OverNeuron: a Spiking Recurrent Cell as Neuromorphic LSTM by Higher-Order Connectivity

With the convergence of neuroscience and Artificial Intelligence (AI) into the highly interdisciplinary field of Neuro-AI, a significant influence of Deep Learning (DL) can be observed in the domain of neuromorphic computing. As a result, established DL architectures have been adapted to more bio-inspired counterparts in the attempt to obtain the advantages offered by event-based computation. In this work, we adopt a deeper approach than just converting activation functions to binarized alternatives, and we look at neuroscience to draw inspiration for the neuromorphic redesign of the Long Short-Term Memory (LSTM) cell. Specifically, we employ the idea of higher-order connectivity in the brain to build a spiking recurrent cell that implements the same structure as the LSTM but only relies on populations of Leaky Integrate-and-Fire (LIF) neurons. For benchmarking purposes, we use the Spiking Heidelberg Digits (SHD) and the event-based Braille letter datasets, and we report on effective execution of our architecture, named OverNeuron, on Intel Loihi 2. With a best on-hardware test accuracy of 88.75% for SHD and 85.93% for Braille, we show that the OverNeuron can outperform the LSTM cell on a temporal and a spatio-temporal event-based task.