Hybrid use cases to dominate machine learning in 2018, part 1

Post Comment

In 2017, the promises of machine learning and data science flooded conversations across IT departments worldwide, generating a sizable demand for data science skills.

Today, as machine learning use cases continue to grow, so do enterprises expectations and challenges. But as industry-wide enthusiasm around machine learning settles into reality, the need for operationalizing and information architecture becomes important, especially in a multicloud environment.

Big Data and Analytics Hub spoke with IBM Distinguished Engineer John Thomas (@johnjaithomas) about some of the importance of tuning information architecture to make algorithms meet enterprise needs, as well as how machine learning can most effectively be applied in hybrid scenarios in 2018. The following is part one of a two part interview (read part 2).

Big Data and Analytics Hub: What do you love about what you do?

John Thomas: Working with clients on real world applications of machine learning and data science, moving beyond the theoretical domain to applying those concepts in ways that transform their business. That’s what’s most exciting to me: the application of technology to challenging use cases.

BDAH: What are some common machine learning use cases you’re seeing now?

JT: Companies across all industries are looking to apply machine learning and data science to improve both how they conduct their business with customers as well as their internal operations, and that’s spawning a huge spectrum of use cases. There are some very popular use cases where companies are trying to engage with customers on a 1:1 basis: customer churn, marketing offers, upsell/cross-sell initiatives, you name it. There are some uses cases in this space that are beginning to get a lot of attention. Optimizing call center operations is an example.

This is about understanding the intent of a customer’s call, thereby reducing the time the agent spends on the phone as well as significantly improving the customer’s experience. Many times you and I have been on calls where we start with one thing and we end up with something else as the call progresses. Then there are use cases like fraud detection, anomaly detection in manufacturing, predicting delays in a supply chain, applying ML to blockchain scenarios, predicting system outages, [and so on].

The intersection of different types of data is where the value is: when you take various data sets and apply them to solve a particular business problem. You could be dealing with structured or unstructured data, on-premises or off-premises data, streaming or static data. It’s about incremental value versus true transformative value. The latter comes when you can successfully combine different types of data; build highly performing, meaningful models; and operationalize them in ways that transform the business.

AI is a very broad umbrella term. Whether you are looking at classic machine learning or getting into deep learning or cognitive services as a part of that, the most important thing to realize is that there’s no magic box with outcomes and answers. Having a systematic approach to data science and AI is essential,  especially when you have a lot of skunkworks projects that happen in companies. Everybody’s enthusiastic to try things out, but how do you actually operationalize in the context of the enterprise?

BDAH: We all can agree that machine learning is becoming more useful, but do C-level executives have misconceptions about what data science can really do?

JT: There’s an expectation that you can just throw data into a black box, and out come these magical predictions of the future. Maybe someday, but right now we’re quite far from that. A lot of the fundamental concepts of information architecture are required in order to make data science successful. There’s a phrase that keeps getting used, “There’s no AI without IA.” Many of the fundamental information architecture constructs like scalable ingest mechanisms, fit for purpose persistence mechanisms, governance, are needed in enterprise machine learning use cases. The notion that you can take the latest algorithm that came from an enthusiastic researcher, throw your data at it and it will magically solve your business problem; that’s not quite true.

BDAH: Are there some common areas of machine learning application that companies might even be missing out on?

JT: Some of these concepts can be applied to a variety of problems companies have, whether its external-facing or internal optimization problems, but opportunities are missed because companies operate in silos. For instance, line of business may not be closely working with IT guys or data science teams. When you have siloed behavior, you miss out on opportunities. Think about applying machine learning to IT domain problems. Are you able to predict an outage in your core system? Are you able to identify a security breach before it happens? There are interesting challenges, but if the business is siloed in terms of skills on the data science side versus the IT staff, you’ll miss out.

BDAH. 2016 seemed to be the year of cloud. 2017 was all about data science and machine learning. Will 2018 be the year machine learning and AI go mainstream?

JT: I don’t think we need AI to predict whether AI will go mainstream in 2018. Likely that will happen. We’re mostly past the hype cycle. As reality settles in, and as companies apply concepts and techniques, they will find out there are enterprise problems and they need AI platforms that address and surround information architecture.

What will happen in 2018 is there will be a lot of people who are skilled in data science techniques, people who have proficiency in using algorithms and understanding data engineering in the context of machine learning. As skill sets reach a critical mass, companies will settle into a few different platforms to build their AI initiatives on. Support for open standards and collaborative environments will be table stakes. Then you will get into challenges around operationalizing machine learning, challenges around deployment, consumption, version management, immutability of machine learning models, proving to others that the predictions being made were based on models you thought you were using. As AI becomes mainstream, platforms that address these challenges will also become mainstream.

Want to learn more about machine learning and hybrid data management? Join IBM and other industry leaders February 27 at 1:00 ET for Machine Learning Everywhere: Build Your Ladder to AI. Sign up today for a calendar reminder. at