When a problem occurs while operating an AI system, it is usually necessary to determine if it is a system-side problem or an AI model problem. This takes a lot of time and effort. Even for AI models which have undergone thorough technological verification by the AI development team, production environments may produce unexpected data. When this happens, the AI model itself needs to be reviewed and revised.
What's more, no matter how high the quality of the AI model that has been developed and deployed, maintaining that quality over the long term is difficult. External factors can, with time, negatively affect the accuracy of AI models. For example, if factors such as air temperatures and equipment aging affect the prediction accuracy of AI models, then AI model relearning will be necessary to reflect these external environmental changes and maintain the accuracy of the model's predictions. In other words, when using AI for business, it is vital to implement an ongoing process of reviewing and revising AI models as necessary.
The system we have developed to continuously implement this process is the MLOps Platform. MLOps stands for "Machine Learning Operations." It refers to all of the processes involved in the management of AI model lifecycles, from development to deployment and operation.
* Deployment: The placement and provision of software (such as AI models) in actual operating environments
The MLOps workflow consists of the following eight processes. Continuously implementing these processes enables AI to evolve in response to changes in the external environment, providing value on an ongoing basis.
- Continuous Integration (CI)：Testing and implementation preparation for 2 and 3
- Learning data acquisition and pre・processing
- AI model learning
- AI model configuration management
- Continuous Delivery (CD)：Creating microservices to enable application programming interfaces (API) to call AI models and deploying these microservices in operating environments
- AI model operation：Operating the AI model microservices
- Usage history/prediction accuracy：Collecting AI model prediction results, etc.
- Monitoring：Remotely monitoring AI models
We developed a platform that supports these MLOps and created a workflow that connects all of the processes involved in AI model creation, system development, and system operation. This partially automates the AI model lifecycle, from deployment to operation. The MLOps Platform makes it possible for AI development teams, system development teams, and system operation teams to cooperate smoothly (Fig. 1).