Automated machine learning (AutoML) comes from a noble goal. By automating the training and deploying machine learning process, we are moving towards democratizing access to machine learning by a broader group of scientists and innovators, including those who do not have years of data science training and a data center at their disposal.
Existing self-proclaimed AutoML tools predominantly focus on automating model development. However, this focus is misguided for two reasons. First, human involvement is crucial for the effective and socially-responsible use of ML models in real-world applications. Second, the major bottleneck in building successful ML applications lies in the process of productionizing models; simply developing the models is insufficient for building end-to-end ML applications. Thus, our focus for automation for ML should be on automating Machine Learning Operations (MLOps), which encompasses the often mechanical and time-consuming engineering tasks involved in taking ML models to production. The unfortunate reality today is that state-of-the-art MLOps tools place more engineering burden on data practitioners and slow development rather than automating mechanical and time-consuming engineering tasks.
In this talk, we will give an overview of the AutoML landscape and present the evidence motivating the need for automated MLOps (AutoMLOps) based on research in AutoML usability. We will then propose user-centric design principles that will overcome the limitations of existing MLOps tools to drive toward true automation. Finally, we will present Linea, a platform for automating critical tasks in the MLOps lifecycle. Linea follows the proposed design principles to automate engineering tasks in the productionization of data science without requiring users to change their existing development workflow. We will present the technical innovations behind the tool and discuss the implications of the Linea approach on the future of MLOps.
Doris Xin is the founder and CEO of Linea, an MLOps startup on a mission to build developer tools that enable organizations to generate value with data rapidly. Doris received a PhD in Computer Science from UC Berkeley. Her thesis focused on designing machine learning systems for developer productivity, research inspired by her experience as a machine learning engineer at LinkedIn. Her career includes engineering and research roles at Databricks, Google, LinkedIn, and Microsoft.
