MLOps for Deep Learning: Drift Detection and Efficient Retraining in Model Serving

Explore cutting-edge MLOps techniques for deep learning in this 48-minute conference talk from MLOps World: Machine Learning in Production. Delve into the challenges of model serving, including when to retrain models and how to do so efficiently. Learn about an ensemble drift detection technique that captures data and concept drifts, considering real-world scenarios where ground truth labels are received with a time lag. Discover a framework that automatically determines retraining data based on drift signals and triggers warnings for potential drift. Examine solutions to life-long retraining challenges, including catastrophic forgetting and efficient retraining, through a novel approach using multi-armed bandits and a new regularization term focusing on synapse and neuron importance. Gain insights into unlocking the true potential of AI in production environments by understanding the importance of a continuous deployment pipeline. Explore an open-source project that integrates unique drift detection and model retrain algorithms for serving deep learning models. Learn how to efficiently deploy, monitor, and maintain deep learning models in production using a Kubernetes native POC solution.