General and stable emulation of finite state machines with spiking neural networks

Abstract Finite state machines (FSMs) are fundamental models widely used in a variety of domains for state control. However, they face challenges in modeling complex especially black-box systems without explicit state descriptions. Neural networks, conversely, excel at modeling implicit and continuous systems but struggling with the temporally stable and precise tasks which FSMs can handle effectively. This work explores the emulation of FSMs with neural networks to harness the strengths of both paradigms. Inspired by the similarity between the spike-based information representation in bio-inspired spiking neural networks (SNNs) and the discrete state transition in FSMs, we propose discrete-time spiking recurrent neural networks (DTSRNNs) to emulate FSMs. We further incorporate one-hot encoding to enhance the discriminability of state vectors, which is beneficial for learning complex behaviors. Then, we build a random-FSM dataset to evaluate model performance. Extensive experiments reveal that, our DTSRNNs surpass conventional discrete-time recurrent neural networks (DTRNNs) with extended decline periods, indicating superior temporal stability. They also exhibit higher robustness against different types of noise. Our work not only presents a significant advancement in the stable emulation of general FSMs with SNNs, but also provides a promising approach for modeling complex temporal systems with especially long sequences.