Data Science & AI

CIOs Can Remove Roadblocks To Adopting ML: Here’s How

Zest AI team

August 29, 2019

The chief information officer occupies a unique position in a bank’s leadership. They stand at the intersection of business and technology, which makes them key players when it comes to adopting machine learning (ML).

And while a few years ago that might have seemed like a minor concern, ML is becoming an increasingly important part of most CIOs’ everyday lives. Last year, Fannie Mae surveyed executives at 200 banks and found most expect to be using ML in some manner by 2020 — more than double the percentage of institutions using it today. ML, an offshoot of artificial intelligence (AI), can help banks make faster, better decisions about underwriting. Using ZAML, ZestFinance’s ML software toolbox, banks see an average 15% increase in loan approval rates without any increase in defaults.

But many bank CIOs get pushback from within their organization when they recommend moving certain functions to ML. While it’s understandable for people to be nervous about new technology, most of the concerns coming from the IT department are overstated. Here are eight roadblocks we’ve heard when talking to CIOs about ML and the reasons why each of these is easy to remove.

“ML needs specialized hardware.”

The biggest perceived problem with adopting ML is that the technology requires hardware that doesn’t match the bank’s current infrastructure. But the truth is that ML can run on regular machines. It doesn’t need fancy tech tools such as field-programmable gate arrays or graphic processing units. ML can be deployed on local servers or in the cloud using the bank’s existing infrastructure.

ML isn’t compatible with my vendor platforms.”

Many bank IT infrastructures operate on multiple third-party hardware and software platforms. The idea of integrating ML across those platforms can be daunting. But again, these worries are overblown. Machine learning works on common software platforms including Apache Yarn and Microsoft SQL. It also works with standalone Unix boxes, Open Database Connectivity (ODBC) APIs for transferring data back and forth as well as almost any microservice architectures that enable flexibility and interoperability.

“ML is self-learning. We’ll lose control.”

Not all ML is self-learning. Since banking is an industry that operates under very restrictive regulations, it’s better for banks to use ML models that don’t automatically update and are instead trained on a discrete data set. At the same time, you need to make sure you work with an ML vendor that provides robust monitoring tools to ensure model performance and your ability to adjust models to reflect changing environments.

“ML is based on open source and therefore can’t be managed.”

While most ML is open source, the open-source version management problem is solvable. Using tools like Docker — which containerizes open source applications in your system – you can pin all versions, packages, and dependencies in a certain place. Your IT department can lock down the network to mitigate open source risks, as opposed to allowing automatic downloads, and update code at their discretion.

“ML forces you to invest big bucks in a data lake.”

Data lakes are a relic of traditional model building and aren’t always needed with ML, which can pull data from the system you already have. ML models get developed on whatever data you have and then operate in that same environment. Where once you needed a data lake to perform feature engineering, now you just build, test and go, without the complicated middle step.
“I need to have my cloud strategy already in place to migrate my data to ML.”

The truth is that ML can run on regular machines. It doesn’t need fancy tech tools such as field-programmable gate arrays or graphic processing units.

Nope. In fact, modern architectures allow you to operate your ML model using on-premise servers or any cloud seamlessly using Kubernetes technology, which automates application deployment, scaling, and management.

“ML bias is inevitable and impossible to detect.”

ML bias is a real concern for banks. If your model is suddenly rejecting potential borrowers who fall into protected classes, you can expect a visit from a government regulator (and a significant loss of business) if you can’t understand and fix the problem. But ZAML offers true explainability giving banks the ability to see why their ML model made each decision. ZAML Fair, which is included in every ZAML model, monitors performance to ensure fair lending practices.

“Well, OK. But ML is expensive.”

Short-term, it may appear cheaper to stick with your tried-and-true model, but ML pays off in the longer run. From a business point of view, ML can boost the bank’s bottom line by improving lending rates and reducing charge offs. Operating ML models is cheaper than using traditional linear regression models because you don’t have to recode the model every time you change it. And modern ML frameworks, like ZAML, enable you to deploy the model directly into production without recoding from SAS, reducing development cycle time from months to weeks. And ZAML includes automated QA testing and validation routines so you don’t need dedicated teams.

In banking, any kind of change is hard. But it doesn’t have to be. Once teams understand the benefits and ease of implementation, their ML concerns tend to fade and the bank can move into the future of technology.

Thank you for subscribing!
Something went wrong while submitting the form.