Schedule: Model lifecycle management sessions

Boris Lublinsky and Dean Wampler walk you through using ML in streaming data pipeline and doing periodic model retraining and low-latency scoring in live streams. You'll explore using Kafka as a data backplane, the pros and cons of microservices versus systems like Spark and Flink, tips for TensorFlow and SparkML, performance considerations, model metadata tracking, and other techniques. Read more.

Abhishek Kumar and Pramod Singh walk you through deep learning-based recommender and personalization systems they've built for clients. Join in to learn how to use TensorFlow Serving and MLflow for end-to-end productionalization, including model serving, Dockerization, reproducibility, and experimentation, and Kubernetes for deployment and orchestration of ML-based microarchitectures. Read more.

Holden Karau, Francesca Lazzeri, and Trevor Grant offer an overview of Kubeflow and walk you through using it to train and serve models across different cloud environments (and on-premises). You'll use a script to do the initial setup work, so you can jump (almost) straight into training a model on one cloud and then look at how to set up serving in another cluster/cloud. Read more.

You've invested heavily in cleaning your data, feature engineering, training, and tuning your model—but now you have to deploy your model into production, and you discover it's a huge challenge. Diego Oppenheimer shares common architectural patterns and best practices of the most advanced organizations who are deploying your model for scalability and accessibility. Read more.

There are many great tutorials for training your deep learning models, but training is only a small part in the overall deep learning pipeline. Tobias Knaup and Joerg Schad offer an introduction to building a complete automated deep learning pipeline, starting with exploratory analysis, overtraining, model storage, model serving, and monitoring. Read more.

Evaluating machine learning models is surprisingly hard, particularly because these systems interact in very subtle ways. Ted Dunning breaks the problem of evaluation apart into operational and function evaluation, demonstrating how to do each without unnecessary pain and suffering. Along the way, he shares exciting visualization techniques that will help make differences strikingly apparent. Read more.

Production ML applications benefit from reproducible, automated retraining, and deployment of ever-more predictive models trained on ever-increasing amounts of data. Kelley Rivoire explains how Stripe built a flexible API for training machine learning models that's used to train thousands of models per week on Kubernetes, supporting automated deployment of new models with improved performance. Read more.

Developing applications that leverage machine learning is difficult. Practitioners need to be able to reproduce their model development pipelines, as well as deploy models and monitor their health in production. Corey Zumar offers an overview of MLflow, which simplies this process by managing, reproducing, and operationalizing machine learning through a suite of model tracking and deployment APIs. Read more.

KubeFlow separates compute and storage to provide the ability to deploy best-of-breed open source systems for machine learning to any cluster running Kubernetes, whether on-premises or in the cloud. Skyler Thomas and Terry He explore the problems of state and storage and explain how distributed persistent storage can logically extend the compute flexibility provided by KubeFlow. Read more.

Ting-Fang Yen details an approach for monitoring production machine learning systems that handle billions of requests daily by discovering detection anomalies, such as spurious false positives, as well as gradual concept drifts when the model no longer captures the target concept. Join in to explore new tools for detecting undesirable model behaviors early in large-scale online ML systems. Read more.

Kevin Moore walks you through how TransmogrifAI—Salesforce's open source AutoML library built on Spark—automatically generates models that are automatically customized to a company's dataset and use case and provides insights into why the model is making the predictions it does. Read more.

Machine learning and data science systems often fail in production in unexpected ways. David Talby shares real-world case studies showing why this happens and explains what you can do about it, covering best practices and lessons learned from a decade of experience building and operating such systems at Fortune 500 companies across several industries. Read more.

Harish Doddi and Jerry Xu share the challenges they faced scaling machine learning models and detail the solutions they're building to conquer them. Read more.