[{"data":1,"prerenderedAt":5539},["ShallowReactive",2],{"post-\u002Fblog\u002F2026\u002F2026-08-19-yan-jiu-guan-xi-de-san-zhong-shi-jiao":3},{"post":4,"nextPost":5259,"prevPost":5269},{"id":5,"title":6,"body":7,"description":13,"draft":5256,"enableComment":5257,"extension":5258,"image":5247,"important":5256,"location":5259,"meta":5260,"navigation":5257,"ogImage":5259,"onday":5261,"path":5262,"seo":5263,"stem":5264,"summary":5265,"tags":5266,"__hash__":5268},"blog\u002Fblog\u002F2026\u002F2026-08-19-yan-jiu-guan-xi-de-san-zhong-shi-jiao.md","研究关系的三种视角",{"type":8,"value":9,"toc":5246},"minimark",[10,14,17,20,25,28,296,975,1170,1174,1177,1579,1583,1586,2063,2342,3155,3158,3161,3221,3569,3664,3939,3942,3945,3948,4059,4447,4450,4453,4526,4529,5240,5243],[11,12,13],"p",{},"哲学里有一种命题：「存在即关系」，在数理研究中同样也有很深的体现。听起来很抽象，但它其实是对世界本质是什么的一种深刻回答。理解事物不是先存在、再去建立关系，而是事物在建立关系的过程中，才得以成为事物。它最著名的提出者是 20 世纪法国哲学家怀特海（过程哲学），但其思想渊源可以追溯到德国哲学家莱布尼茨（单子论）和古希腊的赫拉克利特（万物流动）。",[11,15,16],{},"所以，研究「存在」也是研究「关系」。而关系通常有三种视角：几何、拓扑和因果。",[11,18,19],{},"拓扑提供稳健的骨架，几何赋予精细的尺度，因果指明干预的方向。三者共同构成智能系统理解世界的先验地基。更准确的说法或许是：几何和拓扑是关于「关系的静态结构」（分别是度量化的和非度量化的），而因果是关于「关系的动态\u002F生成机制」，它回答的不是「这两者如何关联」，而是「如果我改变一个，另一个会怎样」。",[21,22,24],"h2",{"id":23},"几何度量结构","几何（度量结构）",[11,26,27],{},"几何关注的是「测量」。基本问题是多接近、多相似，或者一个具体的度量数值。——它需要一个度量结构（距离、内积、相似度函数），通常要求满足对称性、三角不等式等公理。典型例子：嵌入空间里两个向量的余弦相似度、社会心理学里的「亲密度量表」。研究的是关系的强度。几何关系是最重的——信息量最大，但也最脆弱，度量方式一变，结论可能就变。",[11,29,30,31,114,115,295],{},"一个度量空间是二元组 ",[32,33,37,75],"span",{"className":34,"translate":36},[35],"katex","no",[32,38,41],{"className":39},[40],"katex-mathml",[42,43,45],"math",{"xmlns":44},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML",[46,47,48,70],"semantics",{},[49,50,51,56,60,64,67],"mrow",{},[52,53,55],"mo",{"stretchy":54},"false","(",[57,58,59],"mi",{},"X",[52,61,63],{"separator":62},"true",",",[57,65,66],{},"d",[52,68,69],{"stretchy":54},")",[71,72,74],"annotation",{"encoding":73},"application\u002Fx-tex","(X, d)",[32,76,79],{"className":77,"ariaHidden":62},[78],"katex-html",[32,80,83,88,92,98,102,107,110],{"className":81},[82],"base",[32,84],{"className":85,"style":87},[86],"strut","height:1em;vertical-align:-0.25em;",[32,89,55],{"className":90},[91],"mopen",[32,93,59],{"className":94,"style":97},[95,96],"mord","mathnormal","margin-right:0.0785em;",[32,99,63],{"className":100},[101],"mpunct",[32,103],{"className":104,"style":106},[105],"mspace","margin-right:0.1667em;",[32,108,66],{"className":109},[95,96],[32,111,69],{"className":112},[113],"mclose","，其中 ",[32,116,118,161],{"className":117,"translate":36},[35],[32,119,121],{"className":120},[40],[42,122,123],{"xmlns":44},[46,124,125,158],{},[49,126,127,129,132,134,137,139,142],{},[57,128,66],{},[52,130,131],{},":",[57,133,59],{},[52,135,136],{},"×",[57,138,59],{},[52,140,141],{},"→",[143,144,145,149],"msub",{},[57,146,148],{"mathvariant":147},"double-struck","R",[49,150,151,154],{},[52,152,153],{},"≥",[155,156,157],"mn",{},"0",[71,159,160],{"encoding":73},"d: X \\times X \\to \\mathbb{R}_{\\geq 