[{"data":1,"prerenderedAt":578},["ShallowReactive",2],{"content-query-3InqhlGR6R":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":572,"_id":573,"_source":574,"_file":575,"_stem":576,"_extension":577},"/technology-blogs/zh/2267","zh",false,"","MindSpore AI科学计算系列 | 几何深度学习系列——统一视角下的神经网络","在上一篇科学计算系列文章中，我们简单介绍了对称性与群的概念，以及对称性如何在几何学与物理学中如何扮演第一原理的角色。在本篇文章我们将看到，几何深度学习可以视为深度学习领域的埃尔朗根纲领，以对称性为第一原理，统一和分类各种神经网络架构。","2023-05-05","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/05/11/8941362fe8d64b8389eb8138ca0c6c25.png","technology-blogs","大V博文",{"type":15,"children":16,"toc":567},"root",[17,25,44,52,56,66,78,87,92,102,114,119,138,157,164,169,179,198,217,275,294,304,309,319,324,329,334,339,349,359,371,378,383,393,403,413,418,479,484,491,496,501,510,515,520,525,537,547,557],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"mindspore-ai科学计算系列-几何深度学习系列统一视角下的神经网络",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29,31,37,39],{"type":24,"value":30},"**作者：**",{"type":18,"tag":32,"props":33,"children":34},"strong",{},[35],{"type":24,"value":36},"于璠",{"type":24,"value":38}," ",{"type":18,"tag":32,"props":40,"children":41},{},[42],{"type":24,"value":43},"来源：知乎",{"type":18,"tag":26,"props":45,"children":46},{},[47],{"type":18,"tag":32,"props":48,"children":49},{},[50],{"type":24,"value":51},"背景",{"type":18,"tag":26,"props":53,"children":54},{},[55],{"type":24,"value":9},{"type":18,"tag":26,"props":57,"children":58},{},[59,61],{"type":24,"value":60},"**1、**",{"type":18,"tag":32,"props":62,"children":63},{},[64],{"type":24,"value":65},"对称性与神经网络",{"type":18,"tag":26,"props":67,"children":68},{},[69,71,76],{"type":24,"value":70},"随着算力的快速提升，深度学习发展迅速，并在计算视觉、自然语言等诸多领域展现其独有的优势。现如今，从最初的多层感知神经网络(MLP)，到现在各类诸如卷积神经网络(CNN)、循环神经网络(RNN)、图神经网络(GNN)以及Transformer等不同的神经网络架构，我们处于与十九世纪末几何学类似的情形。近几年，研究者以对称性为第一原理，发展出一套统一的神经网络理论框架——几何深度学习。其中，研究者提出“",{"type":18,"tag":32,"props":72,"children":73},{},[74],{"type":24,"value":75},"5G",{"type":24,"value":77},"”概念，即五个在几何深度学习中重要的研究对象：Grids（格网）、Group（群）、Graphs（图）、Geodesics（测地线）、Gauges（规范）。",{"type":18,"tag":26,"props":79,"children":80},{},[81],{"type":18,"tag":82,"props":83,"children":86},"img",{"alt":84,"src":85},"image.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230511072457.42197902386189653693107537872523:50540510081455:2400:D961A959D040D8932085858EC4A7BFC0761770D72D9F26F12AE611972E7FAF8B.png",[],{"type":18,"tag":26,"props":88,"children":89},{},[90],{"type":24,"value":91},"图1：\"5G\"",{"type":18,"tag":26,"props":93,"children":94},{},[95,97],{"type":24,"value":96},"**2、**",{"type":18,"tag":32,"props":98,"children":99},{},[100],{"type":24,"value":101},"空间与特征场",{"type":18,"tag":26,"props":103,"children":104},{},[105,107,112],{"type":24,"value":106},"为了对神经网络有更进一步的理解，我们需要明确两个概念：",{"type":18,"tag":32,"props":108,"children":109},{},[110],{"type":24,"value":111},"空间和特征场",{"type":24,"value":113},"。