November 2021 Summaries

Recently, MLOps World took place in Austin, featuring panels and keynotes on the state of production LLMs. Data versioning tools are essential for data science teams to keep track of all their data, experiments, and model lineage. Top data versioning tools include DAGsHub, DVC, Pachyderm, and MLflow. These tools help reduce errors, improve workflows, and enable easy reproducibility of machine learning models. MLOps is a growing field with more tools designed to simplify, automate, and scale model building and training.

Recently, Zillow experienced significant losses due to their machine learning models' failure to accurately predict housing prices. This incident highlights the importance of monitoring and maintaining ML models in production. Businesses should be aware that a model performing well during training may not perform as expected when deployed, due to constant changes in the environment and data. Ongoing monitoring is essential to detect drift, bias, and integrity issues early on and ensure models are working as intended. Despite many companies utilizing AI in their business, comprehensive monitoring solutions or dedicated resources for building ML monitoring solutions remain limited.

The MLOps space has been growing rapidly with more tools designed to simplify and automate model building and training processes. Training Orchestration is one such tool that enables data science teams to run highly concurrent, scalable, and maintainable training workflows in the cloud. Some popular MLOps tools include Kubeflow, Katonic, MLflow, Spock, Stoke, Valohai, and Spell. These platforms provide various features like automating complex ML infrastructure, operational work for training and deploying AI models, and model monitoring to ensure their accuracy and reliability. Building an ML infrastructure is possible with the right tools that cater to specific needs.