Inside the MLOps lifecycle stages
Developers tasked to train machine learning models are turning to the MLOps lifecycle. The different stages are meant to increase operational speed and efficiency.
Businesses that create and train machine learning models face many complicated aspects in the process, like training models with quality data sets and continuously updating them with new features. The MLOps lifecycle ties these aspects together to make the process easier.
Machine learning operations (MLOps) use DevOps principles and techniques to solve problems associated with producing ML applications. The DevOps approach delivers a constant stream of both bug fixes and feature enhancements to an application or service. DevOps exists to speed up implementation of new code releases for the application in which an ML model is embedded, as well as the ML model itself.
MLOps also optimizes the roles of staff. Because MLOps automates many tasks, employees can spend time performing their core skills and activities. Companies, meanwhile, can avoid seeking employees that have to handle everything from data engineering to model building to application development to production operations.
To adopt an MLOps lifecycle, teams must first understand what this idealized lifecycle is, as well as who it has benefited.
An idealized MLOps lifecycle
MLOps takes the DevOps application lifecycle and lays out a similar lifecycle for ML model development, which consists of gathering and preparing data, designing and training a model, then evaluating and releasing that model. An enterprise typically begins its journey with an intended use case. This use case informs every aspect of data selection and model and application design.
The MLOps lifecycle includes four key sub-cycles:
- The data cycle.
- The model cycle.
- The development cycle.
- The operations cycle.
Each cycle feeds information forward and backward. For example, information from model development can influence successive data cycles. Each cycle also has a specific function. In the first one, teams create and manage data sets for model development. The reason an ML model exists is to distill meaning out of a data set or data flow. It is meant to spot patterns and bring information to the surface. That model can only work when it's trained to perform that function using a carefully curated data set or many data sets.
ML teams must identify and gather data sets from appropriate sources. For most models, meeting training goals depends on high-quality data to reduce incorrect results. Strict data control is vital in early phases of development. It can progressively relax during successive iterations of design until the model can safely ingest data and provide results as expected. Additionally, these teams prepare the data for use and store it in the right format for the model.
When the second cycle is properly implemented, the model is designed, trained, evaluated and released for use in actual application development, where it must meet rigorous accuracy requirements. Along the way, model engineers track data sources, maintain versioning information and augment the input data with additional data points, which allow the model to make better decisions faster. This process is called feature engineering.
The development (third) cycle is for developing, testing and releasing an application, which incorporates the ML model, into production where it will be used by others. Finally, the operations cycle exists so these ML teams can continuously deploy, upkeep and manage or monitor an application that is already in production.
MLOps lifecycle success stories
Diverse organizations have embraced an MLOps approach for application use cases ranging from financials to marketing.
International payment clearing organization Swift runs an AI Center of Excellence that uses MLOps to speed up and improve development of ML tools to spot fraudulent transactions. They aim for not just accurate models, but also AI fairness, auditability and explainability, while preserving data security and privacy.
Indicia Worldwide, which provides marketing and customer retention assistance, uses ML models to help its customers target and fine-tune their marketing campaigns. They have seen increased speed, improved rigor and better modularization in ML-based application development and deployment since they adopted the MLOps approach.
While ML application development is still hard work, the MLOps lifecycle approach provides a company with a clear roadmap for machine learning success. Also, although knowledge of the MLOps lifecycle is currently limited to tech enthusiasts and AI/ML developers, its use will grow as more startups and other tech companies learn of its benefits.