简单的说，空间是指信息在空间尺度上的延展，特征场则在空间上的每一点定义了一个特征矢量，神经网络的本质是一个特征场之间的映射。",{"type":18,"tag":26,"props":115,"children":116},{},[117],{"type":24,"value":118},"我们不妨回顾一下前面几类神经网络架构：MLP没有空间尺度，换句话说空间为一个点，而特征场即为网络层中的特征，MLP将一层中的特征映射到下一层的特征；二维CNN空间尺度为二维离散平面，特征场为平面上每一点的通道信息，包括最初的RGB通道，以及隐藏层中的多个通道信息。类似地，RNN是在时序尺度上延展，GNN是在一般的图结构上延展，而Transformer可以视为一个特殊的完全图GNN。",{"type":18,"tag":26,"props":120,"children":121},{},[122,124,129,131,136],{"type":24,"value":123},"群的表示指群到线性空间的同态映射。空间中点的变换通过",{"type":18,"tag":32,"props":125,"children":126},{},[127],{"type":24,"value":128},"对称变换群",{"type":24,"value":130},"描述，与之对应的特征场之间的变换通过对称变换",{"type":18,"tag":32,"props":132,"children":133},{},[134],{"type":24,"value":135},"群的表示",{"type":24,"value":137},"描述。换句话说，神经网络映射是表示之间的映射。",{"type":18,"tag":26,"props":139,"children":140},{},[141,143,148,150,155],{"type":24,"value":142},"以图片为例，图片上一个点的特征为RGB三个数值，即三个标量的直和，如图5上所示。这里的标量就是变换群的平凡表示，其数值在变换群的作用下保持不变。",{"type":18,"tag":32,"props":144,"children":145},{},[146],{"type":24,"value":147},"平凡表示",{"type":24,"value":149},"是指，对于任意变换群，表示映射都将其映射为一个恒等矩阵，这意味着",{"type":18,"tag":32,"props":151,"children":152},{},[153],{"type":24,"value":154},"对称性信息在映射过程中损失",{"type":24,"value":156},"了。如果要求神经网络保持一定的对称性，那么需要我们用非平凡表示来描述特征场的变换，其表示矩阵的直和分解会存在维度大于1的子矩阵，于是神经网络则为每一块子矩阵保持对称性的映射，如图2下所示。",{"type":18,"tag":26,"props":158,"children":159},{},[160],{"type":18,"tag":82,"props":161,"children":163},{"alt":84,"src":162},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230511072522.83503069666389006883077658075699:50540510081455:2400:4F55DBF988B5DA01981608ADC54EEEE5B41378BBA74E3AB59E8720017D220686.png",[],{"type":18,"tag":26,"props":165,"children":166},{},[167],{"type":24,"value":168},"图2：特征场间变换群的表示",{"type":18,"tag":26,"props":170,"children":171},{},[172,174],{"type":24,"value":173},"**3、**",{"type":18,"tag":32,"props":175,"children":176},{},[177],{"type":24,"value":178},"不变性与等变性",{"type":18,"tag":26,"props":180,"children":181},{},[182,184,189,191,196],{"type":24,"value":183},"正如我们上面提到的，如果特征场的变换只是一些平凡表示，那么神经网络作为一个特征场之间的映射，",{"type":18,"tag":32,"props":185,"children":186},{},[187],{"type":24,"value":188},"对输入特征场的变换不会影响其输出特征场",{"type":24,"value":190},"，则称神经网络对此变换群具有",{"type":18,"tag":32,"props":192,"children":193},{},[194],{"type":24,"value":195},"不变性",{"type":24,"value":197},"。",{"type":18,"tag":26,"props":199,"children":200},{},[201,203,215],{"type":24,"value":202},"而对于非平凡表示，",{"type":18,"tag":32,"props":204,"children":205},{},[206,208,213],{"type":24,"value":207},"对输入特征场的变换，会导致其输出特征场做相应的变换，",{"type":18,"tag":32,"props":209,"children":210},{},[211],{"type":24,"value":212},"变换方式取决于变换群的表示，则称神经网络对此变换群具有",{"type":24,"value":214},"等变性",{"type":24,"value":216},"。可以看到，映射的不变性是等变性的一种特例。",{"type":18,"tag":26,"props":218,"children":219},{},[220,222,227,229,234,236,241,242,247,249,254,255,260,262,267,269,274],{"type":24,"value":221},"直观的理解，神经网络的等变性保持了某种对称性信息。