Supervised learning in spiking neural networks with MLC PCM synapses

Abstract

Spiking neural networks (SNN) are artificial computational models that have been inspired by the brain's ability to naturally encode and process information in the time domain. The added temporal dimension is believed to render them more computationally efficient than the conventional artificial neural networks, though their full computational capabilities are yet to be explored. Recently, computational memory architectures based on non-volatile memory crossbar arrays have shown great promise to implement parallel computations in artificial and spiking neural networks. In this work, we experimentally demonstrate for the first time, the feasibility to realize high-performance event-driven in-situ supervised learning systems using nanoscale and stochastic phase-change synapses. Our SNN is trained to recognize audio signals of alphabets encoded using spikes in the time domain and to generate spike trains at precise time instances to represent the pixel intensities of their corresponding images. Moreover, with a statistical model capturing the experimental behavior of the devices, we investigate architectural and systems-level solutions for improving the training and inference performance of our computational memory-based system. Combining the computational potential of supervised SNNs with the parallel compute power of computational memory, the work paves the way for next-generation of efficient brain-inspired systems.In recent years, deep learning algorithms have become successful in solving complex cognitive tasks surpassing the performance achievable by traditional algorithmic approaches, and in some cases, even expert humans. However, conventional computing architectures are confronted with several challenges while implementing the multi-layered artificial neural networks (ANNs) used in these algorithms, especially when compared against the approximately 20 W power budget of the human brain. The inefficiencies in the von Neumann architecture for neural network implementation arise from the high-precision digital representation of the network parameters, constant shuttling of large amounts of data between processor and memory, and the ensuing limited computational parallelism and scalability. In contrast, the human brain employs billions of neurons that communicate with each other in a parallel fashion, through dedicated, analog, and low-precision synaptic connections. The spike-based data encoding schemes used in these biological networks render the computation and communication asynchronous, event-driven, and sparse, contributing to the high computational efficiency of the brain.The size and complexity of artificial neural networks are expected to continue to grow in the future and has motivated the search for efficient and scalable hadware implementation schemes for learning systems 1 . Spiking neural networks (SNNs) are excellent candidates to implement large learning networks, especially for energy and memory-constrained embedded applications, as they closely mimic some of the key computational principles of...

Federated Neuromorphic Learning of Spiking Neural Networks for Low-Power Edge Intelligence

“…which respectively correspond to the learning signal and eligibility trace, i.e., the running average of the gradients of the log-loss. The global update at the BS is then given by (4). As summarized in Algorithm 1, FL-SNN is based on local and global feedback signals, rather than backpropagation.…”

“…As in [5], the local learning signal ℓ (i) (t), computed every ∆s, indicates to the hidden neurons within the SNN of each device i how effective their current signaling is in maximizing the probability of the desired input-output behavior defined by the selected data (x (i) , y (i) ). In contrast, the global feedback signal θ(t) is given by the global averaged parameter (4), which aims at enabling cooperative training via FL.…”

“…Training state-of-the-art Artificial Neural Network (ANN) models requires distributed computing on large mixed CPU-GPU clusters, typically over many days or weeks, at the expense of massive memory, time, and energy resources [1], and potentially of privacy violations. Alternative solutions for low-power machine learning on resource-constrained devices have been recently the focus of intense research, and this paper proposes and explores the convergence of two such lines of inquiry, namely Spiking Neural Networks (SNNs) for low-power on-device learning [2], [3], [4], [5], and Federated Learning (FL) [6], [7] for collaborative inter-device training.…”

Computing of temporal information in spiking neural networks with ReRAM synapses

“…45 Though a complete training algorithm of a multilayer spatiotemporal network is still missing, training the network to map a spatiotemporal input into a spatiotemporal output (Fig. 8a) is a critical step.…”

Neuromorphic computing with resistive switching memory devices

“…To combine the fast training in crosspoint architectures with the flexibility of GPUs, recently proposed analog/digital hybrid systems implement the crossbar array as an analog computational unit able to largely accelerate the calculation of the MVM[65][66][67][68], thus allowing to relax the computational load on the digital section[68]. For instance, a dot-product engine was proposed where the resistive crossbar array efficiently calculates MVM via vector scalar product a·b = Σai bi, achieving in simulation software-like accuracy on MNIST digit recognition by implementing pre-trained weights and efficiently performing forward inference[65,66].…”