«MLOps is a set of practices that bring people, process and platform together into a stream-aligned process to manage End-2-End Machine Learning lifecycles.»
MLOps is taken about a lot, so I asked an expert what we are actually talking about. Xiaopeng Li is AI business lead at Microsoft for the Western European market, located in Oslo. Xiaopeng is a passionate influencer in the field of Data and AI/ML, who was nominated as AI influencer of the Year at last years DAIR-awards in Stockholm.
Here are my key takeaways:
Patterns in AI adoption
- AI adoption projects are quite diverse, but with some patterns that are visible across. Here are use-cases that a lot of industries are working with:
- Business Process Automation as an AI use case
- Adopting AI to process documents automatically and extract key-values
- Natural Language understanding and processing, but also Natural Language generation
- Chat GPT
- Knowledge Mining
- Unstructured data analysis
«Nordic countries are at the forefront when it comes to adopting AI and ML»
- Some of the most advanced search capabilities used in Microsoft are developed in Norway
- Nordic countries are typically quite tach-savvy
- Nordic countries have very good infrastructure
What is MLOps?
- MLOps is about agility, productivity, consistency and quality
- It is about creating scalability for your Data Science work
- MLOPs is a vage concept and you can probably find a variety of different definitions. Is MLOps at the intersection between DevOps, ML and Software Engineering?
- Scale ML development and deployment with constancy, with quality, with speed
The three elements that are most important are people, process and platform
- People:
- 5 particularly important roles: Stakeholder, Cloud Infrastructure Architect, Data Engineer, Data Scientist, Machine Learning Engineer
- There are many different roles involved in MLOps, from cleaning data to testing a model an implementing it. These roles need to be orchestrated
- Domain experts and stakeholders play a critical role in defining the challenge in the first place. They can formulate what to achieve and what is good enough
- Change Management is important, especially if your ML implementation triggers behavioral change
- Platform:
- You are in need of a secure, scalable infrastructure to would your models on
- Mature organizations who do ML at scale, have most an integrated architecture for Data Management, Analytics and Machine Learning
- Process:
- Data collection,. Data processing and data management are processes you need to focus on in MLOps
- You need a process and the right competencies to gather use-cases in the first place
- Build a backlog of initiatives and then go through prioritization based on eg. Data availability, feasibility of solution given current etch-landscape, value for business, cost, time to marked,..
Path to MLOps
- Always start with assessing your current landscape and maturity
- Start by assessing your platform capabilities
- Ensure you have the right competencies and people
- If you want to operationalize MLOps, don’t look at it as a technological problem, but something that includes the entire organization
- Key is to bring key stakeholders as early as possible into the discussion
Oslo AI:
https://www.linkedin.com/company/oslo-ai/
https://www.meetup.com/oslo-ai/
Link to MS learning:
MLOps Maturity Model