例如，",{"type":18,"tag":32,"props":223,"children":224},{},[225],{"type":24,"value":226},"CNN",{"type":24,"value":228},"具有",{"type":18,"tag":32,"props":230,"children":231},{},[232],{"type":24,"value":233},"空间平移等变性",{"type":24,"value":235},"，图像识别任务中，当图像中一个物体进行平移时，其特征图像也会进行相应的平移，因此网络可以正确地识别这个物体；类似地，",{"type":18,"tag":32,"props":237,"children":238},{},[239],{"type":24,"value":240},"RNN",{"type":24,"value":228},{"type":18,"tag":32,"props":243,"children":244},{},[245],{"type":24,"value":246},"时序平移等变性",{"type":24,"value":248},"，只是这里的平移是有方向性的；",{"type":18,"tag":32,"props":250,"children":251},{},[252],{"type":24,"value":253},"GNN",{"type":24,"value":228},{"type":18,"tag":32,"props":256,"children":257},{},[258],{"type":24,"value":259},"顶点编号的规范等变性",{"type":24,"value":261},"，如果用邻接矩阵表示图的信息，那么对图顶点编号的置换变换，也会同样反应到邻接矩阵上。而对于最基础的",{"type":18,"tag":32,"props":263,"children":264},{},[265],{"type":24,"value":266},"MLP",{"type":24,"value":268},"，其具有",{"type":18,"tag":32,"props":270,"children":271},{},[272],{"type":24,"value":273},"数据集合置换规范等变性",{"type":24,"value":197},{"type":18,"tag":26,"props":276,"children":277},{},[278,280,285,287,292],{"type":24,"value":279},"可以看到，",{"type":18,"tag":32,"props":281,"children":282},{},[283],{"type":24,"value":284},"任何神经网络都是某种对称性的等变映射",{"type":24,"value":286},"。在[2]中，作者提出了另一个关于等变性的重要结论：",{"type":18,"tag":32,"props":288,"children":289},{},[290],{"type":24,"value":291},"等变性 = 坐标无关 + 权值共享",{"type":24,"value":293},"。例如CNN的等变性可以看作空间位置坐标无关，结合卷积核在空间平移中权值共享。特征场这个观察加深了人们对于等变性的理解，对于在更一般的空间结构，例如弯曲空间中设计神经网络非常重要。我们会在后面简单介绍一下如何在黎曼流形上进行卷积。",{"type":18,"tag":26,"props":295,"children":296},{},[297,299],{"type":24,"value":298},"**4、**",{"type":18,"tag":32,"props":300,"children":301},{},[302],{"type":24,"value":303},"范畴论观点",{"type":18,"tag":26,"props":305,"children":306},{},[307],{"type":24,"value":308},"我们还可以进一步抽象，从范畴论的角度来看神经网络的本质。",{"type":18,"tag":26,"props":310,"children":311},{},[312,317],{"type":18,"tag":32,"props":313,"children":314},{},[315],{"type":24,"value":316},"范畴",{"type":24,"value":318},"是数学的一种抽象工具，用来描述不同的数学结构及它们之间的关系。一个范畴由以下三个部分组成：",{"type":18,"tag":26,"props":320,"children":321},{},[322],{"type":24,"value":323},"1、一类对象，可以是集合、群、拓扑空间等任何具有某种结构的数学实体。",{"type":18,"tag":26,"props":325,"children":326},{},[327],{"type":24,"value":328},"2、一类态射，可以是函数、群同态、连续映射等任何保持结构的映射。每个态射都有一个源对象和一个目标对象，表示映射的方向。",{"type":18,"tag":26,"props":330,"children":331},{},[332],{"type":24,"value":333},"3、一种态射的复合运算，满足结合律和单位元的存在性。",{"type":18,"tag":26,"props":335,"children":336},{},[337],{"type":24,"value":338},"一个范畴的例子是集合范畴，它的对象是所有的集合，它的态射是所有的函数，它的复合运算就是函数的复合。",{"type":18,"tag":26,"props":340,"children":341},{},[342,347],{"type":18,"tag":32,"props":343,"children":344},{},[345],{"type":24,"value":346},"函子",{"type":24,"value":348},"是范畴间的一类映射，可以将一个范畴的对象和态射变换到另一个范畴中，同时保持单位态射和态射的复合。例如，群的表示是一种函子，它将群范畴映射到线性空间范畴。",{"type":18,"tag":26,"props":350,"children":351},{},[352,357],{"type":18,"tag":32,"props":353,"children":354},{},[355],{"type":24,"value":356},"自然变换",{"type":24,"value":358},"是一种将一个函子变为另一个函子的映射，同时保持范畴的内在结构不变。",{"type":18,"tag":26,"props":360,"children":361},{},[362,364,369],{"type":24,"value":363},"前面提到，所有的神经网络都是某种对称性的等变映射，如果我们从范畴论的视角来看，这个结论是显然的，因为神经网络的本质是",{"type":18,"tag":32,"props":365,"children":366},{},[367],{"type":24,"value":368},"群表示函子的自然变换",{"type":24,"value":370},"，不同范畴对应于不同的网络架构。