Learning automata based energy-efficient AI hardware design for IoT applications

Abstract

The advancement of sensing devices has heralded the onset of the fourth industrial revolution, necessitating substantial progress in artificial intelligence (AI) for Internet of Things (IoT) applications. This technological shift demands AI systems that can process extensive data volumes in real-time, making energy efficiency a critical concern for hardware designers. Traditional neural network-based AI systems, with their complex arithmetic operations, struggle to meet these energy efficiency requirements, especially within the IoT context.

A significant knowledge gap exists in developing efficient AI hardware that maintains high learning accuracy while minimising energy consumption. This research addresses this gap by proposing an innovative AI hardware architecture founded on Tsetlin machines and Tsetlin automata. Tsetlin machines, a form of learning automata, use propositional logic to reinforce actions based on past outcomes, thus simplifying the learning process. Tsetlin automata, the fundamental components of these machines, operate through discrete-step updates to optimise action decisions.

Our results showcase that the custom-designed integrated circuit, which utilises binarised input data processed through parallel logic blocks, achieves remarkable energy efficiency. The architecture demonstrates a 2.67x reduction in machine size without sacrificing inference accuracy. Extensive validation against contemporary machine learning algorithms indicates that this architecture not only matches but often surpasses them in both accuracy and robustness.

The findings suggest that Tsetlin machine-based AI hardware can revolutionise IoT applications by providing an energy-efficient and reliable alternative to conventional neural network-based systems. This advancement paves the way for the development of sustainable AI solutions capable of real-time decision-making, thereby propelling the continuous evolution and sophistication of IoT technologies.

Read more

The full paper Learning automata based energy-efficient AI hardware design for IoT applications is available from The Royal Society.