Is Neuromorphic MNIST Neuromorphic? Analyzing the Discriminative Power of Neuromorphic Datasets in the Time Domain

Iyer, Laxmi; Chua, Yansong; Li, Haizhou

doi:10.3389/fnins.2021.608567

Cited by 26 publications

(11 citation statements)

References 82 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We compare models using FS and FR coding schemes on classification tasks. We avoid using unrealistic datasets, such as N-MNIST (Orchard et al, 2015 ) and CIFAR10-DVS (Li et al, 2017 ), in which the first spike is not meaningful because events are generated by moving neuromorphic device around static images and there are no significant temporal differences in their sequences (Iyer et al, 2021 ). Instead, we test our model on realistic datasets in which important information is encoded in spike timings.…”

Section: Resultsmentioning

confidence: 99%

First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures

Liu,

Leung,

Dragotti

2023

Front. Neurosci.

View full text Add to dashboard Cite

Spiking neural networks (SNNs) are well-suited to process asynchronous event-based data. Most of the existing SNNs use rate-coding schemes that focus on firing rate (FR), and so they generally ignore the spike timing in events. On the contrary, methods based on temporal coding, particularly time-to-first-spike (TTFS) coding, can be accurate and efficient but they are difficult to train. Currently, there is limited research on applying TTFS coding to real events, since traditional TTFS-based methods impose one-spike constraint, which is not realistic for event-based data. In this study, we present a novel decision-making strategy based on first-spike (FS) coding that encodes FS timings of the output neurons to investigate the role of the first-spike timing in classifying real-world event sequences with complex temporal structures. To achieve FS coding, we propose a novel surrogate gradient learning method for discrete spike trains. In the forward pass, output spikes are encoded into discrete times to generate FS times. In the backpropagation, we develop an error assignment method that propagates error from FS times to spikes through a Gaussian window, and then supervised learning for spikes is implemented through a surrogate gradient approach. Additional strategies are introduced to facilitate the training of FS timings, such as adding empty sequences and employing different parameters for different layers. We make a comprehensive comparison between FS and FR coding in the experiments. Our results show that FS coding achieves comparable accuracy to FR coding while leading to superior energy efficiency and distinct neuronal dynamics on data sequences with very rich temporal structures. Additionally, a longer time delay in the first spike leads to higher accuracy, indicating important information is encoded in the timing of the first spike.

show abstract

Section: Resultsmentioning

confidence: 99%

First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures

Liu,

Leung,

Dragotti

2023

Front. Neurosci.

View full text Add to dashboard Cite

show abstract

“…In addition, the learning mechanism is devoid of a teacher signal, i.e., completely unsupervised, using which a spatiotemporal spiking pattern buried in noise can be spontaneously detected. In a recent study [45], the performance of ANNs, rate-coded SNNs, and time-coded SNNs was empirically evaluated using conventional benchmark tasks. Both rate-coded and time-coded SNNs perform well on conventional two-dimensional benchmark tasks, where inputs have information in the spatial dimensions alone (e.g.…”

Section: Discussionmentioning

confidence: 99%

“…Both rate-coded and time-coded SNNs perform well on conventional two-dimensional benchmark tasks, where inputs have information in the spatial dimensions alone (e.g. MNIST image classification [12], [45]). However, in tasks with spatiotemporal inputs, such as gesture classification (DVS gesture [46]), the performance of time-coded SNNs was demonstrated to be superior.…”

Section: Discussionmentioning

confidence: 99%

Adaptive STDP-based On-chip Spike Pattern Detection

Gautam¹

2022

Preprint

View full text Add to dashboard Cite

<p>In this study, we solved a noisy spatiotemporal spike pattern detection task on an analog neuromorphic chip using an unsupervised learning rule. Spike-timing-dependent plasticity (STDP) is the most widespread unsupervised learning rule implemented in Spiking Neural Networks (SNNs) and neuromorphic chips. It has been shown to perform well in conventional benchmark tasks such as spike pattern classification and image classification in SNN simulations. However, a significant performance gap exists between its ideal model simulation and neuromorphic implementation. The learning rate of STDP learning depends on the resolution of synaptic efficacy, high resolution efficacy leads to a small learning rate and stable performance. In computer simulation, synaptic efficacy is configured using 64-bit floating-point precision whereas in low-power neuromorphic chips the resolution is generally restricted to under 5-bit fixed point precision due to silicon area and power constraints. This leads to a degradation in the performance. To solve this problem we proposed a bioinspired learning rule named adaptive STDP learning in a previous study and demonstrated via numerical simulation that the performance of adaptive STDP learning (using 4-bit fixed point synapses) is similar to STDP learning (using 64-bit floating-point precision) in a noisy spatiotemporal spike pattern detection task. In this study, we present the experimental results for the same. The experimental results are similar to those obtained in our simulation-based study. To our best knowledge, this is the first time that an unsupervised, noisy spatiotemporal spike pattern detection task has been demonstrated to perform well on a mixed-signal CMOS neuromorphic chip with low-resolution synaptic efficacy. </p>