0}",[32,162,164,185,206,225],{"className":163,"ariaHidden":62},[78],[32,165,167,171,174,178,182],{"className":166},[82],[32,168],{"className":169,"style":170},[86],"height:0.6944em;",[32,172,66],{"className":173},[95,96],[32,175],{"className":176,"style":177},[105],"margin-right:0.2778em;",[32,179,131],{"className":180},[181],"mrel",[32,183],{"className":184,"style":177},[105],[32,186,188,192,195,199,203],{"className":187},[82],[32,189],{"className":190,"style":191},[86],"height:0.7667em;vertical-align:-0.0833em;",[32,193,59],{"className":194,"style":97},[95,96],[32,196],{"className":197,"style":198},[105],"margin-right:0.2222em;",[32,200,136],{"className":201},[202],"mbin",[32,204],{"className":205,"style":198},[105],[32,207,209,213,216,219,222],{"className":208},[82],[32,210],{"className":211,"style":212},[86],"height:0.6833em;",[32,214,59],{"className":215,"style":97},[95,96],[32,217],{"className":218,"style":177},[105],[32,220,141],{"className":221},[181],[32,223],{"className":224,"style":177},[105],[32,226,228,232],{"className":227},[82],[32,229],{"className":230,"style":231},[86],"height:0.9341em;vertical-align:-0.2452em;",[32,233,235,239],{"className":234},[95],[32,236,148],{"className":237},[95,238],"mathbb",[32,240,243],{"className":241},[242],"msupsub",[32,244,248,286],{"className":245},[246,247],"vlist-t","vlist-t2",[32,249,252,281],{"className":250},[251],"vlist-r",[32,253,257],{"className":254,"style":256},[255],"vlist","height:0.3011em;",[32,258,260,265],{"style":259},"top:-2.55em;margin-left:0em;margin-right:0.05em;",[32,261],{"className":262,"style":264},[263],"pstrut","height:2.7em;",[32,266,272],{"className":267},[268,269,270,271],"sizing","reset-size6","size3","mtight",[32,273,275,278],{"className":274},[95,271],[32,276,153],{"className":277},[181,271],[32,279,157],{"className":280},[95,271],[32,282,285],{"className":283},[284],"vlist-s","​",[32,287,289],{"className":288},[251],[32,290,293],{"className":291,"style":292},[255],"height:0.2452em;",[32,294],{}," 满足：",[32,297,300],{"className":298,"translate":36},[299],"katex-display",[32,301,303,494],{"className":302,"translate":36},[35],[32,304,306],{"className":305},[40],[42,307,309],{"xmlns":44,"display":308},"block",[46,310,311,491],{},[49,312,313,316],{},[52,314,315],{"fence":62},"{",[317,318,322],"mtable",{"rowspacing":319,"columnalign":320,"columnspacing":321},"0.36em","left left","1em",[323,324,325],"mtr",{},[326,327,328],"mtd",{},[329,330,331],"mstyle",{"scriptlevel":157,"displaystyle":54},[317,332,336,388,432],{"rowspacing":333,"columnalign":334,"columnspacing":335},"0.25em","right left","0em",[323,337,338,344],{},[326,339,340],{},[329,341,342],{"scriptlevel":157,"displaystyle":62},[49,343],{},[326,345,346],{},[329,347,348],{"scriptlevel":157,"displaystyle":62},[49,349,350,352,354,356,359,361,364,366,369,371,375,378,380,382,384,386],{},[49,351],{},[57,353,66],{},[52,355,55],{"stretchy":54},[57,357,358],{},"x",[52,360,63],{"separator":62},[57,362,363],{},"y",[52,365,69],{"stretchy":54},[52,367,368],{},"=",[155,370,157],{},[372,373,374],"mtext",{},"  ",[52,376,377],{},"⟺",[372,379,374],{},[57,381,358],{},[52,383,368],{},[57,385,363],{},[52,387,63],{"separator":62},[323,389,390,396],{},[326,391,392],{},[329,393,394],{"scriptlevel":157,"displaystyle":62},[49,395],{},[326,397,398],{},[329,399,400],{"scriptlevel":157,"displaystyle":62},[49,401,402,404,406,408,410,412,414,416,418,420,422,424,426,428,430],{},[49,403],{},[57,405,66],{},[52,407,55],{"stretchy":54},[57,409,358],{},[52,411,63],{"separator":62},[57,413,363],{},[52,415,69],{"stretchy":54},[52,417,368],{},[57,419,66],{},[52,421,55],{"stretchy":54},[57,423,363],{},[52,425,63],{"separator":62},[57,427,358],{},[52,429,69],{"stretchy":54},[52,431,63],{"separator":62},[323,433,434,440],{},[326,435,436],{},[329,437,438],{"scriptlevel":157,"displaystyle":62},[49,439],{},[326,441,442],{},[329,443,444],{"scriptlevel":157,"displaystyle":62},[49,