Core Concepts
DACAPO enables efficient continuous learning in autonomous systems through hardware-algorithm co-design.
Abstract
The content discusses the challenges faced by autonomous systems in deploying deep neural network video analytics and introduces DACAPO as a solution. It highlights the limitations of existing continuous learning systems and proposes a spatially-partitionable and precision-flexible accelerator architecture for optimal resource allocation. The evaluation shows DACAPO outperforming state-of-the-art GPU-based systems with higher accuracy and lower power consumption.
Introduction to Deep Neural Network Video Analytics in Autonomous Systems.
Challenges Faced by Real-World Deployment.
Proposed Solution: DACAPO - Hardware-Algorithm Co-Designed Acceleration.
Evaluation Results Comparing DACAPO with Baseline Systems.
In-Depth Analysis on Resource Allocation Algorithm.
Stats
Our evaluation shows that DACAPO achieves 6.5% and 5.5% higher accuracy than state-of-the-art GPU-based continuous learning systems, Ekya and EOMU, respectively, while consuming 254× less power.