Tutorial: Neuromorphic spiking neural networks for temporal learning

Abstract

Spiking neural networks (SNN) as time-dependent hypotheses consisting of spiking nodes (neurons) and directed edges (synapses) are believed to offer unique solutions to reward prediction tasks and the related feedback that are classified as reinforcement learning. Generally, temporal difference (TD) learning renders it possible to optimize a model network to predict the delayed reward in an ad hoc manner. Neuromorphic SNNs-networks built using dedicated hardware-particularly leverage such TD learning for not only reward prediction but also temporal sequence prediction in a physical time domain. In this tutorial, such learning in a physical time domain is referred to as temporal learning to distinguish it from conventional TD learning-based methods that generally involve algorithmic (rather than physical) time. This tutorial addresses neuromorphic SNNs for temporal learning from the scratch. It first concerns general characteristics of SNNs including spiking neurons and information coding schemes and then moves on to temporal learning including its general concept, feasible algorithms, and their association with neurophysiological learning rules that have intensively been enriched for the last few decades.Most of neuromorphic systems built using analog building blocks are indeed based on mixed circuit. 1,3,12 The digital neuromorphic prototypes, TrueNorth 2 , SpiNNaker 4 , and Loihi 5 , highlight a fully digital circuit-based strategy in terms of flexibility of network configuration as well as neuron model parameters and learning algorithms. The progress in digital circuit fabrication techniques underpins high-speed and low-power operation of such digital neuromorphic systems. Additionally, a fieldprogrammable gate array (FPGA) is a handy and cost-effective platform for a digital neuromorphic system. To date, various architectures of spiking neurons 13,14 and neuromorphic system architecture for reconfigurable SNN 15 have been built on FPGAs.An emerging strategy is to utilize nonvolatile memory devices, e.g. oxide-based resistive memory, phase-change memory, magnetic tunnel junction, and floating-gate transistor, as synaptic devices. 16,17 Given needs for an enormous number of synapses in a neuromorphic system, replacing even in part mainstream static random access memory or content-addressable memory by such nonvolatile memories can remarkably enhance the areal density of synapses. Additionally, several nonvolatile memories represent multinary states that can further boost the areal density of synapses. 18 Such cutting edge neuromorphic hardware can leverage its capability by the aid of a user-friendly complier, for instance, equipped with graphical user interface (GUI). An example is Nengo 19 , a GUIbased compiler that readily builds an SNN on neuromorphic hardware. Lately, the complier has successfully been applied to Loihi and Braindrop.SNNs largely vary in model complexity. The complexity is generally a measure of fidelity to biological neural network. That is, the more biologically plausible, the more lik...

Emerging Artificial Synaptic Devices for Neuromorphic Computing

“…Like DNNs, there have been several approaches to building SNNs, depending heavily on the desired application as well as the type of synaptic device involved . However, all of these traditional techniques require strict timing mechanisms, as well as extraneous peripheral circuity to accomplish.…”

“…However, all of these traditional techniques require strict timing mechanisms, as well as extraneous peripheral circuity to accomplish. These peripheral circuits not only require large physical chip areas, but also demand massive amounts of memory, making the scalability of these circuits unrealistic . As such, there has been a need for a new type of synaptic device that can inherently incorporate STDP like characteristics for the implementation of the next generation of SNNs.…”

Simplified calcium signaling cascade for synaptic plasticity

“…This may reveal key features to each class, i.e., knowledge discovery. Note that input features encoded using a temporal code or activity code (Jeong, 2018) because the proposed model captures the causality in both domains. However, such direct causality discoveries appear impossible for multilayer SNN, where the input features are connected to the output classes indirectly through hidden neurons whose meanings are unknown.…”

SPSNN: nth Order Sequence-Predicting Spiking Neural Network

“…Spiking neural network (SNN) is a dynamic hypothesis with diverse temporal kernels to express neuronal behaviors in response to synaptic transmission [1][2][3]. The central nervous system (CNS) is based on the SNN, and the SNN has therefore been analyzed theoretically to understand the working principles of the CNS.…”

“…Considering the rich dynamics of SNN, learning with dynamic-domain data perhaps harnesses the full capability of SNNs [3,12]. Dynamic-domain data include time-series data, which embody periodic discrete data points in a time domain.…”