The MLOps ecosystem is maturing to the point where data scientists have the tools they need to ship model-driven products, which would have otherwise required considerable dev ops and frontend engineering efforts. In particular, online machine learning is an area with a growing number of applications, and this talk will dive into building such applications using Flyte, Pandera, and Streamlit.
The MLOps ecosystem is maturing to the point where machine learning engineers and data scientists have the tools that they need to prototype and ship ML-driven products that would have otherwise involved considerable dev ops and frontend engineering efforts.
As a case in point, online learning is an area of machine learning with a growing number of practical applications, such as supply chain management, predictive maintenance, and climate prediction. Online learning is a type of machine learning where the model updates its parameters based on data points it sees only once, and this presents different challenges compared to offline learning because assumptions about the identically and independently distributed nature of the data no longer hold. Another critical aspect of shipping models to production is to ensure the quality of training data, which is important in both offline and online contexts.
This talk will dive into some of the key points to think about when building online machine learning systems in the context of a weather forecasting app and demonstrate how Flyte enables reproducible, robust, and type-safe data pipelines, Pandera integrates with Flyte type system to validate pandas dataframes, and Streamlit eases app development for model-driven user interfaces.