445,446,448,450,452,454,456,459,461,464,466,468,470,472,474,476,479,481,483,485,487,489],{},[49,447],{},[57,449,66],{},[52,451,55],{"stretchy":54},[57,453,358],{},[52,455,63],{"separator":62},[57,457,458],{},"z",[52,460,69],{"stretchy":54},[52,462,463],{},"≤",[57,465,66],{},[52,467,55],{"stretchy":54},[57,469,358],{},[52,471,63],{"separator":62},[57,473,363],{},[52,475,69],{"stretchy":54},[52,477,478],{},"+",[57,480,66],{},[52,482,55],{"stretchy":54},[57,484,363],{},[52,486,63],{"separator":62},[57,488,458],{},[52,490,69],{"stretchy":54},[71,492,493],{"encoding":73},"\\begin{cases}\n\\begin{aligned}\n&d(x,y) = 0 \\iff x = y,\\\\\n&d(x,y) = d(y,x),\\\\\n&d(x,z) \\leq d(x,y) + d(y,z)\n\\end{aligned}\n\\end{cases}",[32,495,497],{"className":496,"ariaHidden":62},[78],[32,498,500,504],{"className":499},[82],[32,501],{"className":502,"style":503},[86],"height:4.2em;vertical-align:-1.85em;",[32,505,508,611,971],{"className":506},[507],"minner",[32,509,511],{"className":510},[91],[32,512,516],{"className":513},[514,515],"delimsizing","mult",[32,517,519,602],{"className":518},[246,247],[32,520,522,599],{"className":521},[251],[32,523,526,541,563,575,587],{"className":524,"style":525},[255],"height:2.35em;",[32,527,529,533],{"style":528},"top:-2.2em;",[32,530],{"className":531,"style":532},[263],"height:3.15em;",[32,534,538],{"className":535},[536,537],"delimsizinginner","delim-size4",[32,539,540],{},"⎩",[32,542,544,547],{"style":543},"top:-2.192em;",[32,545],{"className":546,"style":532},[263],[32,548,550],{"style":549},"height:0.316em;width:0.8889em;",[551,552,559],"svg",{"xmlns":553,"width":554,"height":555,"style":556,"viewBox":557,"preserveAspectRatio":558},"http:\u002F\u002Fwww.w3.org\u002F2000\u002Fsvg","0.8889em","0.316em","width:0.8889em","0 0 888.89 316","xMinYMin",[560,561],"path",{"d":562},"M384 0 H504 V316 H384z M384 0 H504 V316 H384z",[32,564,566,569],{"style":565},"top:-3.15em;",[32,567],{"className":568,"style":532},[263],[32,570,572],{"className":571},[536,537],[32,573,574],{},"⎨",[32,576,578,581],{"style":577},"top:-4.292em;",[32,579],{"className":580,"style":532},[263],[32,582,583],{"style":549},[551,584,585],{"xmlns":553,"width":554,"height":555,"style":556,"viewBox":557,"preserveAspectRatio":558},[560,586],{"d":562},[32,588,590,593],{"style":589},"top:-4.6em;",[32,591],{"className":592,"style":532},[263],[32,594,596],{"className":595},[536,537],[32,597,598],{},"⎧",[32,600,285],{"className":601},[284],[32,603,605],{"className":604},[251],[32,606,609],{"className":607,"style":608},[255],"height:1.85em;",[32,610],{},[32,612,614],{"className":613},[95],[32,615,617],{"className":616},[317],[32,618,621],{"className":619},[620],"col-align-l",[32,622,624,963],{"className":623},[246,247],[32,625,627,960],{"className":626},[251],[32,628,630],{"className":629,"style":525},[255],[32,631,633,637],{"style":632},"top:-4.35em;",[32,634],{"className":635,"style":636},[263],"height:4.35em;",[32,638,640],{"className":639},[95],[32,641,643],{"className":642},[95],[32,644,646,697],{"className":645},[317],[32,647,650],{"className":648},[649],"col-align-r",[32,651,653,689],{"className":652},[246,247],[32,654,656,686],{"className":655},[251],[32,657,659,668,677],{"className":658,"style":525},[255],[32,660,661,665],{"style":632},[32,662],{"className":663,"style":664},[263],"height:2.84em;",[32,666],{"className":667},[95],[32,669,671,674],{"style":670},"top:-2.85em;",[32,672],{"className":673,"style":664},[263],[32,