DaCapo: Accelerating Continuous Learning in Autonomous Systems for Video Analytics
Kernekoncepter
DACAPO proposes a hardware-algorithm co-designed solution for continuous learning, enabling autonomous systems to achieve higher accuracy and energy efficiency.
Resumé
Deep neural network (DNN) video analytics is crucial for autonomous systems like self-driving vehicles and security robots.
Challenges faced include limited computational resources and power constraints.
Continuous learning involves inference, labeling, and retraining tasks to adapt models to changing data distributions.
Existing continuous learning systems overlook computation needs for labeling and inference, relying on power-hungry GPUs.
DACAPO offers spatially-partitionable accelerator architecture and spatiotemporal resource allocation algorithm for optimal performance.
Evaluation shows DACAPO achieves higher accuracy with lower power consumption compared to GPU-based systems.