Chip发表华中科技大学郭新团队最新成果

近日,华中科技大学郭新团队的最新研究成果「Adaptive SRM neuron based on NbOₓ memristive device for neuromorphic computing」发表于全球唯一聚焦芯片类研究的综合性国际期刊Chip第一作者为黄静楠,通讯作者为郭新。


Chip发表华中科技大学郭新团队最新成果


该研究首次基于易失的Pt/NbOₓ/TiN忆阻器实现了漏电整合发放(Leaky integrate-and-fire, LIF)和自适应阈值调节功能的神经元(Spike response model,SRM)。通过控制忆阻器的易失速度,来调节SRM神经元的阈值,以实现神经元的不应期和侧向抑制功能。构造了一个包含SRM神经元的脉冲神经网络(Spiking neural network, SNN),并用无监督学习规则进行训练,成功地实现了带噪声的字母分类,证明了SRM神经元的功能。研究表明,SRM神经元不仅能模拟生物神经元的自适应行为,还丰富了SNN的功能进一步释放其计算能力。


人脑具有存算一体的特性,避免了数据频繁地在存储器和计算单元之间的传输。深度神经网络(Deep neural network, DNN)在编码方式和功能实现上与生物大脑还有巨大的差距,为了获得更高的计算效率和更接近生物大脑功能的计算方式,研究者们提出了下一代的人工神经网络——脉冲神经网络(SNN),其借鉴了生物大脑的结构和工作原理,为神经形态计算的发展提供了一种更有前景的方法。


SNN中所使用的神经元模型来源于计算神经学,具有生物特性。目前,研究人员提出了一些基于忆阻器的人工神经元模型,例如,霍奇金-赫胥黎 (Hodgkin–Huxley, HH) 模型、整合发放(Integrate-and-fire, IF)模型和LIF模型。然而,生物神经元是一个复杂的系统,尽管LIF已被广泛地应用于现有的SNN系统中,但LIF只是其中最基本的功能,仅能构建最简单的SNN网络,其在使用功耗、识别效率等方面还有很大的提升空间。除了LIF功能之外,生物神经元还具有更加高级的功能,例如,对刺激的适应性行为,即生物神经元的活动会随着反复或长时间的刺激而衰退。神经元适应性行为在生物系统中广泛存在,这种适应性行为可以归因于神经元内在的可塑性变化,即神经元的放电阈值具有自适应性。


SRM神经元是一种阈值可调的LIF神经元,其描述了生物神经元对重复或长时间刺激的适应性行为。在研究中,郭新团队开发了一种基于易失性Pt/NbOₓ/TiN忆阻器的SRM神经元。SRM神经元有两个模块:LIF模块和阈值调节模块。LIF模块通过模拟和易失性开关特性来实现,而阈值模块则通过控制器件的易失速度来实现。


SRM神经元的阈值由神经元活动控制。在没有刺激的情况下,阈值保持不变,神经元处于静息状态。当SRM神经元激发时,神经元的阈值增加,以保护其免受强刺激,并为其他神经元提供学习某些特征的公平机会,即SRM神经元在放电后会出现不应期。此外,激发的SRM神经元也会向相邻神经元传递信号以增加其阈值,从而抑制相邻神经元的激发,以实现侧向抑制的功能。


Chip发表华中科技大学郭新团队最新成果

图1 SRM神经元电路图(a),不应期(b)和侧向抑制(c)功能的实现。


该研究团队构建了一个包含SRM神经元和突触的两层全连接SNN,并采用无监督学习规则进行训练。在SNN中,SRM神经元作为输出层以处理突触传递的信号,并产生不同频率的输出信号来识别分类结果。由于SRM神经元的存在,SNN可以实现「赢家通吃规则,并可通过无监督的脉冲频率依赖可塑性(SRDP)学习规则进行训练。利用SRM神经元构建的SNN成功地实现了对带有噪声的四个字母H、U、S、T分类。因此,SRM神经元可以避免信息过载,并支持神经网络的竞争学习。


Chip发表华中科技大学郭新团队最新成果

图2 基于SRM神经元构建的双层全连接的SNN(a),有无SRM神经元的训练结果(b-c)。


Adaptive SRM neuron based on NbOₓ memristive device for neuromorphic computing


The latest research results of our team from Huazhong University of Science and Technology, were recently published in the global emerging comprehensive research journal Chip. The paper¹, titled Adaptive SRM Neuron Based on NbOx Memristive Device for Neuromorphic Computing, was first-authored by Jing-Nan Huang, and co-authored by Tong Wang, He-Ming Huang and myself. An artificial spike resoponse model (SRM) neuron with the leaky integrate-and-fire (LIF) functions and the adaptive threshold was, for the first time, implemented by the volatile memristive device of Pt/NbOₓ/TiN. By modulating the volatile speed of the device, the threshold of the SRM neuron was adjusted to achieve adaptive behaviors, such as the refractory period and the lateral inhibition. To demonstrate the function of the SRM neuron, a spiking neural network (SNN) was constructed with the SRM neurons and trained by the unsupervised spike-rate-dependent plasticity (SRDP) learning rule, which successfully classified letters with noises. This work demonstrated that the SRM neuron not only emulated the adaptive behaviors of a biological neuron, but also enriched the functionality and unleashed the computational power of SNNs.


