Brain neurons exhibit far more sophisticated and powerful information-processing capabilities than the simple integrators commonly modeled in neuromorphic computing. A biological neuron can in fact efficiently perform Boolean algebra, including linear nonseparable operations. Traditional logic circuits require more than a dozen transistors combined as NOT, AND, and OR gates to implement XOR. Lacking biological competency, artificial neural networks require multilayered solutions to exercise XOR operation. Here, it is shown that a single-transistor neuron, harnessing the intrinsic ambipolarity of graphene and ionic filamentary dynamics, can enable in situ reconfigurable multiple Boolean operations from linear separable to linear nonseparable in an ultra-compact design. By leveraging the spatiotemporal integration of inputs, bio-realistic spiking-dependent Boolean computation is fully realized, rivaling the efficiency of a human brain. Furthermore, a soft-XOR-based neural network via algorithm-hardware co-design, showcasing substantial performance improvement, is demonstrated. These results demonstrate how the artificial neuron, in the ultra-compact form of a single transistor, may function as a powerful platform for Boolean operations. These findings are anticipated to be a starting point for implementing more sophisticated computations at the individual transistor neuron level, leading to super-scalable neural networks for resource-efficient brain-inspired information processing.
Keywords: artificial intelligence; boolean algebra; neural network; neuromorphic computing; neuron model; reconfigurable logic.
© 2024 Wiley‐VCH GmbH.