《类脑计算ppt课件.ppt》由会员分享，可在线阅读，更多相关《类脑计算ppt课件.ppt（27页珍藏版）》请在三一办公上搜索。
1、类脑计算,CONTENTS,3,类脑理念与实践,General AI 认知 Narrow AI 感知传统思想：认知科学了解大脑意识的形成模仿人脑设计算法( Empirical )当前思想：神经生物学了解大脑结构结构仿真形成一定规模的神经元网络再思考可以做什么( Heuristic ),类脑理念与实践,感知:PerceptronLeNet AlexNetVGGGoogLeNetResNet认知：BrainScaleS Neurons 4 million Synapses 1 billion Power 10 kWTrueNorth Neurons 16 million Synapses 4 bi
2、llion Power 1 WSpiNNaker Neurons 460 million Synapses 460 billion Power 50 kWHuman Brain Neurons 100 billion Synapses 100 Tera Power 20 W,Deeper than Deeper,Bigger than Bigger,5,上界：认知科学 功能模拟类脑计算下界：神经科学 解析仿真上确界：神经科学 结构模仿类脑实践下确界：传统计算机体系结构,类脑理念与实践,6,神经形态认知计算,神经元状态方程：Hodgkin-Huxley方程,7,神经形态认知计算,CMOS模拟：忆
3、阻器：忆阻器是天然的突触模拟器件，生物突触释放Ca、Na离子改变传导性，忆阻器通过释放氧离子改变传导性。1个忆阻器可以模拟一个突触的行为。当前集成水平可以达到10GB/cm2存储密度（近似人脑神经元密度），主频200MHz（比人脑脑电频率快106倍），未来可达到500G/cm21000MHz。,8,Spiking Neurons and Spiking Neural Network Functions and Consumption,神经形态认知计算,9,ANNAssumed that neurons represent information through their mean rates
4、 of action potential firing. NonLinear(Sum() SNN How Spikes convey information ? Rate : spike count within a time window Temporal : precise time = time array,神经形态认知计算,10,FeatureMap Spatiotemporal Patterns Classification Solved. How to Code ? Unsolved.,神经形态认知计算,11,Implementing Instances,神经形态认知计算,12,I
5、mplementing Instances Convolution Max Pooling STDP (Unsupervised Learning in Feature Extraction) Supervised Learning rules to fine tuning Spiking Network,神经形态认知计算,13,STDP The process adjusts the connection strengths based on the relative timing of a particular neurons output and input action potenti
6、als (or spikes).,神经形态认知计算,14,Implementing Instances,神经形态认知计算,15,Neuron Software 分子层面精细建模 最大程度模拟神经元与突触行为 对视网膜中央区域 2mm 范围内所有神经元进行建模 天河2号超算辅助运算,类脑理念与实践,16,Spiking Array 时空关系上的连续性 超高速视屏 20,000 Frame/s Line Motion Illusion 视网膜与V1层功能区的协作产生线运动错觉 V1区含有方向选择细胞 V1区脑皮层之间的互联形成对位置方向的感知 V1区脑皮层在无刺激输入时也会产生自发性放电现象,类脑
7、理念与实践,17,Neural Encoding of Motion Direction Single NeuronNeural Population,计算神经科学,CANN Continuous Attractor Neural Network,18,Dynamical and Adaptive Information ProcessingFrom Rate to Correlation CodeSFASpike frequency AdaptationNeuronal response attenuates after experiencing long time firing.Slow
8、negative feedback modulation to neuronal response.Traveling Wave in CANNA moving bump in the network without relying on external drive.,计算神经科学,19,The strategy of the brainEvery animal adopts to its own optimal time scale suitable for its own survival in the natural environment.The brain co-evolves s
9、trategies to compensate delays. Why Neural Delays? To implement temporal code. To integrate multiple sensory cues. To integrate temporal information over time for reliable responding. Why Adaptive?Responding fast to a novel object with strong independent firing-rate code.Encoding information efficie
10、ntly with low frequency but correlation code lately.,计算神经科学,20,深度学习处理器,图像处理 : GPU信号处理 : DSP智能处理 : DianNao Family未来的硬件设备可能都需要一个深度学习算法的硬件处理器目前GPU芯片每年出货量：0.64 Billion产值：50 Billion $ NameSpeed upEnergy EfficiencyDianNao :1.1x GPU100 x CPUDaDianNao: 21x GPU330 x GPU PuDianNao: 1.1x GPU200 x GPUShiDianNao
11、: 28.94x GPU4688x GPU 320mW,21,深度学习处理器,如何加速？把数据从内存搬运到硬件运算单元，甚至比运算本身更耗能量。小尺度但支持大规模神经网络速度：把访存带宽用起来，尽可能提高性能能耗：通过优化片上存储层次尽量减少访存次数,22,深度学习处理器,DianNao16 NeuronsWin 8KBWout 8KBWeight 32KBDisadvantage:Memory ExchangeASPLOS Best Paper,23,深度学习处理器,DaDianNao16 Cores16 Neurons全网络模型片上存储无内存访问EDRAMMICRO Best Paper,24,深度学习处理器,PuDianNao关注数据局部性设计专用功能器件三根柱子：InstBuf 指令缓存HotBuf ColdBuf常见运算：向量内积 向量距离计数 非线性函数排序,25,深度学习处理器,ShiDianNao集成在摄像头中的视觉处理器直接对输入图像进行识别理解消除DRAM读写操作 ?DaDianNao + 可编程VLIW处理器结构,26,忆阻器,阻变式存储器，RRAM被认为是电阻、电容、电感外的第四种基本原件耐久性高、传输速度快、更低的功耗、可擦写,Q&AThank You,
