What does it take to build a machine learning capacity? Less than you think
Machine learning is the new game changer in business technology. In a world where digital information volumes are doubling every two years on average, machine learning allows organizations to extract highly valuable information from enormous data stores at heretofore unimaginable speeds.
Building and deploying machine learning solutions can be expensive, requiring investment in servers and storage, expanded networks, and data scientists.
Alternatively, companies can invest in none of the above and turn to one of the many new machine learning as-a-service solutions. Getting started with machine learning in this way basically requires what virtually every organization is awash in today: data.
Machine learning as a service
The “machine” in question here is a computer. Historically, computers spat out results in accordance with the programming instructions fed into them. With machine learning, the computer essentially figures things out on its own, using complex algorithms to identify patterns and trends in data sets. Then the machine produces reports, predicts human behavior, and performs other useful tasks.
These machines can have fun, too. Earlier this year, a machine learning program called AlphaGo from Google DeepMind beat one of the world’s best players of Go, an ancient and highly complex Chinese board game. AlphaGo was not programmed to play Go, but rather learned to play on its own.
Many organizations are experiencing "unheard of improvements" by applying machine learning to various aspects of their business, according to Allan Alter, a senior research fellow at Accenture: “Practical applications of machine learning are very real, spread across just about every vertical.”
Today, scores of vendors sell machine learning solutions that require nothing more than your data. There is no immediate need for additional infrastructure or high-end skill sets, such as predictive analytics experts.
"All that matters to us is the quality of the data we get to put into our machine learning models, and what improves our results over time is even more data from which our models learn," says David Glueck, vice president of data science at Sailthru. The New York-based firm uses machine language algorithms that mine customer data to help companies of all sizes create more effective email and promotional campaigns.
Sailthru offers its solutions as a service built around user-friendly APIs that allow customers to directly download reports and business insights gained by Sailthru's machine learning programs. Glueck notes that managers from outside the IT function frequently initiate engagements with Sailthru. This can create pushback from IT, which often likes to build its own solutions.
“If IT wants to invest and build something similar, it can,” Glueck says, noting however that the latter approach requires IT to allocate resources and incur the opportunity cost of not doing other mission-critical work.
Why buy when you can rent?
One Sailthru customer is Rent the Runway, an online service that offers designer dresses and accessories for rent. As a young company attempting to disrupt a venerable industry, Rent the Runway’s customer acquisition team engaged Sailthru to spread the word about its service. Other Rent the Runway teams worked with Sailthru to optimize email campaigns.
Armed with the company's customer data, Sailthru's algorithms helped identify the customers most likely to rent a dress in the next 30 days. Sailthru drove new customer acquisitions by using Facebook’s lookalike modeling service to identify prospects that fit the customer profiles that matter to a specific business.
Rent the Runway reaped the benefits, including a 40 percent reduction in mobile subscriber acquisition costs and a 28 percent reduction in desktop subscriber acquisition costs. “We help set a very low bar for entry into machine learning,” Sailthru’s Glueck says. “It’s data science as a service. All you need is your own data.”
Healthcare's front line
Machine learning as a service isn't just for startups. With 21 hospitals, four teaching hospitals, and 29,000 employees, Edinburgh-based NHS Lothian is the second largest healthcare provider in the U.K. Its neonatal intensive care unit (NICU) captures enormous volumes of vital patient data that must be shared among doctors, nurses, dieticians, parents, and others. Speed is of the essence: In the NICU, the difference between life and death is often measured in minutes. Yet historically it took two hours or more for NHS Lothian to generate a single report manually.
NHS Lothian needed a machine language application that could scan patient data rapidly and generate automated reports. The hospital chain engaged Arria NLG a London-based company that delivers machine language applications as a service. The results have been striking. Report generation times have fallen from two hours to nearly real time, giving doctors and nurses the information they need to deliver the right care to their small patients.
Arria’s effectiveness comes down to how much data it receives, and how good that data is, says Ehud Reiter, chief scientist at Arria. Thousands of data sets are needed to “train” the system, as machine learning is an iterative process that improves over time with more data. But you don’t necessarily need more infrastructure or analytics experts to leverage it.
Picturing a smart solution
Style Me Pretty is also reaping the benefits of machine learning as a service to help take it to the next business level. With 25 million page views per month, this destination site for wedding inspirations is leveraging a machine learning visual recognition solution from Clarifai—a machine learning firm specializing in visual recognition—to make the leap from wedding blog to full-blown wedding platform featuring vendors and other nuptial resources.
Style Me Pretty is a highly visual site that relies on the 100,000-plus images it gets every week from hundreds of weddings and wedding vendors. Using manual image tagging, Style Me Pretty was able to categorize and publish photos from only 50 of the 600 weddings it receives weekly. With Clarifai’s visual recognition tool, all 100,000 images are now tagged and uploaded, significantly increasing the images available to site visitors.
Meanwhile, the Clarifai solution curates the photos in finer, more discriminate ways so that visitors are served up more refined image content ("hot pink, high bodice, white trimmed lace, knee-length dresses," for example).
Go ahead and build it
Some organizations will choose to build all or some of their machine learning resources in-house. The first challenge they’ll run up against is today's mad scramble to hire analytics experts. The fact is that machine learning vendors typically can and will outspend user organizations for this talent.
“There is a huge demand for machine learning experts,” says Matt Zeiler, who leveraged a PhD in machine learning to found Clarifai. “You don’t need that expertise today to get the benefits of machine learning.”
On the infrastructure side, it takes enormous computing power to crunch today's stupendous volumes of machine learning data. A lot of that power is consumed in the solution testing phases. That is a perfect application of cloud technology, notes Accenture’s Alter. And cloud vendors by and large offer their own machine learning tool sets. Amazon Web Services (AWS) recently announced its Deep Scalable Sparse Tensor Network Engine, an open source version of its deep learning library. Experts generally agree building out full support for machine learning is not for the faint of heart.
Whether you choose to buy, build, or rent ultimately determines whether you have what it takes to build a machine learning operation. The point is that the options are many and growing fast. That’s good news, given the potential of machine learning to transform your business.
Machine learning: Lessons for leaders
- Machine learning as-a-service solutions can be an affordable alternative to expensive investments in technology and personnel.
- Getting started with a machine learning vendor requires little more than what virtually every organization is awash in today: data.
- Machine learning can deliver significant operational improvements in just about every division of your business.