尖端的CMOS-NVM混合信号神经形态ConvNet与电路和训练优化的时间分样采集
Anuar Dorzhigulov1, Vishal Saxena1
1AMPIC Lab, Department of Electrical and Electronic Engineering, University of Delaware, Newark, DE, United States.
Frontiers in neuroscience
|August 3, 2023
概括
研究人员开发了一种新的尖端神经网络 (SNN) MaxPool 层,用于神经形态硬件上的节能视觉处理. 这种方法可以实现高精度,降低延迟,克服将深度神经网络 (DNN) 转换为尖端模型的局限性.
科学领域:
- 神经形态工程的神经形态工程
- 计算机视觉 计算机视觉
- 人工智能的人工智能
背景情况:
- 深度神经网络 (DNN) 对于视觉数据处理至关重要,但在常规硬件上部署时耗费大量能量.
- 尖端神经网络 (SNN) 和神经形态硬件提供了节能替代方案,但转换DNN架构,如卷积神经网络 (CNN) 是具有挑战性的.
- 一个关键的障碍是CNN函数缺乏高效的尖端实现,例如MaxPool,这阻碍了神经形态硬件开发.
研究的目的:
- 提出和评估一个新的MaxPool层,用于Spiking CNNs (SCNNs) 的时间复杂化.
- 为了使SCNNs能够在节能,混合信号的神经形态硬件上实现.
- 为了应对将 DNN 功能转换为 spiking 域的挑战,以最小的准确性损失.
主要方法:
- 利用Integrate & Fire神经元的时间动态,为MaxPool在增强领域的决策提供帮助.
- 在SCNN架构中使用aihwkit框架实现和测试拟议的MaxPool.
- 建议在非挥发性内存 (NVM) 交叉条阵列中硬件实现尖端的MaxPool,并使用CMOS外围电路.
主要成果:
- 拟议的尖端MaxPool在CIFAR10上达到61.74%的精度,时空输入为10%,与100%输入 (62.78%相比) 显示最低1%的下降 (二位重量分辨率).
- 该方案允许在延迟和准确性之间进行选择性权衡,使决策能够在完整的时空输入被处理之前做出决定.
- 能源效率估计表明,与现有的神经形态芯片设计相比,其性能具有竞争力.
结论:
- 拟议的时 multiplexing MaxPool 适用于神经形态硬件上的混合信号电路实现.
- 这项工作促进了CNN转换为节能SCNN,推进了神经形态计算.
- 开发的技术为专用硬件上的低功耗,高精度的视觉数据处理提供了一个有前途的方法.
相关概念视频
Neural Circuits
1.3K
Neural circuits and neuronal pools are two of the main structures found in the nervous system. Neural circuits are networks of neurons that work together to carry out a specific task or process. They consist of interconnected neurons and glial cells, which provide structural and metabolic support.
Neuronal pools are collections of nerve cells with similar functions and interact through chemical and electrical signals. These pools include both interneurons (the central neural circuit nodes that...
Neuronal pools are collections of nerve cells with similar functions and interact through chemical and electrical signals. These pools include both interneurons (the central neural circuit nodes that...
1.3K
Downsampling
187
When considering a sampled sequence with zero values between sampling instants, one can replace it by taking every N-th value of the sequence. At these integer multiples of N, the original and sampled sequences coincide. This process, known as decimation, involves extracting every N-th sample from a sequence, thereby creating a more efficient sequence.
The Fourier transform of the decimated sequence reveals a combination of scaled and shifted versions of the original spectrum. This...
The Fourier transform of the decimated sequence reveals a combination of scaled and shifted versions of the original spectrum. This...
187
Upsampling
264
Managing signal sampling rates is essential in digital signal processing to maintain signal integrity. A decimated signal, characterized by a reduced frequency range due to its lower sampling rate, can be upsampled by inserting zeros between each sample. This upsampling process expands the original spectrum and introduces repeated spectral replicas at intervals dictated by the new Nyquist frequency. To refine this zero-inserted sequence, it is passed through a lowpass filter with a cutoff...
264
Sampling Continuous Time Signal
278
In signal processing, a continuous-time signal can be sampled using an impulse-train sampling technique, followed by the zero-order hold method. Impulse-train sampling involves the use of a periodic impulse train, which consists of a series of delta functions spaced at regular intervals determined by the sampling period. When a continuous-time signal is multiplied by this impulse train, it generates impulses with amplitudes corresponding to the signal's values at the sampling points.
In the...
In the...
278
MOS Capacitor
839
A Metal-Oxide-Semiconductor (MOS) capacitor is a fundamental structure used extensively in semiconductor device technology, particularly in the fabrication of integrated circuits and MOSFETs (metal-oxide-semiconductor field-effect transistors). The MOS capacitor consists of three layers: a metal gate, a dielectric oxide, and a semiconductor substrate.
The metal gate is typically made from highly conductive materials such as aluminum or polysilicon. Beneath the metal gate lies a thin layer of...
The metal gate is typically made from highly conductive materials such as aluminum or polysilicon. Beneath the metal gate lies a thin layer of...
839


