Unlike other software, AI models need constant monitoring and regular retraining. AI models may start to deteriorate as soon as they begin to interact with the external environment. This is because the statistical properties of a target variable (that the model is trying to predict) may change over time in unforeseen ways (Pechenizkiy & Zliobaite, 2010). This is called concept drift. Concept drift needs to be monitored post-deployment, and experts (ideally people who are knowledgeable about both data science and the business domain) should be assigned the responsibility. Recalculations of the model should happen when it reaches a particular pre-set threshold, for instance, on criteria such as accuracy or stability. 

There are various ways to create a robust monitoring system and provide feedback loops. A monitoring system can include but is not limited to: a continuous examination of dashboards showing key indicators; the monitoring of gaps between the distributions of training datasets and online datasets; quality measures of new incoming data; accuracy measurements of model predictions; and changes in the use of AI products or services by internal or external customers. 

Additional monitoring measures should be established, such as having individuals keep an eye on the business environment to see if it has significantly shifted from when a model was first trained. If circumstances have substantially changed, it is essential to run tests to see what changes have happened to the model recommendations before the new circumstances vs. after. Further, regular re-testing of models (especially under different scenarios/simulations) would allow an organization to stay one step ahead of adversarial attacks.

It is also essential to maintain sensitivity towards customer reactions through various other monitoring mechanisms, such as customer feedback surveys. Such surveys may assist an organization in knowing whether an AI product or service is fulfilling customer needs effectively and consistently. 

An organization needs to set up guardrails to avoid unintended consequences after AI model deployment. When a monitoring system indicates that an AI model is not working effectively, timely actions need to be taken. These actions should be increasingly stricter as an AI model moves away from its given thresholds. These actions can range from exception reporting to escalation to recalibration of the model to immediate decommissioning. AI model can also be set up to not provide any answer when its level of confidence goes below a certain threshold. Further, as changes are made to an AI model, there may also be changes needed within the automated business processes using a particular AI model. 

Similar to data life cycle management, for AI activities, there is AI life cycle management. That means determining the process of bringing a particular algorithm on board, managing it, and then retiring it or replacing it when it does not serve effectively anymore. 

Boards need to realize that real work with AI comes after deployment and operationalization. Constant monitoring results in higher marginal costs of operating AI products compared to regular IT software products. 

Author: Dr. Jodie Lobana

References Included:

Pechenizkiy, M., & Zliobaite, I. (2010, October). Handling concept drift in medical applications: Importance, challenges and solutions. In 2010 IEEE 23rd International Symposium on Computer-Based Medical Systems (CBMS) (pp. 5-5). IEEE Computer Society. 

Music Producer photo created by senivpetro – www.freepik.com