Session
Poster Session 3
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
Deploying deep learning models onboard spacecraft is limited by strict size, weight, power (SWaP), and bandwidth constraints. Standard object detection networks like YOLO or Faster R-CNN are too computationally heavy for real-time use in space. We present a pruning framework using PEEK (Probabilistic Explanations for Entropic Knowledge extraction), an entropy-based interpretability method that pinpoints low-information layers for removal. Applied to YOLOv5 for Rendezvous, Proximity Operations, and Docking (RPOD), this approach cuts model size by up to 30% and compute cost by 25%, with only minor accuracy trade-offs—enabling faster, reliable detection on low-SWaP spacecraft platforms.
Document Type
Event
PEEK-Guided Neural Network Pruning for Deployment on Low SWaP Hardware
Salt Palace Convention Center, Salt Lake City, UT
Deploying deep learning models onboard spacecraft is limited by strict size, weight, power (SWaP), and bandwidth constraints. Standard object detection networks like YOLO or Faster R-CNN are too computationally heavy for real-time use in space. We present a pruning framework using PEEK (Probabilistic Explanations for Entropic Knowledge extraction), an entropy-based interpretability method that pinpoints low-information layers for removal. Applied to YOLOv5 for Rendezvous, Proximity Operations, and Docking (RPOD), this approach cuts model size by up to 30% and compute cost by 25%, with only minor accuracy trade-offs—enabling faster, reliable detection on low-SWaP spacecraft platforms.
