An Efficient Analog Convolutional Neural Network Hardware Accelerator Enabled by a Novel Memoryless Architecture for Insect-Sized Robots
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionInsect-sized robots have gained considerable attention due to their applications such as ambient monitoring. However, shrinking robots' dimensions reduces energy availability. Thus, it has prohibited successful technologies in larger-scale robots from application in insect-sized ones making their autonomy an open challenge. One of these technologies is convolutional neural networks (CNN). This paper presents novelty in different levels of abstraction that lowers the CNN's power. Analog computation is utilized for its compactness, and an architecture is devised to simplify analog circuitry. Proposed convolutional filters consume 1.5 nW/image with 92% accuracy and promise application of CNN-based controllers in insect-sized robots.