When machines learn how to learn

It sounds like science fiction: instead of having humans create a computer algorithm to solve a problem, computers teach themselves by being shown examples, and then generating their own models to solve it. Such solutions can work better, faster and more comprehensively than those made by their human “teachers” ever could. Today, there is certainly no shortage of use cases for this future technology. The enormous potential for efficiency also makes this technology attractive to online banks like ING. In order to support its specialist departments in leveraging this technology, Volkan and his team are developing a comprehensive solution as a service: “Machine learning as a Service (MLaaS) is not only a tool for building models, but also a package that includes everything necessary to take a model live: from preparation through to production deployment.”

Computer making themselves useful: the use cases

Machine learning use cases at ING are as diverse as the specialist areas themselves: data extraction, fraud detection, IT monitoring, or document classification. Just this last example has traditionally involved a rather high – and not exactly productive – degree of manual effort. Documents are received via different channels and then sorted into types, such as contracts or documents on construction financing, and assigned criteria, such as origin and destination. However, in 40 percent of the cases, traditional tools fail. This then requires a team of 32 employees to manually process the remaining documents. With machine learning, a trained model can reduce this failure rate to 20 percent – a significant reduction that frees up manpower for more productive tasks. In other fields, such as fraud detection, human supervision of the algorithm results remains necessary. Suspicious cases are only flagged and then sent to a “humanoid” processor for final classification. The results can then be fed back to the system as new training examples to improve the models.

Putting theory into practice

The MLaaS offering now ensures that use cases, like the ones described above, not only sound good in theory but also can be implemented in practice. The model that a data scientist is working on may look interesting, but deploying it in the real world is a completely different matter. As Volkan explains, “The questions arise again and again as to how a model should be deployed in production, who will be maintain it once it’s deployed, and who will be responsible for its ongoing governance. And this is repeated with every project.” All too often, projects fall by the wayside because of such problems. With MLaaS, Volkan can offer a central point of contact for his “customers”, i.e. the data scientists from throughout the organization who want to create solutions for their specialist departments. In the future, they will no longer need to plan the operation, maintenance and updating of solutions, because MLaaS inherently offers them secure and tested environments. The “productization” of projects can then run smoother and faster, reducing costs significantly. This approach also increases the independence of data scientists from other teams and units, makes it easier to build models and simplifies deployment. Last but not least, Volkan and his team can offer comprehensive end-to-end support for the entire life cycle of the product. Pure DevOps in practice – with all the advantages.

A vibrant mosaic: tools & modules

How can MLaaS ensure a smooth project flow? MLaaS provides a wide range of tools and proven workflows as a catalogue of services. Customers can then pick what they require almost as simply as “select, order, use”. As Volkan puts in, “We support the project from start to production, deliver an environment with tools specifically tailored to the project, and also offer standardized pipelines.” This ensures that models can go live with minimal effort. The basis of the MLaaS solution is jupyter, a coding environment in which the data scientists design their models. If jupyter's resources are not sufficient for a specific project, optional modules such as the calculation clusters Spark or Dask can be added. In addition, the MLaaS workflow uses MLFlow, a model repository that stores finished models for future use. The components can be exchanged or extended at any time: a modularity that is possible because the entire project environment runs in a private cloud (OpenShift). Since ING is subject to banking-specific regulations on data handling, some applications have been developed in house, such as an S3 proxy that can be installed “on premise” as well as a system for convenient and secure loading and saving of data.

A question of organization: IT engineering for ING

Fast, affordable, modular and future-proof: the MLaaS approach makes a lot of sense. Of course, there are also hurdles to implementation. “It's not easy,” summarizes Volkan. One particular challenge, for example, is collaboration. In a global corporation like ING with its many international subsidiaries, any project which integrates different modules and business areas requires extra effort to coordinate changes. This dynamic bank’s consistently agile way of working certainly helps. Engineering professionals like Volkan work together with data scientists and subject matter experts in squads on a project-by-project basis. The agility of the organization is one of the many plus points that all IT jobs at ING offer: in addition to a global IT community, there is plenty of opportunity for further development and a generous budget for further education, as well as exciting professional challenges. In the case of MLaaS, this includes further expansion of the service in areas such as lifecycle management, automated deployment and the repository. ING’s offices in Frankfurt am Main and Nuremberg are constantly seeking new employees. If you are an IT specialist curious about innovative topics, find out more about IT opportunities at ING on our jobs page.