Deep learning systems have matured considerably, but are these tools ready to bring business value?
For years, companies have been dipping a toe into artificial intelligence, initially with machine learning. But today, as the need for actionable data becomes imperative to maintain a competitive edge, it's important to understand how traditional machine learning has evolved from deep learning (and other advanced approaches), and whether artificial intelligence technologies have value for your business.
Manufacturing companies, for instance, have used traditional machine learning for predictive maintenance and to identify equipment failure, enabling technicians to head off potential problems. But these older systems pale in comparison with what can be done today, using more advanced algorithms that take advantage of state-of-the-art computing power.
While both traditional machine learning and deep learning systems teach computers to learn and accomplish tasks previously performed by human, they differ in complexity. While machine learning can handle only a small number of labeled data streams, a deep learning system can read much more data -- and the data doesn't have to be categorized for the system to understand it. As a result, a deep learning system can make much more rich, contextually based conclusions.
A deep learning algorithm can identify the features of a photo, learn from them and make sense of the photo as a whole. Feed it a picture of you standing near a license plate with a sunny beach in the background, and a deep learning system might deduce that you vacationed on a beach in Florida last summer.
Deep learning systems -- and their application in enterprise artificial intelligence (AI) -- are certainly the subject of a lot of hype these days. But underneath all the noise, experts say, is as set of techniques that will one day transform most businesses. According to Gartner's January 2017 report Innovation Insight for Deep Learning, "Deep learning is here to stay. It is currently the most promising technology in predictive analytics for previously intractable data types for machine learning, such as images, speech and video. It also can deliver higher accuracy than other techniques for problems that involve complex data fusion."
Some organizations have already used deep learning systems to solve pressing problems. For example, NASA developed a deep learning framework for satellite image classification and segmentation, and Insilico Medicine developed a deep learning platform to help find treatments for cancer and other diseases.
But the market is in a very early phase, and only a handful of organizations have a problem that's worth the pioneering costs that come with trying this new and complex technology. The three biggest challenges early adopters have to overcome include the following:
1. You need a lot of data. Deep learning systems don't work without large volumes of data, and the data needs to be appropriate, accurate and free from bias. But huge volumes of quality data can be difficult to come by.
If you're a large company, you might invest in another company just to get information. For example, IBM bought the Weather Company because it wanted to provide Watson, its artificial intelligence technology, with weather data. It also bought several healthcare companies for billions of dollars to obtain hundreds of thousands of patient records and a whole lot of information on things like costs and insurance claims.
But what about the smaller companies that can't go out and buy another company just to get information? "Startups find creative ways to get large data sets," said Aditya Kaul, research director at Tractica, an AI advisory service in Boulder, Colo. "For example, companies trying to develop autonomous car algorithms give away a free app on a mobile phone that sits on your dashboard and records what it sees and passes the data to servers. In return, they give you some freebies, like spotting speed cameras or alerting you if there's an accident."
2. You need to train a deep learning system. Once you get the data, you have to feed it to the system so the system can learn from it. This training is a long process that one rarely gets right the first time.
"Training is a critical aspect of deep learning implementations, which is often a compute-intensive process that can take many days of computer time," according to Gartner's August 2017 report Questions to Ask Vendors That Say They Have 'Artificial Intelligence'. "And once completed, only small modifications are possible, and the entire training must typically be rerun to accommodate new data. Furthermore, the training sets are large and extremely difficult to curate. Output errors driven by errors in the input data are extremely hard to correct, because correlations are often invisible to humans."
3. You need tools to implement a deep learning system. "Enterprise-grade deep learning tools aren't plentiful and robust enough at this time," Kaul said. "So many enterprises need to develop their own tools using open source tools as a foundation layer. Very few organizations have the expertise to do so. For vendors, this need for more robust deep learning platforms that allow you to create production-scale AI solutions for the enterprise should be seen as an opportunity."
"As the market develops, two distinct approaches are beginning to emerge," Kaul said. "The first is the platform approach. . . . Most of these platforms, such as Google's TensorFlow, are useful for researchers, but not as useful in the enterprise. For the time being, very few available platforms work well in the enterprise -- and even those that are out there still need a lot of engineering and tweaking."
"The second approach is point solutions," he continued. "In most cases, point solutions are developed by startups using open source tools, most of the time borrowing well-tested algorithms from open libraries. These boutique vendors usually solve a very specific set of problems for a very specific sector. Their intellectual property is not necessarily the algorithm, but the data they collect and their knowledge of the sector. What they sell is domain knowledge, problem-area expertise and the data sets that fuel the algorithms."
Industry observers agree the market is still young, with only the most sophisticated companies positioned to derive value from deep learning. "Initially, this corporate impact will materialize only for the most advanced organizations with fitting tasks," according Gartner's January 2017 report, Innovation Insight for Deep Learning .
In the meantime, Gartner advises, consider much simpler solutions: "Apply the simplest solution -- for many business problems, 'shallow' machine learning will remain the best approach, avoiding the complexities intrinsic to deep learning implementation."
Deep learning will change the world one day. But for now, too few vendors specialize in the data gathering and cleansing that business strategists need. What’s more, the tools still need to mature before most IT pros will find it practical to use these powerful algorithms. Until the ecosystem of vendors develops further, deep learning will be a stretch goal for many enterprises.
Affiliated professor at Grenoble École de Management, and author of the book Master the Moment: Fifty CEOs Teach You the Secrets of Time Management, Pat Brans writes and teaches about cutting-edge technology and the business surrounding technological innovation. Previously, Brans worked in high tech for 22 years, holding senior positions in three large organizations (Computer Sciences Corp., then-HP, and Sybase).