Use of Machine Learning models is associated with their own set of risks. Like any other risks to an organization, the risks associated with Machine Learning models needs to be managed. MLOps practices address these risks to a great extent. In this chapter of the MLOps tutorial, you will learn about Risk Management side of MLOps.
Apart from the usual risks that are associated with data and software, use of Machine Learning in any organization has addition risks associated with it. There are several sources from which risk originates, the most notable ones being-
- Risk due to loss of Human Resource
- Regulatory Risk
- Errors and Misuse in Production
Risk due to loss of Human Resource
The risk due to Human Resource is not just limited to Machine Learning model, but is amplified in the field of Machine Learning. The risk can originate due to a human resource, such as a Data Scientist or a Machine Learning Engineer leaving the organization, or transferring in the same organization.
This risk can be mitigated by creating and maintaining proper documentation and versioning of all softwares and models being used. Regular Knowledge Transfer sessions to other members of the team about the code and models can prove helpful to address the risks associated with loss of Human Resource.
Regulatory Risk
The Machine Learning models being used in production are often in scrutiny for regulatory reasons. This is mostly because their black-box nature. The business and organization can have adverse effects from models being biased not only financially, but in intangible ways such as loss of reputation, etc. This is even more true in case of heavily regulated industries such as healthcare, finance, etc.
Use of MLOps best practices such as version control and along with the concepts such as Reproducible AI, Fair AI, Responsible AI, and Explainable AI can help to cope up with regulatory risks associated with productionizing Machine Learning models
Errors and Misuse in Production
Any Machine Learning model in production is prone to making errors and occasionally, subject to deliberate exploitation of these errors. Apart from the standard errors associated with software such as bugs in the software while running in the production environment, Machine Learning models are also subject to the effects of input drift in the data- the difference in the inference data as compared to training data. The model is also exposed to deliberate exploitation by the use of Adversarial Attacks. Apart from these, the model can, in general, have lower accuracy in production as compared to what was expected of it after testing.
The use of MLOps practices such as constant monitoring and feedback can reduce the errors and misuse of Machine Learning models. Thereby helping in the management of the risk associated with errors and misuse of Machine Learning models.