Metareasoning Approaches for Thermal Management During Image Processing

Abstract

Resource-constrained electronic systems are present in many semi- and fully-autonomous systems and are tasked with computationally heavy tasks such as neural network image processing. Without sufficient cooling, these tasks often increase device temperature up to a predetermined maximum, beyond which the task is slowed by the device firmware to maintain the maximum. This is done to avoid decreased processor lifespan due to thermal fatigue or catastrophic processor failure due to thermal overstress. This paper describes a study that evaluated how well metareasoning can manage a Raspberry Pi 4B’s central processing unit (CPU) temperature while it is performing image processing (object detection and classification) on the Common Objects in Context (COCO) dataset. We developed and tested two metareasoning approaches: the first maintains constant image throughput, and the second maintains constant expected detection accuracy. The first approach switched between the InceptionV2 and MobileNetV2 image classification networks with a Single Shot Multibox Detector (SSD) attached. The second approach was tested on each network for a range of parameter values. The study also considered cases that used the system’s built-in throttling method to control the temperature. Both metarea-soning approaches were able to stabilize the device temperature without relying on throttling.

MLAE2: Metareasoning for Latency-Aware Energy-Efficient Autonomous Nano-Drones

“…Metareasoning as an anytime algorithm improves the agent's decision-making process based on the current situation [14], [15], [16], [17], [18], [19], [20]. As a result, the drone which is the agent can dynamically switch between cloud-based and edge computing implementation while considering the power consumption, latency constraints, and model accuracy metric.…”