675],{"className":676},[95],[32,678,680,683],{"style":679},"top:-1.35em;",[32,681],{"className":682,"style":664},[263],[32,684],{"className":685},[95],[32,687,285],{"className":688},[284],[32,690,692],{"className":691},[251],[32,693,695],{"className":694,"style":608},[255],[32,696],{},[32,698,700],{"className":699},[620],[32,701,703,952],{"className":702},[246,247],[32,704,706,949],{"className":705},[251],[32,707,709,789,855],{"className":708,"style":525},[255],[32,710,712,716],{"style":711},"top:-4.51em;",[32,713],{"className":714,"style":715},[263],"height:3em;",[32,717,719,722,725,728,731,734,737,741,744,747,750,753,756,759,762,765,768,771,774,777,780,783,786],{"className":718},[95],[32,720],{"className":721},[95],[32,723,66],{"className":724},[95,96],[32,726,55],{"className":727},[91],[32,729,358],{"className":730},[95,96],[32,732,63],{"className":733},[101],[32,735],{"className":736,"style":106},[105],[32,738,363],{"className":739,"style":740},[95,96],"margin-right:0.0359em;",[32,742,69],{"className":743},[113],[32,745],{"className":746,"style":177},[105],[32,748,368],{"className":749},[181],[32,751],{"className":752,"style":177},[105],[32,754,157],{"className":755},[95],[32,757],{"className":758,"style":177},[105],[32,760],{"className":761,"style":177},[105],[32,763,377],{"className":764},[181],[32,766],{"className":767,"style":177},[105],[32,769],{"className":770,"style":177},[105],[32,772,358],{"className":773},[95,96],[32,775],{"className":776,"style":177},[105],[32,778,368],{"className":779},[181],[32,781],{"className":782,"style":177},[105],[32,784,363],{"className":785,"style":740},[95,96],[32,787,63],{"className":788},[101],[32,790,792,795],{"style":791},"top:-3.01em;",[32,793],{"className":794,"style":715},[263],[32,796,798,801,804,807,810,813,816,819,822,825,828,831,834,837,840,843,846,849,852],{"className":797},[95],[32,799],{"className":800},[95],[32,802,66],{"className":803},[95,96],[32,805,55],{"className":806},[91],[32,808,358],{"className":809},[95,96],[32,811,63],{"className":812},[101],[32,814],{"className":815,"style":106},[105],[32,817,363],{"className":818,"style":740},[95,96],[32,820,69],{"className":821},[113],[32,823],{"className":824,"style":177},[105],[32,826,368],{"className":827},[181],[32,829],{"className":830,"style":177},[105],[32,832,66],{"className":833},[95,96],[32,835,55],{"className":836},[91],[32,838,363],{"className":839,"style":740},[95,96],[32,841,63],{"className":842},[101],[32,844],{"className":845,"style":106},[105],[32,847,358],{"className":848},[95,96],[32,850,69],{"className":851},[113],[32,853,63],{"className":854},[101],[32,856,858,861],{"style":857},"top:-1.51em;",[32,859],{"className":860,"style":715},[263],[32,862,864,867,870,873,876,879,882,886,889,892,895,898,901,904,907,910,913,916,919,922,925,928,931,934,937,940,943,946],{"className":863},[95],[32,865],{"className":866},[95],[32,868,66],{"className":869},[95,96],[32,871,55],{"className":872},[91],[32,874,358],{"className":875},[95,96],[32,877,63],{"className":878},[101],[32,880],{"className":881,"style":106},[105],[32,883,458],{"className":884,"style":885},[95,96],"margin-right:0.044em;",[32,887,69],{"className":888},[113],[32,890],{"className":891,"style":177},[105],[32,893,463],{"className":894},[181],[32,896],{"className":897,"style":177},[105],[32,899,66],{"className":900},[95,96],[32,902,55],{"className":903},[91],[32,905,358],{"className":906},[95,96],[32,908,63],{"className":909},[101],