MyML: User-Driven Machine Learning
TimeWednesday, December 8th2:10pm - 2:30pm PST
Event Type
Research Manuscript
Virtual Programs
Presented In-Person
Embedded System Design Methodologies
Embedded Systems
DescriptionMachine learning (ML) on edge devices is expensive and requires offloading computation to the cloud, which compromises the user privacy. However, the data processed at edge devices is user specific and limited to few classes. In MyML, we create small, user-specific ML models, rather than utilizing a generic, compute-intensive one. To build user-models, we present a hardware-friendly pruning, which leverages compute sharing between pruning and inference, customizes the backward-pass and the choice of pruning granularity for efficient processing. We then propose architectural support to prune user-models on systolic edge ML accelerators. We demonstrate that user-models provide a 2.3× speedup.