Running hardware-aware neural architecture search on embedded devices under 512MB of RAM
Researchers have developed a new hardware-aware neural architecture search (HW-NAS) technique that enables the creation of small convolutional neural networks (CNNs) suitable for resource-constrained embedded devices. This approach allows for the on-device tailoring of CNN architectures, enhancing privacy by eliminating the need for external servers. The method has demonstrated state-of-the-art performance on benchmarks for tasks like human recognition and tiny computer vision, particularly on ultra-low-power microcontrollers. AI
IMPACT Enables more powerful AI capabilities on low-power, resource-constrained devices, expanding the reach of AI into IoT and wearables.