Symbolic Reservoir Computing within Memristive Crossbar Arrays as a Cellular Automata

Abstract

Symbolism with better explainability and connectionism with better learning ability are the two basic formalisms of artificial intelligence (AI). Though AI is born symbolic, connectionist AI, represented by deep learning, has recently achieved unprecedented success with the support of various hardware accelerators. As connectionist AI has caused a paradigm shift in diverse fields and its applications are being extended to high-stakes domains, its opacity has become a growing concern, giving rise to a rethinking of the need for integrating symbolic logical reasoning with neural networks, i.e., neuro-symbolic computing. However, the very different mathematical nature of logical reasoning and neural networks, i.e., localist and distributed representations, respectively, has impeded their fusion, which has been exacerbated by the lack of suitable hardware. Here, reservoir computing (RC), a typical connectionist model, is experimentally demonstrated to be implemented symbolically within memristor crossbar arrays (MXAs) as a discrete-time 1D rule-based cellular automata (CA). CA-based RC (CARC) within the MXAs achieves up to 86.5% and 100% accuracy in the image recognition task and 20-bit memory task, respectively. Because of hyperdimensionality, CARC requires fewer training samples and is more robust against errors in the emerging MXA hardware. This work expands the functionality of MXAs for AI applications.