Improving ING's Machine Learning as a Service
Machine learning is a very imprortant feature for several ING products. For this reason, ING is investing in Machine Learning as a Service, an internal team to make ML a generally available feature for all interested ING teams. There are various ways that the ING’s MLaaS can improve, some topics of interested are listed below:
Combining live and historical data in ML models: How can we perform analytics operations that combine live and historical data in an efficient way? How can we live-retrain ML models when the current model fails with current live data. A potentially interesting approach would be to use TU Delft’s Codefeedr project for this reason.
Monitoring ML models: Data provided as inputs to ML models can change, as a result of new services being deployed or due to temporal effects. How can we effectively monitor that the currently deployed ML model is suitable for the data it processes?
ML pipeline deployments: How can we efficiently train, test-deploy and really deploy ML models? How can we A/B test ML models? How can we collect analytics and trace data for currently deployed models? The purpose of this task is to remove the human factor from deploying and monitoring ML pipelines.
Contacts about the project
- Georgios Gousios (TU Delft)
- Arie van Deursen (TU Delft)
- Hennie Huijgens (ING)