In the past, successful use of machine learning algorithms required bespoke algorithms and huge R&D budgets, but all that is changing. IBM Watson, Microsoft Azure, Amazon and Alibaba all launched turnkey cloud based machine learning SaaS solutions in 2015. At the same time startups like Idibon, MetaMind, Dato and MonkeyLearn have built machine learning products that companies can take advantage of.
Gartner already puts machine learning at the top of its hype curve, and no: machine learning won’t replace all of your employees with computers or suddenly double your revenue. But that doesn’t mean that it can’t give every business a competitive advantage. There are plenty of business processes that can significantly benefit from machine learning.
So how does machine learning change the way businesses operate
First thing’s first: Machine learning needs training data and training data costs money. Especially training data labelled by humans.
Let me explain. To make machine learning work for business, the algorithm needs to see lots and lots of examples of what it’s supposed to be doing. If you want an algorithm to tell you if a sales lead is good, you need to show it lots and lots of examples of good sales leads and bad sales leads. If you want an algorithm to tag your support tickets you need to show it many examples of support tickets. If you localize your algorithm to a new language you probably need to collect lots of examples in that language.
In some instances, a company may have those training sets in house. For example, a bunch of disqualified or qualified leads. But say you haven’t labelled each of your support tickets as they’ve come in over the year. You’d need to have people -- either in-house or en masse via a data enrichment platform -- label those tickets. The machine will then look at those judgments and start finding connections and patterns it can learn from.
Machine learning is much cheaper and more efficient than people when it works well. The downside is that it often it works well in 80 percent of the cases and badly in 20 percent of the cases, and lowering the 20 percent error rate is hard, if not impossible.
But even an 80 percent accurate algorithm can save you a lot of money because good machine learning algorithms know where they are accurate and where they are more likely to have errors. Smart companies take the cases where the algorithm has high confidence and uses those directly while sending low confidence cases to humans.
Banks have been doing this for years. When you put a check in an ATM, an algorithm tries to decipher the numbers on the check. If you have really sloppy handwriting or the ink is smudged the algorithm passes the task to a human. This design pattern saves banks lots of money while preserving a very high level of accuracy.
A huge benefit of machine learning is that it can turn part of your variable cost into more of a fixed cost. If you use humans to handle cases where that algorithm is struggling, you are creating the perfect training data to feed into your algorithm. This is a well studied technique called active learning -- it turns out that training data labels collected on cases where the algorithm has low confidence helps the algorithm learn much, much more efficiently.
As the algorithm becomes increasingly more accurate, the unit economics of your business process become better -- and as the machine learning becomes able to handle more cases, the expensive humans are only called in on the toughest, rarest situations. That means you use the best of both human and machine intelligence in tandem: leveraging the speed and reliability of computers for the easy judgments and the fluency and expertise of humans for the difficult ones. And if that sounds like smart business, it’s because it is.