• Machine Learning -- A Systems Problem - Markus Weimer


  • Abstract:
    In this talk, I will introduce ML.NET, a machine learning library for .NET developers as well as some recent research projects done on and around it. ML.NET is a machine learning toolbox in the same equivalence class as scikit-learn: Users author training pipelines consisting of a DAG of its 100+ operators. Those training pipelines are then fit to data, which yields a predictive pipeline or model. ML.NET differs from its peers in its ease of deployment (models can run anywhere .NET is available). Also, it has been carefully designed for memory and CPU efficiency. I will then give a brief summary of recent research around ML.NET: (1) Pretzel (OSDI 2018) is a compiler that compacts many ML.NET models for deployment, (2) Distributed Learning with codes (NeurIPS 2018) allows for efficient fault tolerant training of ML.NET models at scale, and (3) Learning (from) graphs allows for DAGs to form the input to machine learning pipelines. I will conclude the talk with open questions and opportunities to collaborate.

     

    Bio:
    I am an architect in Microsoft’s Cloud and AI division. My group develops ML.NET, Microsoft’s machine learning toolkit. I am also a member of the Apache Software Foundation and was the inaugural PMC chair (VP) of Apache REEF. My work focuses on machine learning techniques, systems therefore and applications thereof. Prior, I lead the machine learning research group of the Cloud Information Services Laboratory (CISL) at Microsoft and prior to that, was a researcher at Yahoo! Research.