例如集合范畴对应于MLP，而图范畴对应于GNN。",{"type":18,"tag":26,"props":372,"children":373},{},[374],{"type":18,"tag":82,"props":375,"children":377},{"alt":84,"src":376},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230511072545.77471245932998293625778797976559:50540510081455:2400:8BB69851FA293BDB7C0B01660EC87A247C5AB21F35B7EA890AE95993A39AC8C2.png",[],{"type":18,"tag":26,"props":379,"children":380},{},[381],{"type":24,"value":382},"图3：群表示函子ρ(g)的自然变换η",{"type":18,"tag":26,"props":384,"children":385},{},[386,388],{"type":24,"value":387},"**5、**",{"type":18,"tag":32,"props":389,"children":390},{},[391],{"type":24,"value":392},"神经网络设计蓝图",{"type":18,"tag":26,"props":394,"children":395},{},[396,398],{"type":24,"value":397},"至此，我们对神经网络有了一个统一的整体理解，由此可以获得一般性的",{"type":18,"tag":32,"props":399,"children":400},{},[401],{"type":24,"value":402},"神经网络设计蓝****图：",{"type":18,"tag":26,"props":404,"children":405},{},[406],{"type":18,"tag":32,"props":407,"children":408},{},[409],{"type":18,"tag":82,"props":410,"children":412},{"alt":84,"src":411},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230511072605.71967395996558933693056256407037:50540510081455:2400:5824EB1A2CD214B1F9BE8DCEF5D827DCFC84605BA010F7E758CBD70D2C2614C4.png",[],{"type":18,"tag":26,"props":414,"children":415},{},[416],{"type":24,"value":417},"图4：神经网络设计蓝图",{"type":18,"tag":26,"props":419,"children":420},{},[421,423,428,430,435,437,442,444,449,451,456,458,463,465,470,472,477],{"type":24,"value":422},"一个神经网络需要明确信息延展的",{"type":18,"tag":32,"props":424,"children":425},{},[426],{"type":24,"value":427},"空间",{"type":24,"value":429},"/域(domain)，空间的",{"type":18,"tag":32,"props":431,"children":432},{},[433],{"type":24,"value":434},"变换群",{"type":24,"value":436},"(G)，空间上的",{"type":18,"tag":32,"props":438,"children":439},{},[440],{"type":24,"value":441},"特征场",{"type":24,"value":443},"/信号(signals)，特征场的",{"type":18,"tag":32,"props":445,"children":446},{},[447],{"type":24,"value":448},"变换的表示",{"type":24,"value":450},"(ρ(G))。网络包含以下几个模块：",{"type":18,"tag":32,"props":452,"children":453},{},[454],{"type":24,"value":455},"线性等变",{"type":24,"value":457},"层，",{"type":18,"tag":32,"props":459,"children":460},{},[461],{"type":24,"value":462},"非线性",{"type":24,"value":464},"（等变）层，",{"type":18,"tag":32,"props":466,"children":467},{},[468],{"type":24,"value":469},"局部池化",{"type":24,"value":471},"层（粗粒化），",{"type":18,"tag":32,"props":473,"children":474},{},[475],{"type":24,"value":476},"线性不变",{"type":24,"value":478},"层（全局池化）。",{"type":18,"tag":26,"props":480,"children":481},{},[482],{"type":24,"value":483},"以CNN为例，空间是二维离散平面，变换群是二维平移群，特征场是图片的通道信息，变换的表示是平凡表示，线性等变层是卷积层，非线性层可以是任意非线性函数（因为特征场由标量组成，不会对非线性层做出约束），局部池化层是CNN中的池化层，线性不变层是网络末端的MLP。",{"type":18,"tag":26,"props":485,"children":486},{},[487],{"type":18,"tag":82,"props":488,"children":490},{"alt":84,"src":489},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230511072636.02468238796893644948619901225314:50540510081455:2400:0A9CEB34178BE6D349E3DA7404B4D112A952700E7F9DC7816CD55B2633A65DC0.png",[],{"type":18,"tag":26,"props":492,"children":493},{},[494],{"type":24,"value":495},"图5：CNN网络架构",{"type":18,"tag":26,"props":497,"children":498},{},[499],{"type":24,"value":500},"至此，我们从宏观统一的视角回顾了各类神经网络架构，并介绍了具有对称性的神经网络设计蓝图。