Spike-based local synaptic plasticity: a survey of computational models and neuromorphic circuits

Abstract

Understanding how biological neural networks carry out learning using spike-based local plasticity mechanisms can lead to the development of real-time, energy-efficient, and adaptive neuromorphic processing systems. A large number of spike-based learning models have recently been proposed following different approaches. However, it is difficult to assess if these models can be easily implemented in neuromorphic hardware, and to compare their features and ease of implementation. To this end, in this survey, we provide an overview of representative brain-inspired synaptic plasticity models and mixed-signal CMOS neuromorphic circuits within a unified framework. We review historical, experimental, and theoretical approaches to modeling synaptic plasticity, and we identify computational primitives that can support low-latency and low-power hardware implementations of spike-based learning rules. We provide a common definition of a locality principle based on pre- and post-synaptic neural signals, which we propose as an important requirement for physical implementations of synaptic plasticity circuits. Based on this principle, we compare the properties of these models within the same framework, and describe a set of mixed-signal electronic circuits that can be used to implement their computing principles, and to build efficient on-chip and online learning in neuromorphic processing systems.

ETLP: event-based three-factor local plasticity for online learning with neuromorphic hardware

“…In addition to the accuracy benchmarking, we demonstrated a proof-of-concept hardware implementation on FPGA to show how it is possible to map the local computational primitives of ETLP on digital neuromorphic chips. It is also a way to better understand the simplicity of the plasticity mechanism in hardware, where the required information is two Hebbian pre-synaptic (spike trace) and post-synaptic (voltage put in a simple function) factors [29] and a third factor in the form of an external signal provided by the target or label. Therefore, ETLP is compatible with neuromorphic chips such as Loihi2 which supports three-factor local plasticity on the chip [13,45].…”

“…Recent works have shown important gains in computation delay and energy-efficiency when combining event-based sensing (sensor level), asynchronous processing (hardware level), and spike-based computing (algorithmic level) in neuromorphic systems [8,14,41,62]. On the other hand, adaptation with online learning is still ongoing research because neuromorphic hardware constraints prevent the use of exact gradient-based optimization with (BP) and impose the use of local plasticity [29,43,64].…”

“…This constraint is further relaxed by replacing the global loss with several local loss functions [28,39,43] as in the DECOLLE learning rule [28] or by using fixed random learning signals as in the direct random target projection (DRTP) learning rule [19]. Other local learning rules have been proposed for biological plausibility such as the prescribed error sensitivity (PES) learning rule [18,34], or from a close biological inspiration using two-factor plasticity mechanisms which are inherently compatible with neuromorphic circuits [29].…”

A bio-inspired hardware implementation of an analog spike-based hippocampus memory model

“…An important aspect of neural networks is given by the learning rules that govern the process of learning and storing information. While Spike-Timing-Dependent Plasticity (STDP) learning rules have been investigated to a great extent in SNNs [42]- [44], more recent spike-based synaptic plasticity mechanisms that take into account additional factors (such as the neuron's membrane potential or it's recent firing activity) have been shown to be more powerful (see [9] for an overview of rules that are also compatible with neuromorphic hardware).…”

“…In these networks, information is transmitted among neurons in the form of asynchronous pulses (spikes), signals are represented as mean spike rates, calculated either over time (many spikes) or space (many neurons) [8]. Computation is carried out by creating networks with multiple layers and/or recurrent connections, and by implementing learning algorithms based on local synaptic plasticity mechanisms [9]. This approach has great advantages in terms of energy consumption [10], noise robustness [11]- [13] and real-time operation compared to conventional computing systems [14], [15].…”