An Efficient Analog Convolutional Neural Network Hardware Accelerator Enabled by a Novel Memoryless Architecture for Insect-Sized Robots

Abstract

For decades, miniaturization of robots has gained considerable attention due to the exciting applications of insect-sized robots, such as ambient monitoring. However, scaling down the robots’ dimensions reduces energy availability drastically for sensors and controllers. It has prohibited many successful technologies tested in larger-scale robots from application in insect-sized ones. As a result, insect-sized robots’ power and sensor/control autonomy is an open field of research. One of these technologies is Convolutional Neural Networks (CNN). This paper presents novelty in different levels of abstraction from architectural to transistor-level that drastically reduces the CNN power to comply with the low power budget of insect-sized robots. Analog computation is utilized for its compactness, and an architecture is devised to simplify the analog circuitry. Proposed convolutional filters, showing four orders of magnitude higher efficiency with respect to the state-of-the-art, consume merely 1.5 nW/image with 92% accuracy and promise application of CNN-based controllers in insect-sized robots.