype, a lot of buzz around AI and machine learning. And much like big data from five years ago, people are going to start entering the 'trough of disillusionment,' to steal Gartner's term, because getting this stuff right is incredibly hard. Even though places like Amazon or some of the tools vendors are trying to make the underlying algorithms easier to use, that's actually not the hard part.

Jeff Erhardt

AI systems or machine learning systems are really hard for three reasons. First, they're hard to structure. And by structure I mean, how do you understand what you're trying to optimize for, how do you understand bias in the data, how do you understand that the world may be changing. That is not trivial to get right. Second, the systems are hard to implement. There's the old cliché that in data science, data prep and data cleaning is 80% of the work. That's still true, especially in this space. Third, machine learning systems are hard to maintain.

All of those things in context mean a lot of effort is required to get it right. And it's the reason why we -- and companies in general -- should be selective about where and how we implement this type of capability. And it's the reason why, at GE, we're doing this for use cases that have large-scale applications delivering outcomes as a service to our customers such that we can do all of the work that I just described one time and then sell the results to many customers instead of having to go and repeat that every time one by one by one.

Why are AI systems hard to maintain?

Erhardt: When we built old-school software, we'd build a software application and put it on floppy discs, shrink-wrap it and send it to somebody. That was hard. But once you developed, finished and tested the software, it was locked down and it never changed. Then we started developing data-driven software where not only did we have to design a good application, but we also had to account for the fact that incoming data being consumed by that software was potentially changing. That was more complicated because you've got more things to account for, more things that are changing, and so on.

But with machine learning or with an intelligent system, not only do you need a good software application, not only is the incoming data potentially changing, but now the software itself is changing. And keeping track of that -- monitoring and maintaining it over time -- is really hard, especially in the industrial context.

If you're a consumer-facing brand and you show a customer a recommendation for a product that he or she doesn't want to buy, no big deal. But in our space, the cost of false positives or false negatives -- for example, missing a potential defect in an oil and gas pipeline inspection -- has potentially catastrophic consequences. It means we need to be more certain that the recommendations or the decisions within these systems are accurate not just [now], but ongoing into the future.

How does this emphasis on software development influence the way GE thinks about building AI systems?

Erhardt: You have to take great care in designing these systems upfront to be reliable, robust, maintainable, secure and stable. That's why one of our core theses and the way we're building our team and the capability within GE is this: Machine learning is not a traditional data scientist endeavor; it is not for the realm of the citizen data scientist, and, in fact, it is much more of a software engineering or a systems engineering challenge than it is an algorithmic challenge.

We treat this like we're developing true enterprise software where we need to build in quality, testing, security, reliability from ground zero of when these systems are designed and built. And then we're putting in monitoring, controls, etcetera. All of those things need to come together, and, yes, the AI piece is important, but the algorithmic AI is really a small component to the overall success of that business.

Of the things you just listed -- security, reliability, testing and so on -- what area is the most challenging?

Erhardt: Building in fault tolerance and resiliency. Some of the things that I mentioned are knowable and solvable, but many data scientists and data science organizations don't think about them.

Few people think about the concept of doing [continuous integration and continuous delivery] in the context of data science. Google, Facebook, modern companies do, but 99% of the data science organizations I've worked with don't understand what it means to do testing and continuous release.

So, that stuff is there, it's mechanical, but the hard part -- and it's why AI is still a black art and why there are still so few people who can do this -- is how do you think about the problem structure in the context of the application so that not only is the math correct and stable over time, but considers the end-user experience and the way these predictions are consumed if things drift, change, if we explore and exploit and make a prediction that might actually be wrong.