Stephen Fry – a much-loved British comedian with a ferocious love of new technology – commented recently that artificial intelligence and other new technology innovations will transform society even more than the industrial revolution or the introduction of the printing press.
He may be right. AI, and more specifically machine learning, are generating enormous buzz in many areas, with different solutions claiming to place the perfect advert or detect every single fraudulent transaction.
But are these claims realistic?
Gartner commented last year that machine learning had become a buzzword – or presumably buzz phrase – and was at the peak of inflated expectations within their hype cycle. People rarely shout about AI and machine learning failures – but they certainly exist.
Machine learning works by recognizing patterns using complex computer techniques and algorithms. A core reason it is now gaining such notoriety is that there are a wealth of openly available technologies that make it simpler and cheaper to build and run machine learning algorithms.
That’s great news for data scientists. Many of the tools are cheap if not free, and public clouds like Microsoft Azure and Amazon Web Services mean even massive data crunching exercises can be run efficiently without the need to buy tons of hardware.
The success stories can be mind-blowing. At FICO, we see examples all the time. A large bank in the UK using FICO Falcon Fraud Manager regularly processes 300+ transactions a second for sustained periods with average fraud screening time of 80 milliseconds and the fastest checks taking only 5 milliseconds. Machine learning is doing a brilliant job of balancing fraud detection and customer experience.
So where’s the downside? Where’s the over-inflated hype? Let’s start with a dose of reality.
What Machine Learning Can't DoMachine learning isn’t magic – it won’t solve every problem – and it won’t configure itself. The current excitement will dwindle as the wilder claims are unwound.
For instance, spotting card and account fraud is a very complex problem and domain knowledge is a very important ingredient in building effective models. Without a healthy respect for how AI can go wrong, you can learn relationships that are totally spurious or overtrain a model so it is blind to emerging patterns.
Be wary of anyone relying too heavily on the computer taking the strain. Some think they can point their technology at any problem and find all the anomalies. They think they can find the needle in any haystack. However we see exceptions all around us. Just ask Siri. Or think about Facebook and Twitter’s inability to detect fake news and the devastating consequences it has had on our politics. These are the visible manifestations of what happens when a model doesn’t cut it.
While machine learning is incredibly powerful, it is not a 300 tps Sherlock Holmes. Consumers do the oddest things and fraudsters can closely mirror normal behavior. Spotting the difference requires more than pure processing horsepower. FICO combines deep domain knowledge with behavioral analytics to maximize results.
As our Chief Analytic Officer, Scott Zoldi, said recently in this blog post on data leaders: “Machine learning is touted to solve the malaise of Big Data. When machine learning finds patterns in data, data leaders ensure it’s not arbitrary correlations but look for causation."
You should be very excited about machine learning and what it can bring to your organization, but exercise some restraint. If something sounds too good to be true, it probably is.