Machine Learning Engineering and Operations (MLOps)
Whether physics-based, domain-knowledge-driven process models or wholly empirical data-driven models are used (or a combination of the two) for environmental system evaluation, these models become stale as new data becomes available. ML Engineering (or MLOps) combines machine learning model development with DevOps principles for continuous improvement/integration and continuous delivery/deployment (CI/CD). This is a relatively new field of practice and study, having just emerged in the past decade at tech flagships like Google, Facebook and Amazon.
In this section I’ll collect the best resources I find in my efforts to understand the objectives and requirements of MLOps, and begin to outline my own take on the subject as it intersects with the more traditional model development and management lifecycle of water-quality and environmental infrastructure modeling.
Topics to be considered:
- Technical debt
- MLOps maturity model(s)
- Standards development
- Data drift
- Concept drift
- Model tracking systems
- Containerization
- Distributed computing
- Connection with SimOps
References and Resources
Are compiled for this section here.
Helpful resources
My first resource for getting a handle on MLOps was Andrew Ng’s online certificate program for MLOps. I highly recommend it, and many of the particular technical references discussed throughout this section were brought to my attention in that course.