在下一篇文章中，我们将具体来看看三类典型具有对称性的神经网络：图神经网络、群等变神经网络和流形上的神经网络。",{"type":18,"tag":502,"props":503,"children":505},"h2",{"id":504},"参考文献",[506],{"type":18,"tag":32,"props":507,"children":508},{},[509],{"type":24,"value":504},{"type":18,"tag":26,"props":511,"children":512},{},[513],{"type":24,"value":514},"[1]Bronstein, Michael M., et al. \"Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.\" arXiv preprint arXiv:2104.13478 (2021).",{"type":18,"tag":26,"props":516,"children":517},{},[518],{"type":24,"value":519},"[2]Weiler, Maurice, et al. \"Coordinate Independent Convolutional Networks--Isometry and Gauge Equivariant Convolutions on Riemannian Manifolds.\" arXiv preprint arXiv:2106.06020 (2021).",{"type":18,"tag":26,"props":521,"children":522},{},[523],{"type":24,"value":524},"往期回顾",{"type":18,"tag":26,"props":526,"children":527},{},[528],{"type":18,"tag":529,"props":530,"children":534},"a",{"href":531,"rel":532},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247594679&idx=1&sn=9c128e792b6a6d19bc029ce53c7b4495&chksm=c11e62f8f669ebee13848ba4790f2ae94c53054e057088eac13ab5711e44344ffefbc531043f&scene=21#wechat_redirect",[533],"nofollow",[535],{"type":24,"value":536},"MindSpore AI科学计算系列 | 几何深度学习系列——对称性与守恒律",{"type":18,"tag":26,"props":538,"children":539},{},[540],{"type":18,"tag":529,"props":541,"children":544},{"href":542,"rel":543},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247590517&idx=1&sn=d6c1ae0237eca9df594c153e672ae24a&chksm=c11e723af669fb2c046132f28ec79f28950f4b3627ac7fd1dbcefe73388d94c324cab83ae675&scene=21#wechat_redirect",[533],[545],{"type":24,"value":546},"MindSpore AI科学计算系列 | AI与“科学”如何分工与协作",{"type":18,"tag":26,"props":548,"children":549},{},[550],{"type":18,"tag":529,"props":551,"children":554},{"href":552,"rel":553},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247589601&idx=1&sn=ec6c3b0063b67ba2871e623f83c00dcd&chksm=c11e8eaef66907b8446f05ab2abe7394dad5e749b095bc20fb42aff762629be54021603c351e&scene=21#wechat_redirect",[533],[555],{"type":24,"value":556},"MindSpore AI科学计算系列 | 非定常气动力代理模型及在增升装置的应用",{"type":18,"tag":26,"props":558,"children":559},{},[560],{"type":18,"tag":529,"props":561,"children":564},{"href":562,"rel":563},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247585665&idx=1&sn=084030e97a7ad59155aae97d13030f83&chksm=c11e9fcef66916d8a7062ac95a3033e09715de4dba69ded78fc2df423e5573ba3bb8c1a83645&scene=21#wechat_redirect",[533],[565],{"type":24,"value":566},"MindSpore AI科学计算系列 | 等变神经网络与e3nn数学库",{"title":7,"searchDepth":568,"depth":568,"links":569},4,[570],{"id":504,"depth":571,"text":504},2,"markdown","content:technology-blogs:zh:2267.md","content","technology-blogs/zh/2267.md","technology-blogs/zh/2267","md",1776506121534]