This podcast, “Caveat Emptor,” is the second episode in a five-part series featuring Webster and Madhu.
“A lot of companies throw the ‘AI’ or ‘machine learning’ moniker onto their brand when they’re not necessarily using the current, state-of-the-art systems,” said Sunil Madhu, CEO and president of digital identity verification company Socure. “They’re still relying on technology created nine or 10 years ago.”
The difference? Older machine learning is part of a rules-based approach – that is, it follows rules defined by its human creators – while advanced machine learning means algorithms can learn by themselves. Rules-based learning is reactive, said Madhu, requiring people to go in and re-program the system when new patterns are discovered. The new algorithms re-program themselves.
For example, a bank may teach its algorithm the rule that opening more than 10 accounts a day is an activity that signals fraud. But when fraudsters find a way around that – perhaps creating one account a day for 10 days – then a person has to create a new rule before the machine can recognize that pattern. After a few years, the system gets bloated with rules and becomes unwieldy to manage.
Modern systems don’t need someone to define the rules
The machine writes them itself, and it writes them better than a human.
“Human beings create rules based primarily on human intuition,” Madhu said. “Machines can go beyond human intuition and see patterns that a human being might not be able to intuit, and that’s the thing that separates these two types of solutions at the market.”
Patterns are hard to find. You’ve got to comb through a lot of data to determine correlation and causality. For a human, that process can be long and laborious, even infinite. For a machine, it’s possible to take the bird’s-eye view and spit out a nominal outcome, such as “fraud” or “not fraud.”
So, which AI is the best? It depends what you want it to do.
An artificial neural network is good for pattern recognition, computer vision and speech or text processing, which sounds like a good candidate for fraud recognition – but the system lacks transparency, a key selling point for financial institutions.
“Neural networks are kind of a magical system as to internally what happens – the neurons learning and the activations triggering in those neurons,” Madhu said. “No one knows what’s going on inside a neural network, inside its hidden layers, and that becomes problematic.”
For fraud prevention, you’re better off using an AI with linear pattern recognition. This is also useful for email filtering and network attack detection. By contrast, non-linear pattern recognition is good for calibrating inputs that don’t linearly affect the output of the system.
By now it’s clear that there’s a spectrum of technologies and approaches available to buyers. Socure achieved that variety in the same way it achieves pattern translation: with robots. That’s right: It built a robot to handle development of its robots.
Development happens in three stages: data engineering, model development and training, and deployment.
According to Madhu, “Getting the data right is the hardest part.” Stage one is manually intensive, requiring Socure to distill data, determine whether it’s good data, transform it and extract predictive features – the red flags that indicate fake identities or fraud.
Over the years, Madhu said, “We’ve essentially honed in on what constitutes good data, what are the range of problems we can see in data, how the data should be transformed to get effective features. So, what we’ve essentially done is, having studied all that, we’ve created robots.”
“Our company has uniquely created an identity verification robot, and that robot has little minion robots working for it,” Madhu explained. “One of the minions does data engineering. We’ve taken the repetitive work and trained the machine to be able to do that work.”
At stage two, a different robot handles model development. The process is trial-and-error. Luckily, a robot can try hundreds or thousands of models and can weed out the ones that aren’t useful much faster than a human, thus discovering the most effective machine learning method for training new algorithms to achieve their goal.
Finally, at stage three, the process has been completely automated as an AI takes the predictive model, packages it as a service and applies it to infrastructure. This bot also watches to see how the machine is working and to ensure that the model’s performance will always be predictive, even if real-time events get in the way (for instance, if a vendor’s output has timed out).
If robots making robots sounds futuristic, just wait: “We are extricating human beings from the process of data science and letting machines and robots do the job for us,” Madhu said. “Humans are and will still be involved, but transferring tasks better suited for machines will just continue to improve our results.”
“We hope to get to a point where we’ve made that robot smart enough that it can automatically trigger the model retraining or generation robot by itself so we can have an endless feedback loop with these robots constantly ensuring that our models are performing,” he said.
In the next episode, Madhu will introduce Webster and listeners to his “robot army” and will delve into the data engineering process, shedding light on the decisions that must be made about data inputs and known sources.