Autonomously Adaptive Experimentation-Driven Pipeline

What is it?

MLOps pipeline takes care of the ML model, including various tasks such as model training, deployment, and serving. Continuous training (CT) and continuous deployment (CD) are needed for ML models.  CT enables automatic model retraining, and CD automatically deploys retrained models to production. They enable ML systems to respond to changes in production by keeping models fresh. Retraining can be triggered periodically or by model monitoring results or repository updates. A commonly used strategy during CD in today's software engineering is A/B testing, meaning experimenting with a redeployed model on a small percentage of user traffic. A/B testing can validate the performance of a retrained model in production and mitigate the risk of deploying a poorly performing model, further elevating the effectiveness of model CT and CD.

Why is it necessary?

Especially when changes are frequent, uncertainty is high or many models are being served, CT and CD may need to operate autonomously to adapt the ML model requiring advanced tools on top of the MLOps pipeline to handle automation. However, simply CT and CD are not enough, the resulting retrained ML model needs to be validated so that it, at least, outperforms the existing model requiring additional infrastructure to handle validations.

How does it work?

CTCD-e (continuous-training-and-continuous-deployment-enabling) pipeline works on top of the IML4E OSS pipeline so that it can autonomously adapt ML systems to changing data by providing flexible CT and CD support for models. It can automatically start to retrain a model when its performance degrades, and automatically A/B test the retrained model against its predecessor in production.