This is the third of a multi-part series explaining the fundamentals of deep learning by long-time tech journalist Michael Copeland.
You may not be a rabid fan of Formula 1 racing (it’s the one with the cars that happens mostly outside the U.S.), but it’s hard not to appreciate the machines. Sleek, beautiful, wicked fast, and filled with exotic technology that someday might show up in your relatively pokey commuter car.
If you think about the concept of machine learning today, it’s a little like that Formula 1 car—powerful, exotic, fast, and very impressive in stacking up wins against opponents in Jeopardy! and the maniacally complex game of Go.
That makes for great spectacle. But machine learning adds an extra element—a chance to get behind the proverbial wheel ourselves and mash the pedal. We can apply machine learning techniques, not to television contests and board games, but to our own complex problems and businesses.
In other words, it can make you money.
Even more important, compared to F1, it’s not just the bits and pieces we can get our hands on, like the paddle shifter or variable valve timing. It’s the entire car, driver included.
For businesses, this begs the question: where do you want to go? Or more precisely, what problems do you need solved where machine learning can help? How can a technology that helps researchers win games, help you win new customers.
Deep Learning and Game Theory
Think about what AlphaGo did in its series of Go matches. It had an opponent, Grandmaster Lee Sedol, and an objective – to win the match given the rules and confines of the game. AlphaGo trained itself to learn the game, and to find a path that led to winning more times than not (Alpha go won four of five matches).
In that context, Lee can be considered an agent, while Go provides the constraints and context of how to interact with—and ultimately beat—that agent. Now put those same elements into the realm of self-driving cars.
Driving is a more sophisticated example in that it’s a multi-agent environment. The other “agents” are the other drivers. They might be human, or they could be other AI drivers. The board is the road with all its variability and changing conditions.
To learn more about where AI is going next, see NVIDIA CEO Jen-Hsun Huang’s report on The Intelligent Industrial Revolution.
“Winning” in this scenario might be getting from home to work safely, and without running out of gas. Or it might be getting to work as quickly (and safely) as possible. Different drivers will pick different constraints and have different objectives. The objectives will be achieved (or not) while interacting with other agents, and within the rules set to operate.
Swap out driverless cars for drones, and you can see how Amazon is thinking about applying machine learning to the problem of package delivery by autonomous flying machines. Winning for Amazon is the fast, safe airborne drop-off of a pair of shoes to the correct address.
If you believe in the theory of gamification – that pretty much anything that you do in human behavior, or even machine behavior, can be represented as a game with rules and objectives – then it’s not hard to make the leap and see almost infinite applications for machine learning.
Deep Learning to Optimize Manufacturing
Take a manufacturing line. If winning is to crank out the greatest number of finished products that pass the quality control test, a machine learning approach can be used to train a system — a deep neural net (DNN) in technical parlance – to finely tune all the parameters in the manufacturing line to do just that. You could even extend the parameters back to include your entire supply chain and outside suppliers.
With the right data, a DNN can continuously turn the dials of a business to optimize for the objective. It might be throughput during peak season, or efficiency during slower periods. The price of some component goes through the roof, and the calculus changes yet again, and the dials get tuned automatically.
Deep Learning to Improve Industrial Design
Machine learning can also help with designing the next version of a product. General Electric uses data flowing from sensors and machine learning techniques to optimize the design and performance of everything from oil wells to jet engines.
But they don’t stop there. The company also uses it to engineer improved versions for the future. It’s all about turning those dials and seeing impact relative to what winning is.
Better Beer, Bigger Profits, Smarter Predictions
Analytical Flavor Systems, a software startup based in Pennsylvania, uses machine learning to help its clients brew more quaffable beer based on the data it extracts from helpful beer tasters. The company’s version of a DNN, called Gastrograph, can identify dozens of beer styles, and drill down on the characteristics that beer drinkers are liable to love, or send them off to a competing brand of suds.
Both the NBA’s Golden State Warriors and the Cleveland Cavaliers are using machine learning techniques embedded in a service from L.A. -based Second Spectrum. The service uses computer vision and statistics to predict how individual players will perform in different game scenarios (it would have been nice if they could have spotted Curry’s three-point shooting slump in the finals). While it didn’t take a sophisticated machine learning approach to suggest it, no doubt the Dubs saw a gap in the data that Kevin Durant could fill.
Financial technology startup Affirm, based in San Francisco, uses machine learning techniques at every phase in product development — from prototype through to production. The resulting credit and fraud models help the company decide who to lend to when they click on an offer to finance that Casper mattress. In Affirm’s case, winning is offering the right priced loan to the right person who, based on its models, will pay it back.
How Can Deep Learning Help My Business?
It’s easy to understand that for companies like Google, Facebook, Amazon, Netflix and others, machine learning is a powerful tool. They’re in the data business. But as you can see from the above examples – beer, personal loans, oil wells, basketball – most businesses these days are, too, or will be soon.
The buzz phrase “big data” has been invoked repeatedly for several years. Most companies have jumped on that bandwagon. CEOs and CTOs recognized there was value in collecting all this data around their business processes.
That was phase one. Phase two was about finding more sophisticated ways to query that data – the basic business analytics trend. Phase three, which we’re entering now, brings to bear machine learning techniques on the data.
Machine learning will help businesses develop models that are less backwards looking and more predictive in terms of outcome and more instructive when it comes to what to do or build or offer in your business.
The difference between the drumbeat of the early big data days and now is that a lot of this stuff is really starting to work in surprising and valuable ways. It’s not just beating skilled humans at games, but the more mundane victories of an Apple Siri that works really well, or image recognition that can spot a potential cancer on an MRI far faster and with greater accuracy than a human. It’s improvements in the accuracy or the capability of your favorite SaaS platform running HR or data center management. Machine learning techniques are enabling those, too.
Another surprise is that companies like Google, Facebook, Amazon and yes, NVIDIA, have been relatively open about making available a lot of their software tools to get started. They’ll even offer engineering help (for a price) to get you started.
This isn’t to say these are simply humanitarian gestures, there is clearly a business to build in machine learning from a services and hardware side. But you don’t have to bring a team of data scientists to the party; you only have to bring a problem to solve – something every company has.
As the CEO or CTO of a company, you already know what will make a difference in your business. Rather than get distracted by some of the shinier applications of machine learning in business, start with a simple business problem that can be automated perhaps, or simply improved.
Machine learning is that wickedly fast car, but first you have to get into it and drive.
To learn more about the evolution of AI into deep learning, listen to our in-depth interview with NVIDIA’s own Will Ramey.