Change for the Better: Improving Predictions by Automating Drift Detection in Machine Learning Models

Learn how to improve machine learning predictions through automated drift detection in this 17-minute conference talk from tinyML EMEA. Explore a comprehensive solution for handling data drift challenges in machine learning systems, demonstrated through a real-world predictive maintenance example. Discover how to detect model performance degradation, generate accurate models from new data, and seamlessly deploy updates into existing ML pipelines. Master the implementation of data drift detection algorithms that evaluate observation variability and prediction accuracy, while leveraging physics-based simulation models for precise data labeling. Gain insights into automating both drift detection and data labeling processes to reduce operational costs and expertise requirements. Follow along with a practical example involving electric vehicle battery management, where streaming data, Kafka integration, and thermodynamic modeling combine to create an efficient ML pipeline. Learn to implement a production system that monitors deployed models, detects drift, and automatically triggers new model training when needed, all while maintaining continuous integration and deployment standards.