AI Model Monitoring & AI Governance
Satisfy Risk Management & Compliance Requirements by Monitoring Models for Drift, Bias, & Performance (AKA “Explainability”)
Why AI Model Monitoring & AI Governance?
ModelOps and getting your models into production is just the beginning. Soon, your Compliance & Risk team will be asking you difficult questions. Graduating to a sophisticated AI/ML program means you are prepared for the inevitable audit. But what does that mean?
Sophisticated AI/ML programs monitor their production models for drift, bias, performance, and anomalies, ensuring they identify potential issues and correct them immediately before they become a business problem.
Governing your models, ensuring they are Reliable, Explainable, and Responsible is paramount to the longevity and profitability of your AI/ML program.
Enterprises leverage the following list of features to govern and monitor their AI/ML models in production.
AI Monitoring & AI Governance Features
The ability to quickly get an overview of the health of your AI models is a crucial requirement to ensure that you are not exposing your business to unintended risks with severe consequences.
The Dashboard provides a high-level overview of your AI/ML program without having to decipher the nuances of all the AI models running in the environment.
For the DevOps/IT/Engineering teams, Datatron provides infrastructure monitoring as an out-of-the-box feature for all machines in a cluster. Activities monitored include CPU usage, memory usage, health check, and more.
Overall Health Score
A proprietary “Health Score” provides an intuitive, easy-to-understand measure of the overall health of all the models in the system.
Bias, Drift, Performance, Anomaly Detection
Bias is calculated as the difference of a particular variable’s distribution between a group of interest and the rest of the population.
Datatron supports bias monitoring in four scenarios:
- Regression without feedback
- Regression with feedback
- Classification without feedback
- Classification with feedback
- Data Drift – Also known as covariate shift, is changed after a model is deployed. Each data point lives in a high dimensional space whose dimension is defined by the number of input features. If data drift occurs after the model is deployed, the new data points will likely come from a region of the input space less populated by training data.
- Concept Drift – Concept drift occurs when the decision boundary changes, and is often observed in time-series or streaming data.
Performance metrics for each model measure show how well the model is performing in production. Typical metrics monitored include R2 , root mean squared error (RMSE), mean absolute error (MAE), explained variance, and more.
Datatron’s proprietary anomaly detection mechanism identifies potential issues via an ensemble approach by gathering data and metrics from multiple sources, including model performance, bias, drift, model logs, system logs, and more. Thus enabling Datatron to more swiftly and accurately identify potential issues not found in traditional metrics calculations.
Customers can set alerts and/or automatic shutdown defaults triggered when the model varies from pre-defined performance thresholds. These notification mechanisms act as a continuous feedback loop – when the KPI is impacted, the platform notifies the user.
You can select the type of monitoring and how to be notified, such as via email, Slack, PagerDuty, or other methods. Or, if you prefer, leverage Datatron’s API to integrate with your own preferred alerting tools.
Monitoring and reporting can be customized based on an organization’s KPI metrics. Datatron provides easy connection mechanisms to pull data from customer data sources.
Activity Log & Audit Trail
Datatron maintains thorough information on models and datasets, such as versioning, history, and user information. Datatron can efficiently track down old information for review. Datatron also maintains complete logs of request responses for all models.
Leverage logs generated by Datatron to produce a complete audit trail. Users can now observe how their system behaved at a given point in time, as well as the circumstances around the system’s behavior, dramatically accelerating the complete audit process.
Go back in time to narrow down the model used for a particular prediction, along with the corresponding information:
- Dataset used to train the model
- Version of the model deployed
- Feature vector of the model
- Model prediction value
- User who approved deployment of model
- Time of deployment
- Comments on discussion and reviewer forums at the time of model deployment
Why Business Executives, BU Leaders,
and CDO’s Love Datatron
Business leaders are tasked with driving profitability from their AI/ML program, but often encounter numerous growing pains, including model operationalization, team dynamics, and performance, not to mention inquiries from Compliance and Risk partners. Here are just a few of the harsh realities:
- Not getting ROI from AI/ML
- Team, role and expertise challenges
- Program-wide audits
With Datatron, business leaders empower their teams to create more models, operationalize their models faster, and understand how their models are performing – when the inevitable audit happens, they are prepared. All of this ensures a viable, explainable, and profitable AI/ML program.
Your AI Program Deserves Liberation.
Datatron is the Answer.
See what major Enterprise Brands have already discovered about
Datatron’s production-proven, Enterprise-grade AI Platform.
Our Latest *** RELEASE ***
Datatron continues to innovate and improve data scientist workflows with our latest release, which includes a JupyterHub Integration, Kubernetes Management and more Enterprise-grade features.