Hardware–Software Co-design for Ultra-Resource-Constrained Embedded Machine Learning Inference: A Printed Electronics Use Case

In this chapter, we discuss hardware–software co-design approaches that enable machine learning inference on ultra-resource-constrained embedded systems. As our use case, we consider the printed electronics. The latter form an extreme example of embedded machine learning application. Printed electronics constitute a promising solution to bring computing and smart services in application domains that require sub-cent cost and conformality and have not seen yet significant penetration of computing. Printed electronics form a rapidly growing market but also feature several prevalent limitations. Integration density and performance of printed electronics are of order of magnitude lower than those in silicon VLSI systems, and implementing complex circuits, e.g., ML classifiers, poses a great challenge. Considering the a priori requirement of embedded ML for acceptable accuracy and low latency within the limitations of the ultra-resource-constrained printed devices, custom-designed circuits as well as software–hardware co-design and optimization combined with non-conventional computing approaches (e.g., approximate and stochastic computing) are becoming mandatory for the realization of such circuits.