[32,911],{"className":912,"style":106},[105],[32,914,363],{"className":915,"style":740},[95,96],[32,917,69],{"className":918},[113],[32,920],{"className":921,"style":198},[105],[32,923,478],{"className":924},[202],[32,926],{"className":927,"style":198},[105],[32,929,66],{"className":930},[95,96],[32,932,55],{"className":933},[91],[32,935,363],{"className":936,"style":740},[95,96],[32,938,63],{"className":939},[101],[32,941],{"className":942,"style":106},[105],[32,944,458],{"className":945,"style":885},[95,96],[32,947,69],{"className":948},[113],[32,950,285],{"className":951},[284],[32,953,955],{"className":954},[251],[32,956,958],{"className":957,"style":608},[255],[32,959],{},[32,961,285],{"className":962},[284],[32,964,966],{"className":965},[251],[32,967,969],{"className":968,"style":608},[255],[32,970],{},[32,972],{"className":973},[113,974],"nulldelimiter",[11,976,977,978,1054,1055,1169],{},"三条公理分别对应「可分辨性」「对称性」「三角不等式」。注意对称性是内嵌在定义里的——这一点后面会成为几何\u002F拓扑与因果的分水岭。在 ML 里，我们常用的是更强的结构：内积空间 ",[32,979,981,1014],{"className":980,"translate":36},[35],[32,982,984],{"className":983},[40],[42,985,986],{"xmlns":44},[46,987,988,1011],{},[49,989,990,992,994,996,999,1002,1004,1006,1009],{},[52,991,55],{"stretchy":54},[57,993,59],{},[52,995,63],{"separator":62},[52,997,998],{"stretchy":54},"⟨",[52,1000,1001],{},"⋅",[52,1003,63],{"separator":62},[52,1005,1001],{},[52,1007,1008],{"stretchy":54},"⟩",[52,1010,69],{"stretchy":54},[71,1012,1013],{"encoding":73},"(X, \\langle \\cdot,\\cdot\\rangle)",[32,1015,1017],{"className":1016,"ariaHidden":62},[78],[32,1018,1020,1023,1026,1029,1032,1035,1038,1041,1044,1047,1050],{"className":1019},[82],[32,1021],{"className":1022,"style":87},[86],[32,1024,55],{"className":1025},[91],[32,1027,59],{"className":1028,"style":97},[95,96],[32,1030,63],{"className":1031},[101],[32,1033],{"className":1034,"style":106},[105],[32,1036,998],{"className":1037},[91],[32,1039,1001],{"className":1040},[95],[32,1042,63],{"className":1043},[101],[32,1045],{"className":1046,"style":106},[105],[32,1048,1001],{"className":1049},[95],[32,1051,1053],{"className":1052},[113],"⟩)","，度量由内积诱导 ",[32,1056,1058,1097],{"className":1057,"translate":36},[35],[32,1059,1061],{"className":1060},[40],[42,1062,1063],{"xmlns":44},[46,1064,1065,1094],{},[49,1066,1067,1069,1071,1073,1075,1077,1079,1081,1085,1087,1090,1092],{},[57,1068,66],{},[52,1070,55],{"stretchy":54},[57,1072,358],{},[52,1074,63],{"separator":62},[57,1076,363],{},[52,1078,69],{"stretchy":54},[52,1080,368],{},[57,1082,1084],{"mathvariant":1083},"normal","∥",[57,1086,358],{},[52,1088,1089],{},"−",[57,1091,363],{},[57,1093,1084],{"mathvariant":1083},[71,1095,1096],{"encoding":73},"d(x,y) = \\|x-y\\|",[32,1098,1100,1136,1157],{"className":1099,"ariaHidden":62},[78],[32,1101,1103,1106,1109,1112,1115,1118,1121,1124,1127,1130,1133],{"className":1102},[82],[32,1104],{"className":1105,"style":87},[86],[32,1107,66],{"className":1108},[95,96],[32,1110,55],{"className":1111},[91],[32,1113,358],{"className":1114},[95,96],[32,1116,63],{"className":1117},[101],[32,1119],{"className":1120,"style":106},[105],[32,1122,363],{"className":1123,"style":740},[95,96],[32,1125,69],{"className":1126},[113],[32,1128],{"className":1129,"style":177},[105],[32,1131,368],{"className":1132},[181],[32,1134],{"className":1135,"style":177},[105],[32,1137,1139,1142,1145,1148,1151,1154],{"className":1138},[82],[32,1140],{"className":1141,"style":87},[86],[32,1143,1084],{"className":1144},[95],[32,1146,358],{"className":1147},[95,96],[32,1149],{"className":1150,"style":198},[105],[32,1152,1089],{"className":1153},[202],[32,1155],{"className":1156,"style":198},[105],[32,1158,1160,1163,1166],{"className":1159},[82],[32,1161],{"className":1162,"style":87},[86],[32,1164,363],{"className":1165,"style":740},[95,96],[32,1167,1084],{"className":1168},[95],"，这才有了余弦相似度、点积注意力这些操作。