Power-Efficient Spiking Neural Networks

高能效尖峰神经网络

• 批准号: 576712-2022
• 负责人: Han, JieJ
• 金额: $ 1.46万
• 依托单位: University of Alberta
• 依托单位国家: 加拿大
• 项目类别: Idea to Innovation
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=751835
• 关键词: Power; Efficient; Spiking; Neural; Networks

Inspired by the neurons and synapses of human's brain, we focus on exploring the advantages and disadvantages of spiking neural networks (SNNs). Then scalable and energy-efficient SNNs have been proposed to use time (spikes) of the signal to process information. The spike is essentially a binary event, it is either 0 or 1. The main advantage of SNNs is that it can make the full use of time and space information. Using time as an additional input dimension, SNNs record valuable information in a sparse manner. The neuron in SNNs is in an active state only when receiving or generating a peak signal, which means that it is driven by events, so it can save energy. If there is no spike coming, the neuron will remain idle. In addition, the input value in SNN is 1 or 0, which also reduces the multiplication operation, to a less computation load. There are several ways to encode the input data of SNNS. When looking into ways to reduce the hardware cost of neural networks, stochastic computing (SC) becomes appealing due to the simple logic used for complex computation. Unlike the conventional binary encoding, SC operates on randomly generated binary sequences. Using random numbers to encode input data, the SNNs are called stochastic SNNs. In this project, we focus on the design, evaluation and implementation of stochastic SNNs. We will also explore the market opportunity and work toward the commercialization of the developed technology.

受人脑神经元和突触的启发，我们重点探索尖峰神经网络(SNN)的优势和劣势。然后提出了可扩展的、能量高效的SNN来利用信号的时间(尖峰)来处理信息。脉冲本质上是一个二元事件，它要么是0，要么是1。SNN的主要优势是它可以充分利用时间和空间信息。SNN使用时间作为额外的输入维度，以稀疏的方式记录有价值的信息。SNN中的神经元只有在接收或产生峰值信号时才处于激活状态，这意味着它是由事件驱动的，所以它可以节省能量。如果没有尖峰出现，神经元将保持空闲状态。此外，SNN中的输入值为1或0，这也减少了乘法运算，从而减少了运算量。有多种方法可以对SNN的输入数据进行编码。在寻找降低神经网络硬件成本的方法时，随机计算(SC)由于用于复杂计算的简单逻辑而变得有吸引力。与传统的二进制编码不同，SC对随机生成的二进制序列进行操作。使用随机数对输入数据进行编码，这些SNN称为随机SNN。在本项目中，我们重点研究了随机SNN的设计、评估和实现。我们还将探索市场机会，并努力将已开发的技术商业化。