Abstract:Spiking neural networks (SNNs) have recently garnered increasing attention for both their potential to “understand” and “utilize” the human brain. Consequently, there is considerable promise in developing an SNN-based neuron array that targets both of these applications. However, such neuron arrays pose challenges that must be overcome [1]. Firstly, the primary applications of spiking neural networks - simulation and inference - require neuronal models with different levels of computational complexity [2]. On the one hand, complex models (e.g., lzhikevich(IZ) model) are needed for simulating large-scale neural networks to better understand brain function and create real-world applications. On the other hand, simple and well-studied models (e.g., Leaky Integrate and Fire (LIF) model) are used to improve the capabilities and energy efficiency of machine learning. However, due to the diverse circuit implementations of different neuron models, existing neural arrays are mainly built on simple LIF model and target only inference applications. Secondly, densely connected networks with non-uniform spike event distributions can cause Network-on-Chip (NoC) congestion, which in turn leads to inevitable delays in worst-case scenarios [3]. In particular, if spike events are processed in a serial manner, the potential delays makes it a challenging problem to optimize the throughput/area trade-off of synapses in large-scale neuron arrays.

A 5.37-TSOPS/W Reconfigurable Neuron Array with Dual-mode Neurons and Asynchronous Synapses for Energy-Efficient Inference and Biological Neural Network Simulation