度量结构的信息量最大，因为它不仅告诉你「是否有关系」，还告诉你「关系有多强」——但代价是它对表示方式极度敏感：换一个嵌入模型、换一个归一化方式，同一批数据算出来的相似度可以完全不同。",[21,1171,1173],{"id":1172},"拓扑连通结构","拓扑（连通结构）",[11,1175,1176],{},"拓扑关注的是「连通」。把度量忘掉，只问是否有路径可达、结构是否连通、是否存在环\u002F洞。在同胚变换（连续拉伸、不撕裂）下不变的东西才算拓扑性质。典型例子：社交网络里的连通分量、图的最短路径存在性（而非具体长度）。拓扑关系比几何轻，研究的是关系的骨架，但换来的是更强的稳健性——不在乎具体尺度，只在乎结构骨架。",[11,1178,1179,1180,1234,1235,1265,1266,1294,1295,1362,1363,1367,1368,1444,1445,1473,1474,1578],{},"一个拓扑空间是二元组 ",[32,1181,1183,1206],{"className":1182,"translate":36},[35],[32,1184,1186],{"className":1185},[40],[42,1187,1188],{"xmlns":44},[46,1189,1190,1203],{},[49,1191,1192,1194,1196,1198,1201],{},[52,1193,55],{"stretchy":54},[57,1195,59],{},[52,1197,63],{"separator":62},[57,1199,1200],{},"τ",[52,1202,69],{"stretchy":54},[71,1204,1205],{"encoding":73},"(X, \\tau)",[32,1207,1209],{"className":1208,"ariaHidden":62},[78],[32,1210,1212,1215,1218,1221,1224,1227,1231],{"className":1211},[82],[32,1213],{"className":1214,"style":87},[86],[32,1216,55],{"className":1217},[91],[32,1219,59],{"className":1220,"style":97},[95,96],[32,1222,63],{"className":1223},[101],[32,1225],{"className":1226,"style":106},[105],[32,1228,1200],{"className":1229,"style":1230},[95,96],"margin-right:0.1132em;",[32,1232,69],{"className":1233},[113],"，",[32,1236,1238,1252],{"className":1237,"translate":36},[35],[32,1239,1241],{"className":1240},[40],[42,1242,1243],{"xmlns":44},[46,1244,1245,1249],{},[49,1246,1247],{},[57,1248,1200],{},[71,1250,1251],{"encoding":73},"\\tau",[32,1253,1255],{"className":1254,"ariaHidden":62},[78],[32,1256,1258,1262],{"className":1257},[82],[32,1259],{"className":1260,"style":1261},[86],"height:0.4306em;",[32,1263,1200],{"className":1264,"style":1230},[95,96]," 是 ",[32,1267,1269,1282],{"className":1268,"translate":36},[35],[32,1270,1272],{"className":1271},[40],[42,1273,1274],{"xmlns":44},[46,1275,1276,1280],{},[49,1277,1278],{},[57,1279,59],{},[71,1281,59],{"encoding":73},[32,1283,1285],{"className":1284,"ariaHidden":62},[78],[32,1286,1288,1291],{"className":1287},[82],[32,1289],{"className":1290,"style":212},[86],[32,1292,59],{"className":1293,"style":97},[95,96]," 的一族「开集」，满足对任意并、有限交封闭，且 ",[32,1296,1298,1322],{"className":1297,"translate":36},[35],[32,1299,1301],{"className":1300},[40],[42,1302,1303],{"xmlns":44},[46,1304,1305,1319],{},[49,1306,1307,1310,1312,1314,1317],{},[57,1308,1309],{"mathvariant":1083},"∅",[52,1311,63],{"separator":62},[57,1313,59],{},[52,1315,1316],{},"∈",[57,1318,1200],{},[71,1320,1321],{"encoding":73},"\\emptyset, X \\in \\tau",[32,1323,1325,1353],{"className":1324,"ariaHidden":62},[78],[32,1326,1328,1332,1335,1338,1341,1344,1347,1350],{"className":1327},[82],[32,1329],{"className":1330,"style":1331},[86],"height:0.9444em;vertical-align:-0.1944em;",[32,1333,1309],{"className":1334},[95],[32,1336,63],{"className":1337},[101],[32,1339],{"className":1340,"style":106},[105],[32,1342,59],{"className":1343,"style":97},[95,96],[32,1345],{"className":1346,"style":177},[105],[32,1348,1316],{"className":1349},[181],[32,1351],{"className":1352,"style":177},[105],[32,1354,1356,1359],{"className":1355},[82],[32,1357],{"className":1358,"style":1261},[86],[32,1360,1200],{"className":1361,"style":1230},[95,96],"。