AutoPNN (Automated Printed Neural Network via Software-Hardware Approximation and Codesign for Machine Learning Classification in Printed Electronics) is a 2-year research project targeting the development of a fully automated framework for enabling printed self-powered classification circuits.
Printed electronics is increasingly recognized as a key enabler for the Internet of Things (IoT) as part of the “Fourth Industrial Revolution”. Ultimately, printed electronics form the key technology to outreach the limited and physically bounded potential of silicon-based IoT systems and realize a “true” pervasive computing to its full potential, thus fulfilling its deepest promise. Printed electronics is expected to grow over the next years into a key technology with an enormous economic and social impact. Quoting the European Commission report on Printing Electronics : for European businesses, flexible and printed electronics offer opportunities in high value-added products, with most potential in several specific application areas. These application areas are healthcare and medical devices, smart packaging and logistics, sensors for IoT, industry and environmental monitoring, and automotive. The fact that in such applications the core task is mainly classification highlights the utmost impact of our project, i.e., automated generation of high-accuracy, battery-powered printed ML classifiers. Moreover, considering that partnerships and collaboration platforms play a critical role to enable European players to stay competitive and develop the market for flexible electronics in Europe , our objective to open-source AutoPNN is underlined. Finally, we must note that startups play a central role in the flexible and printed electronics ecosystem due to the fact that innovation in this field is very risky and it is usually not targeted by larger companies. The low equipment cost of printed electronics and open-sourcing AutoPNN will enable start-ups to do business on printed ML classification.
In order to achieve the objectives of AutoPNN, the work plan is divided into the following Work Packages (WP):
WP1 and WP6 regard the management and dissemination of the project. In WP2 and WP3 the approximation libraries, dedicated to printed circuits, of AutoPNN are generated. WP2 targets hardware-driven software approximations, while WP3 targets approximate bespoke circuits. In WP4, the required models (area and accuracy) are developed to enable fast, high-level operation of AutoPNN. WP2-WP4 form AutoPNN's offline stage. Finally, WP5 is AutoPNN's online process in which, given an input dataset and user-defined constraints, an optimization search is executed in order to generate the corresponding approximate PNN design.
