Over eighty percent or more of companies that attempt to integrate machine learning into operational applications fail. How could this be? Many organizations underestimate the difficulty of implementing ML. This talk emphasizes the significance of machine learning operations (MLOps) in scaling TinyML to enterprise-scale deployments that provide real-world value. Training and deploying a machine learning model on a single tiny embedded device is one thing; it is quite another to scale to thousands of devices. TinyML adds a number of embedded ecosystem-specific impediments to the conventional machine learning deployment pipeline, hence considerably complicating ML deployment even further. To address these myriad issues, the talk introduces a seven-stage MLOps architecture for operationalizing TinyML successfully. These stages range from ML model development for a fleet of heterogeneous devices to continuous monitoring for detecting data drift and everything in-between. The framework is a comprehensive end-to-end workflow for scaling TinyML deployments from a proof of concept to a real-world solution.
Schedule
Timezone: PDT
MLOps for TinyML: Challenges & Directions in Operationalizing TinyML at Scale
Vijay JANAPA REDDI, Associate Professor
Harvard University
Vijay JANAPA REDDI, Associate Professor
Harvard University
Vijay Janapa Reddi is an Associate Professor at Harvard University, Inference Co-chair for MLPerf, and a founding member of MLCommons, a nonprofit ML organization that aims to accelerate ML innovation. He also serves on the MLCommons board of directors.
Schedule subject to change without notice.