拓扑结构不需要度量：任何度量空间都能诱导出一个拓扑（用开球生成开集），但反过来不成立——拓扑空间不一定可度量化。关键的不变量是在",[1364,1365,1366],"strong",{},"同胚","（连续双射且逆映射也连续）下保持不变的性质，比如连通性、紧致性、亏格（洞的个数）。用代数拓扑的语言，贝蒂数 ",[32,1369,1371,1391],{"className":1370,"translate":36},[35],[32,1372,1374],{"className":1373},[40],[42,1375,1376],{"xmlns":44},[46,1377,1378,1388],{},[49,1379,1380],{},[143,1381,1382,1385],{},[57,1383,1384],{},"b",[57,1386,1387],{},"k",[71,1389,1390],{"encoding":73},"b_k",[32,1392,1394],{"className":1393,"ariaHidden":62},[78],[32,1395,1397,1401],{"className":1396},[82],[32,1398],{"className":1399,"style":1400},[86],"height:0.8444em;vertical-align:-0.15em;",[32,1402,1404,1407],{"className":1403},[95],[32,1405,1384],{"className":1406},[95,96],[32,1408,1410],{"className":1409},[242],[32,1411,1413,1435],{"className":1412},[246,247],[32,1414,1416,1432],{"className":1415},[251],[32,1417,1420],{"className":1418,"style":1419},[255],"height:0.3361em;",[32,1421,1422,1425],{"style":259},[32,1423],{"className":1424,"style":264},[263],[32,1426,1428],{"className":1427},[268,269,270,271],[32,1429,1387],{"className":1430,"style":1431},[95,96,271],"margin-right:0.0315em;",[32,1433,285],{"className":1434},[284],[32,1436,1438],{"className":1437},[251],[32,1439,1442],{"className":1440,"style":1441},[255],"height:0.15em;",[32,1443],{}," 数的是 ",[32,1446,1448,1461],{"className":1447,"translate":36},[35],[32,1449,1451],{"className":1450},[40],[42,1452,1453],{"xmlns":44},[46,1454,1455,1459],{},[49,1456,1457],{},[57,1458,1387],{},[71,1460,1387],{"encoding":73},[32,1462,1464],{"className":1463,"ariaHidden":62},[78],[32,1465,1467,1470],{"className":1466},[82],[32,1468],{"className":1469,"style":170},[86],[32,1471,1387],{"className":1472,"style":1431},[95,96]," 维「洞」的个数，这是持久同调（persistent homology）在数据分析里常用的工具——它告诉你数据云的「形状骨架」，而完全不关心具体的尺度参数。这解释了为什么拓扑比几何「轻」：它是几何结构经过遗忘函子 ",[32,1475,1477,1518],{"className":1476,"translate":36},[35],[32,1478,1480],{"className":1479},[40],[42,1481,1482],{"xmlns":44},[46,1483,1484,1515],{},[49,1485,1486,1489,1491,1503,1505],{},[57,1487,1488],{},"U",[52,1490,131],{},[49,1492,1493,1497,1500],{},[57,1494,1496],{"mathvariant":1495},"bold","M",[57,1498,1499],{"mathvariant":1495},"e",[57,1501,1502],{"mathvariant":1495},"t",[52,1504,141],{},[49,1506,1507,1510,1513],{},[57,1508,1509],{"mathvariant":1495},"T",[57,1511,1512],{"mathvariant":1495},"o",[57,1514,11],{"mathvariant":1495},[71,1516,1517],{"encoding":73},"U: \\mathbf{Met} \\to \\mathbf{Top}",[32,1519,1521,1540,1564],{"className":1520,"ariaHidden":62},[78],[32,1522,1524,1527,1531,1534,1537],{"className":1523},[82],[32,1525],{"className":1526,"style":212},[86],[32,1528,1488],{"className":1529,"style":1530},[95,96],"margin-right:0.109em;",[32,1532],{"className":1533,"style":177},[105],[32,1535,131],{"className":1536},[181],[32,1538],{"className":1539,"style":177},[105],[32,1541,1543,1547,1555,1558,1561],{"className":1542},[82],[32,1544],{"className":1545,"style":1546},[86],"height:0.6861em;",[32,1548,1550],{"className":1549},[95],[32,1551,1554],{"className":1552},[95,1553],"mathbf","Met",[32,1556],{"className":1557,"style":177},[105],[32,1559,141],{"className":1560},[181],[32,1562],{"className":1563,"style":177},[105],[32,1565,1567,1571],{"className":1566},[82],[32,1568],{"className":1569,"style":1570},[86],"height:0.8805em;vertical-align:-0.1944em;",[32,1572,1574],{"className":1573},[95],[32,1575,1577],{"className":1576},[95,1553],"Top"," 投影之后剩下的东西。",[21,1580,1582],{"id":1581},"因果约束结构","因果（约束结构）",[11,1584,1585],{},"因果关注的是「约束」。