IBM this week revealed it has added a drift detection capability to the Watson OpenScale platform to govern artificial intelligence (AI) models that will become a foundational piece of IBM’s approach to defining best DevOps practices for building AI-infused applications.
Rohan Vaidyanathan, program director for IBM Watson OpenScale, said one of the biggest AI issues organizations face today is determining when to update or replace an AI model. Announced at the IBM Data and AI Forum event, the drift detection software added to Watson OpenScale provides a continuous monitoring capability that detects how far an AI model has moved from its original parameters, Vaidyanathan said.
Drift in AI models usually occurs over time, especially as use cases change in ways that are unexpected. Once that drift is detected, organizations eventually want to either retrain that AI model or replace it with a new one. The process challenge organizations face when building AI applications is first assembling the right data to train the model and then integrating that model within an application as part of a continuous integration pipeline.
With each interaction with an AI model, the result returned to the application might be different as the machine learning algorithms embedded in the AI model continually learn about the process the application abstracts.
As it turns out, Vaidyanathan said, only one in 20 of the AI models being built make it into production applications, largely because organizations have not bridged the divide between DevOps teams and data scientists effectively. The drift detection capability IBM has developed is a “missing link” that will enable those two teams to begin collaborating more closely, he said.
IBM has already defined a set of best practices for building AI models as part of its Watson Anywhere initiative, which seeks to drive adoption of the various elements of the Watson platform on multiple clouds. Rather than forcing organizations to load data into an IBM Cloud to use Watson, IBM is working to bring the various engines and application programming interfaces (APIs) to wherever data resides.
Naturally, IBM would prefer organizations employ the opinionated DevOps and DataOps processes it defines as best practices for building AI models using its tools. However, Vaidyanathan acknowledged there is likely to be plenty of instances in which organizations will prefer to define their own DevOps processes. The hope is those organizations will still access Watson services and metrics using the APIs that IBM has exposed.
The challenge IBM is encountering in terms of driving AI adoption are manifold. In addition to trying to get data scientists to appreciate DevOps, IBM recognizes AI models are only as good as the data employed to train them. Unfortunately, many organizations have been lax when it comes to data management. As a result, many of them are now revisiting broken DataOps processes. IBM Watson OpenScale and IBM Cloud Paks for Data, a set of portable data virtualization and management tools deployed as microservices, are part of an effort to bring both explainability and order to the building of AI models.
Obviously, it will take some time for organizations to extend best DevOps practices to incorporate AI models. But at this juncture, it’s more a matter of when rather than if that goal will be achieved.