Event-Based, Timescale Invariant Unsupervised Online Deep Learning With STDP

Abstract

Learning of hierarchical features with spiking neurons has mostly been investigated in the database framework of standard deep learning systems. However, the properties of neuromorphic systems could be particularly interesting for learning from continuous sensor data in real-world settings. In this work, we introduce a deep spiking convolutional neural network of integrate-and-fire (IF) neurons which performs unsupervised online deep learning with spike-timing dependent plasticity (STDP) from a stream of asynchronous and continuous event-based data. In contrast to previous approaches to unsupervised deep learning with spikes, where layers were trained successively, we introduce a mechanism to train all layers of the network simultaneously. This allows approximate online inference already during the learning process and makes our architecture suitable for online learning and inference. We show that it is possible to train the network without providing implicit information about the database, such as the number of classes and the duration of stimuli presentation. By designing an STDP learning rule which depends only on relative spike timings, we make our network fully event-driven and able to operate without defining an absolute timescale of its dynamics. Our architecture requires only a small number of generic mechanisms and therefore enforces few constraints on a possible neuromorphic hardware implementation. These characteristics make our network one of the few neuromorphic architecture which could directly learn features and perform inference from an event-based vision sensor.

UAV Detection: A STDP Trained Deep Convolutional Spiking Neural Network Retina-Neuromorphic Approach

“…DVS Trained Network -During initial testing the net-work trained on real DVS event data struggled to converge to useful feature within the second layer, due to the sparse feature maps that were learned in the rst layer. A set of pre-trained weights representing Gabor features, shown in Figure 6, indicative of that seen in other rst layer SNNs [10,18], allowed all the networks to have better building blocks to create more complex features in the second and third layers. Throughout all of the further testing, this method was used, using the four features presented in Figure 6 as the rst layer of each network.…”

“…This allows it to act as a detection layer. In this situation, it is able to forgo usage of a fully connected layer [18] or support vector machine [10] as classi cation isn't required. The network's evaluation will be based upon the number of successful detections and its robustness to a range of highly spiking noisy inputs, replicating low light conditions.…”

Design Space Exploration of Hardware Spiking Neurons for Embedded Artificial Intelligence

“…Indeed, our network is first trained in formal domain, and its weights are then exported to be used in an SNN with the same topology, which is then directly ready for inference. Note that there exist learning methods directly in spike domain, such as SpikeProp or STDP [18][39] [19] [40]. Additional information concerning spiking learning methods is available in [1], which presents a complete survey of Spiking Neural Network training techniques.…”

Toward One-Shot Learning in Neuroscience-Inspired Deep Spiking Neural Networks

“…Deep SNNs (DSNNs) are brain-inspired information processing systems, which have shown interesting capabilities such as fast inference and event-driven information processing which make them excellent approaches for deep neural network architectures [27]. Event-driven means that SNNs generate spikes in response to stimulation from other neurons and show very small firing activity when they receive sparse inputs, such strategy results in power-efficient computing [28]. DSNNs have been developed for supervised learning [29], unsupervised learning [30,28] and reinforcement learning paradigms [31].…”

“…Event-driven means that SNNs generate spikes in response to stimulation from other neurons and show very small firing activity when they receive sparse inputs, such strategy results in power-efficient computing [28]. DSNNs have been developed for supervised learning [29], unsupervised learning [30,28] and reinforcement learning paradigms [31].…”