Close

Presentation

ARGOS: an Adaptive and ReGion-scale knowledge distillation for Object recognition Systems
TimeTuesday, December 7th6:00pm - 7:00pm PST
Event Type
Networking Reception
Work-in-Progress Poster
Virtual Programs
Presented In-Person
DescriptionObject recognition research communities have achieved great success using neural networks. Reducing these networks' computational cost remains a challenge, which impedes their deployment on embedded devices. Recent studies employ a keyframe selection and temporal knowledge distillation for detecting objects in related frames to reduce these networks' computational cost. Still, keyframe selection models' processing time substantially increases the system's response time. Consequently, we present an online knowledge distillation framework (ARGOS) that improves the performance and the response time by decomposing input frame into a few sub-processes executed in parallel. Extensive experiments demonstrate its effectiveness with an average 89% reduction in energy.