The difficulty of transitioning from research to production is a prevalent issue in the machine learning development life cycle. An ML team may need to modularize and rework their code to work more effectively in production. Occasionally, depending on whether the application requires offline, online, or streaming predictions, this can necessitate re-implementing and maintaining feature engineering or model prediction logic in several locations.

The audience will learn about an open-source microframework for creating machine learning applications in this session. UnionML, developed by the Flyte team, offers a straightforward, user-friendly interface for specifying the fundamental components of your machine learning application, from dataset curation and sampling to model training and prediction. UnionML automatically generates the procedures required to fine-tune your models and release them to production in various prediction use cases, such as offline, online, or streaming settings using these building blocks. There will be a live demonstration by taking an end-to-end machine learning-based example written in Python.

We can look to the web for ideas while we consider a solution to this issue. For instance, the HTTP protocol, which provides a backbone of techniques with clearly defined but flexible interfaces, standardizes the way we move data across the internet. We were interested in posing the question, "What if we could develop, automate, and monitor data and ML pipelines at scale?" as machine learning systems proliferate across industries. https://github.com/unionai-oss/unionml