Seamlessly govern AI models with AI Factsheets and IBM OpenPages
This post has been co-authored by Shashank Sabhlok (Product Manager — AI Governance) and Marc Cassagnol (Product Manager - OpenPages MRG), IBM Data & AI
W e are seeing prolific growth of AI models in all industries as more and more organizations are making use of AI to solve business problems in areas like human resources, customer service, marketing, consumer behaviour, fraud detection and finance. In the financial industry for example, where institutions already have mature model risk governance programs, there is a growing shift to replace legacy models with AI to help reduce costs, improve efficiency, and remain competitive with their peers. Model risk is no longer something that only financial institutions need to worry about though. AI models are inherently more complex and difficult to understand, which increases their risk — making it important to implement Trustworthy AI technologies that constantly govern and monitor models for aspects such as fairness, explainability and data drift.
A critical pillar of Trustworthy AI is AI governance, which is an organization’s ability to manage and monitor AI activities. Today, this requires:
- Reconfiguring teams to support AI/ML development
- Implementing technologies to monitor and validate whether the AI is operating as the organization intends and as its stakeholders expects
- Constantly improving risk assessment methodologies to ensure that teams have all the information at their disposal to accurately assess model risk
- Robust plans to address governments expanding regulations in this space. A categorical example being the SR 11–7 regulation set out by the US Federal Reserve Bank that provides guidance on elements of comprehensive model validation.
Model validators have an increasing challenge in performing their validation, and with the added complexity of AI models, having all the information needed at their disposal to perform their validation is necessary. In the past validators have had to go out and gather facts about a model manually, which is a time-consuming and error-prone process. Our latest integration between AI Factsheets and OpenPages changes that by facilitating end to end automated AI model risk governance.
Introducing the AI Factsheets — OpenPages Integration on Cloud Pak for Data 4.5
AI Factsheets, now available on-prem with Cloud Pak for Data 4.5, automatically captures metadata across the model development lifecycle thus facilitating subsequent enterprise validation or external regulation. We are excited to announce its game-changing integration with IBM Open Pages Model Risk Governance (MRG). MRG facilitates end to end AI Governance for risk and compliance through its powerful workflow capability that allows users to setup their own custom steps and processes within the tool. The marriage with AI Factsheets means model and deployment metadata is automatically synced with OpenPages, allowing model validators, risk analysts and business users to make informed decisions about an AI model and it’s lifecycle faster.
MRG now allows for the creation of Model Entries, which are essentially model use cases that can be created by a data science leader or model owner within your organization. The creation of the model entry in OpenPages automatically creates an equivalent model entry in Cloud Pak for Data and sends out an e-mail notification to the data scientist, asking them to associate a model to the model entry (or model use case). Once the data scientist is done developing the model (and has automatically captured model facts using the AI Governance Facts Python SDK), they can add the model to this model entry and also view its Factsheet.
Back in OpenPages, the addition of the newly created model to the model entry triggers the model candidate workflow, where the model goes through the necessary approvals before being added to the inventory. As the data scientist continues development of the model, the model development workflow enables the data scientist to collaborate with the model owner to document the definition and planning of the model development. During development, facts are being replicated from AI Factsheets to OpenPages to aid in the risk assessment process which usually kicks off during development and is performed by the model owner.
Once development is completed, pre-production deployments can be created in Watson Machine Learning and monitored with Watson OpenScale, which also integrates with AI Factsheets and OpenPages to bring in model performance metrics such as bias, quality and data drift. The model validation team performs its pre-implementation review against the pre-production deployment, and documents this process in OpenPages using the model validation workflow. After the model passes, the head of model validation gives the final approval that the model is ready for deployment to production. At this point, it’s also good to note that the ability to set thresholds in OpenPages means that metric non-compliance (e.g. severe data drift) can easily be discovered on the platform and factored into the validators decision making process.
As the steps above indicate, with OpenPages and AI Factsheets, the model validator always has easy access to all model metadata required to be able to make an informed decision and ensure their organization’s AI Governance is comprehensive.
If you are already an IBM Cloud Pak for Data-as-a-Service customer, visit the Gallery to try out the MLOps and Trustworthy AI tutorial. Learn more about the advantages of Trustworthy AI, and how it drives responsible business transformation here. To understand how OpenPages supports model risk compliance click here.