show abstract

“…MNIST is a database that has been widely used to evaluate the ability of NNs to undergo supervised learning. [ 227 ] MNIST consists of 28 × 28 pixel images of handwritten digits, 60 000 for training and 10 000 for testing. [ 228 ] To test an ANN that uses multilayer perceptrons, each image of 28 × 28 pixels is converted to 784 inputs and transmitted to the input neurons, and through the hidden layers that are composed of fully‐connected hidden neurons, then output to the output layer that is composed of 10 neurons that represent digits 0–9.…”

Section: Neuromorphic Skin That Uses Artificial Synapses For Memory A...mentioning

confidence: 99%

Neuromorphic Skin Based on Emerging Artificial Synapses

Lee

2022

Adv Materials Technologies

View full text Add to dashboard Cite

biological neural systems such as brains and nerves. [1,2] Although the complementary metal-oxide-semiconductor (CMOS) technology has been remarkably developed and enabled digital revolution during the last decades, the current computing and electronic systems are expected to encounter limitations in the upcoming era of artificial intelligence (AI). [3] Moore's law may be no longer applicable, [4] so further downscaling and integration, and decrease in the energy consumption of processors is becoming difficult; consequently, supercomputers for AI to process big data require huge numbers of processing chips. Graphic processing units (GPUs) have been used as core elements for efficient parallel implementation of vectormatrix multiplication (VMM) in deep learning, [5] but in the classic von Neumann structure, processing units are separated from memory cells, so data must be shuttled through buses; this process constrains speed and reduction of energy usage. This is the "von Neumann bottleneck"; it restricts the efficiencies of time and energy in current digital computing system. [6] Moreover, numerous central processing units (CPUs) and GPUs make supercomputers bulky and heavy.A brain has slow computational speed and low calculation accuracy compared to the digital supercomputer, but efficiently performs comprehensive functions such as learning, recognition, judge, memory, and controlling homeostasis and somatic/ autonomic nervous systems, while consuming little power (≈20 W). [7] Neurons and synapses are the fundamental components of biological nervous system including the central nervous system (CNS; i.e., the brain and spinal cord), and peripheral nervous system (PNS; i.e., sensory and motor nerves). A human brain is composed of ≈10 12 neurons are interconnected by ≈10 15 synapses. The neurons are entangled with high compactness in three-dimensional and massively-parallel networks. [8] The brain implements processing and memory together with extremely low energy consumption of ≈10 fJ per synaptic event. [9] Moreover, the brain use asynchronous event-responsive operation that spends power only with input events, whereas most computing systems that use CPUs with von Neumann architecture operate with a synchronous clock that consumes power periodically, regardless of whether the unit is active. [10] A brain communicates with sensory and motor organs in the body by transmitting neural signals through neurons and synapses of afferent and efferent nerves in the PNS. With sensory input, an afferent nerve transfers neural signals from sensory cells to CNS, then the brain makes decisions, which Neuromorphic skin is an emerging electronic skin that demonstrates sensory, memory, learning, and motor responses in a neuromorphic way by using advanced artificial synaptic devices that emulate the biological nervous system and its neural plasticity. Many types of artificial synapses that emulate brain-inspired computing have been developed and are being integrated in electronic skin to demonstrate artificial sensory and motor ner...

show abstract

Is Neuromorphic MNIST Neuromorphic? Analyzing the Discriminative Power of Neuromorphic Datasets in the Time Domain

Cited by 26 publications

References 82 publications

First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures

First-spike coding promotes accurate and efficient spiking neural networks for discrete events with rich temporal structures

Adaptive STDP-based On-chip Spike Pattern Detection

Neuromorphic Skin Based on Emerging Artificial Synapses

Contact Info

Product

Resources

About