The Business Guide to Machine Learning

By Rob Marvin Features PCmag

Machine Learning (ML) algorithms are embedded in the fabric of much of the technology we use every day. ML innovations spanning computer vision, deep learning, natural language processing (NLP), and beyond are part of a larger revolution around practical artificial intelligence (AI). Not autonomous robots or sentient beings but an intelligence layer baked into our apps, software, and cloud services that combines AI algorithms and Big Data under the surface.

Continue Reading Below

The trend is even more pronounced in business. ML is no longer solely used for specialized research projects undertaken by a team of data scientists. Enterprises now make use of ML to gain actionable business intelligence (BI) and predictive analytics from ever-increasing amounts of data. That's why it's more important than ever to be aware not solely of what ML is but also the most effective strategies in which to use it for tangible value.

Ted Dunning, Ph.D., is the Chief Application Architect at enterprise Hadoop vendor MapR, and co-author of two books on what he refers to as "Practical Machine Learning." The Silicon Valley veteran has worked in the field for decades, watching the AI techniques and the space evolve to the point where advances in cognitive computing and the availability of open-source tools has truly brought ML to the mainstream. Dunning spoke to PCMag to cut through the jargon and explain what ML actually means, and impart some wisdom and best practices on how businesses can make the most of their ML investment.

A Practical Definition

The straight definition of ML is giving systems the ability to act and to iteratively learn and make adjustments, without any explicit programming. Dunning said ML is a branch of statistics but a branch that's very practical. He stressed that, in a real-world business context, you need to be pragmatic and realistic with how you apply it. The core task of ML is to create a business process that's repeatable, reliable, and executable.

"Machine learning isn't about looking backwards at scientific data and trying to decide what conclusions are viable," said Dunning. "It's about looking forward, and asking what we can predict about the future and what will happen in various scenarios. When it comes down to doing business with this data, we're talking about very limited situations where you want replicability."

Continue Reading Below


Image credit: Todd Jaquith at Futurism.com. Click to expand full infographic

Deep Learning vs. Cheap Learning

You can break down that basic idea into a number of different fields within ML, but Dunning pointed to two in particular on either end of the spectrum: deep learning and what he calls "cheap learning." Deep learning is the more complicated concept.

"We wanted machine learning to go deeper. That's the origin of the term," said Dunning. "Over the past 10 or 15 years, techniques have been developed that actually do it. [Machine learning] used to require a lot of engineering work to make relationships in the data visible to algorithms, which, for a long time, weren't as clever as we wanted them to be. You had to hand algorithms this palatable data on a plate, so we used to hand-code all these features that systems now do on their own."

Deep learning is where much of the innovation around neural networks lies. It combines sophisticated techniques such as computer vision and NLP into layers of "deeper" learning that have led to huge strides in areas such as image and text recognition. This is great for complex modeling but can be overkill for simpler, everyday business uses that can rely upon established ML frameworks and techniques with far fewer parameters.

Cheap learning, Dunning explained, means simple, effective, tried-and-tested techniques where businesses don't need to invest expensive resources to reinvent the wheel.

"In computing, we talk a lot about low-hanging fruit. The availability of data and the massive increase in computational capacity means we've lowered the entire tree," he explained. "Simple machine learning isn't just for data scientists anymore."

How Does Cheap Learning Work?

Basic ML algorithms can identify correlations and make recommendations, or make experiences more contextual and personalized. Dunning said there's an opportunity in pretty much every aspect of how we interact with computers for them to use cheap learning to simply make things work better.

One example of cheap learning in practice is in fraud detection. Banks and merchants deal with widespread fraud, but it's often dispersed and concerning low enough values that it doesn't get reported. Dunning explained that, by using a cheap learning algorithm (that is, an existing ML test programmed for this specific task), merchants can more easily identify the common points of compromise that put users at risk, and catch fraud patterns that wouldn't otherwise be visible.

"Suppose that you want to find which merchants seem to be leaking data that leads to fraud. You can use a G^2 test to simply find which merchants are over-represented in the transaction histories of victims of fraud versus consumers without fraud," said Dunning. "This seems too simple to be called machine learning but it finds bad guys in real life. Extensions of this technique can be used to augment somewhat more advanced techniques allowing simpler learning algorithms to succeed where they might fail otherwise."

Cheap learning can be used in all sorts of different ways, so Dunning gave another example of how an online business might use it. In this instance, he explained how an existing ML algorithm can solve a simple comment ranking problem.

"Suppose you have an article with a number of comments on it. What order should they be placed in? How about ordering the comments according to how interesting people think they are? You can count the number of times people read the comment, and how many times they upvote it, but there is a little bit of magic still needed," said Dunning.

"One upvote from one reader probably isn't actually better than eight upvotes out of 10 readers," he explained. "Even worse, if you put early winners on top, the other comments never see the light of day and so you never learn about them. A tiny bit of machine learning called Thompson sampling can solve this in a way that gathers data on new comments and where rankings are uncertain, but generally orders them in a way that gives users the best experience."

Dunning also laid out a set of best practices for how your business can make the most of ML. For a breakdown of how logistics, data, and an arsenal of different algorithms and tools factor into a successful business strategy, check out our 7 Tips for Machine Learning Success.

This article originally appeared on PCMag.com.