The human brain exhibits the characteristics of in-memory computing, frequent and massive data shuttling between memories and processing units are not necessary. The human brain outperforms deep neural networks (DNNs) in cognitive tasks in terms of speed and energy efficiency. Spiking neural networks (SNNs)², which are believed to be the next generation of neural networks, emulate the structure and working principles of the human brain, and offer a promising approach to the development of intelligent computing.


The neuron model used in SNN is derived from computational neurology and has biological properties. At present, artificial neurons based on memristive devices, such as the Hodgkin–Huxley (HH) neuron, the integrate-and-fire (IF) neuron and LIF neuron, have been developed to emulate the rich dynamics of biological neurons. The LIF neuron³ is most widely used for the hardware implementation of SNNs, because of its simple circuit and powerful processing ability. In addition to the LIF functions, however, biological neurons have advanced functions which are adapted to stimuli, for example, neuronal activities decay in response to repeated or prolonged stimulation. The neuron adaptation is widely observed in biological systems; such adaptive behaviors can be attributed to changes in the intrinsic plasticity of a neuron whose firing threshold is adaptive.


An SRM neuron is a LIF neuron with adjustable threshold, which describes the adaptive behavior of a biological neuron to repetitive or prolonged stimuli. In this work, Xin Guo's team developed a SRM neuron based on the volatile memristive device of Pt/NbOx/TiN. The SRM neuron has 2 modules: a LIF module and a threshold adjustment module. The LIF module is realized by using analogue and volatile switching property, while the threshold module is realized by controlling the volatile speed of the device.


The threshold of an SRM neuron is controlled by neuronal activities. Without stimulation, the threshold keeps constant and thus the neuron remains in the resting state. When an SRM neuron fires, the threshold of the neuron increases to protect it from strong stimuli, while providing fair chances for the other neurons to learn some features; in other words, an SRM neuron shows a refractory period after firing. Besides, a firing SRM neuron also transmits signals to adjacent neurons to increase their thresholds, therefore, adjacent neurons are directly inhibited and learn nothing from current stimuli; such a phenomenon is called lateral inhibition (Fig. 1).


To demonstrate the function of the SRM neuron in a neural network, our research team constructed a two-layer fully connected SNN with the SRM neurons and synapses (Fig. 2). In the SNN, the SRM neurons are implemented in the output layer to process the synaptic signals, and generate output signals with different frequencies to indicate classification results. Thanks to the SRM neurons, the SNN can realize the "winner-takes-all" rule and be trained by the unsupervised spiking-rate dependent plasticity (SRDP) learning rule. The SNN successfully classifies the letters "H, U, S, T" with noises. Therefore, SRM neurons can avoid information overload and support neural networks for competitive learning.


关于Chip


Chip全球唯一聚焦芯片类研究的综合性国际期刊,已入选由中国科协、教育部、科技部、中科院等单位联合实施的「中国科技期刊卓越行动计划高起点新刊项目」,为科技部鼓励发表「三类高质量论文」期刊之一。


Chip期刊由上海交通大学与Elsevier集团合作出版,并与多家国内外知名学术组织展开合作,为学术会议提供高质量交流平台。


Chip秉承创刊理念: All About Chip,聚焦芯片,兼容并包,旨在发表与芯片相关的各科研领域尖端突破性成果,助力未来芯片科技发展。迄今为止,Chip已在其编委会汇集了来自13个国家的68名世界知名专家学者,其中包括多名中外院士及IEEE、ACM、Optica等知名国际学会终身会士(Fellow)。


Chip第二期将于2022年7月在爱思唯尔Chip官网以金色开放获取形式(Gold Open Access)发布,欢迎访问阅读文章。


爱思唯尔Chip官网:

https://www.journals.elsevier.com/chip


论文链接:

https://www.sciencedirect.com/science/article/pii/S2709472322000132


引用文献

1.Huang, J.-N., Wang, T., Huang, H.-M. & Guo, X. Adaptive SRM neuron based on NbOₓ memristive device for neuromorphic computing. Chip 1, 100015 (2022).

2.Ghosh-Dastidar, S. & Adeli, H. Spiking neural networks. Int. J. Neural Syst. 19, 295-308 (2009).

3.Hunsberger, E. & Eliasmith, C. Spiking deep networks with LIF neurons. Preprint at https://arxiv.org/abs/1510.08829 (2015).

发表评论
留言与评论(共有 0 条评论) “”
   
验证码:

相关文章

推荐文章