这一支和前两支有本质区别：几何和拓扑通常是对称的（A 到 B 的距离=B 到 A 的距离；A 连着 B 等价于 B 连着 A），而因果关系本质上是非对称的——「X 决定\u002F限制 Y 的取值范围」不等于反过来成立。因果结构需要额外的信息才能识别（干预、反事实、时间先后），单纯的相关性数据（哪怕是几何或拓扑意义上完整的）不足以确定因果方向。这也是 Pearl 那套 do-calculus 要解决的问题。在 ML 中，几何\u002F拓扑模型在独立同分布（i.i.d.）下表现很好，但只要测试分布变了（协变量偏移），它们就会崩。而因果结构（比如有向无环图 DAG）告诉你的是「机制」——只要干预机制不变，哪怕输入分布变了，模型依然能泛化。这也是为什么说因果是稳健性的终极来源。",[11,1587,1588,1589,114,1679,1707,1708,1886,1887,2062],{},"结构因果模型（SCM）是一个三元组 ",[32,1590,1592,1624],{"className":1591,"translate":36},[35],[32,1593,1595],{"className":1594},[40],[42,1596,1597],{"xmlns":44},[46,1598,1599,1621],{},[49,1600,1601,1603,1605,1607,1609,1611,1614,1616,1619],{},[57,1602,1496],{},[52,1604,368],{},[52,1606,998],{"stretchy":54},[57,1608,1488],{},[52,1610,63],{"separator":62},[57,1612,1613],{},"V",[52,1615,63],{"separator":62},[57,1617,1618],{},"F",[52,1620,1008],{"stretchy":54},[71,1622,1623],{"encoding":73},"M = \\langle U, V, F \\rangle",[32,1625,1627,1645],{"className":1626,"ariaHidden":62},[78],[32,1628,1630,1633,1636,1639,1642],{"className":1629},[82],[32,1631],{"className":1632,"style":212},[86],[32,1634,1496],{"className":1635,"style":1530},[95,96],[32,1637],{"className":1638,"style":177},[105],[32,1640,368],{"className":1641},[181],[32,1643],{"className":1644,"style":177},[105],[32,1646,1648,1651,1654,1657,1660,1663,1666,1669,1672,1676],{"className":1647},[82],[32,1649],{"className":1650,"style":87},[86],[32,1652,998],{"className":1653},[91],[32,1655,1488],{"className":1656,"style":1530},[95,96],[32,1658,63],{"className":1659},[101],[32,1661],{"className":1662,"style":106},[105],[32,1664,1613],{"className":1665,"style":198},[95,96],[32,1667,63],{"className":1668},[101],[32,1670],{"className":1671,"style":106},[105],[32,1673,1618],{"className":1674,"style":1675},[95,96],"margin-right:0.1389em;",[32,1677,1008],{"className":1678},[113],[32,1680,1682,1695],{"className":1681,"translate":36},[35],[32,1683,1685],{"className":1684},[40],[42,1686,1687],{"xmlns":44},[46,1688,1689,1693],{},[49,1690,1691],{},[57,1692,1488],{},[71,1694,1488],{"encoding":73},[32,1696,1698],{"className":1697,"ariaHidden":62},[78],[32,1699,1701,1704],{"className":1700},[82],[32,1702],{"className":1703,"style":212},[86],[32,1705,1488],{"className":1706,"style":1530},[95,96]," 是外生变量，",[32,1709,1711,1753],{"className":1710,"translate":36},[35],[32,1712,1714],{"className":1713},[40],[42,1715,1716],{"xmlns":44},[46,1717,1718,1750],{},[49,1719,1720,1722,1724,1726,1733,1735,1738,1740,1747],{},[57,1721,1613],{},[52,1723,368],{},[52,1725,315],{"stretchy":54},[143,1727,1728,1730],{},[57,1729,1613],{},[155,1731,1732],{},"1",[52,1734,63],{"separator":62},[52,1736,1737],{},"…",[52,1739,63],{"separator":62},[143,1741,1742,1744],{},[57,1743,1613],{},[57,1745,1746],{},"n",[52,1748,1749],{"stretchy":54},"}",[71,1751,1752],{"encoding":73},"V = \\{V_1, \\dots, V_n\\}",[32,1754,1756,1774],{"className":1755,"ariaHidden":62},[78],[32,1757,1759,1762,1765,1768,1771],{"className":1758},[82],[32,1760],{"className":1761,"style":212},[86],[32,1763,1613],{"className":1764,"style":198},[95,96],[32,1766],{"className":1767,"style":177},[105],[32,1769,368],{"className":1770},[181],[32,1772],{"className":1773,"style":177},[105],[32,1775,1777,1780,1783,1824,1827,1830,1833,1836,1839,1842,1883],{"className":1776},[82],[32,1778],{"className":1779,"style":87},[86],[32,1781,315],{"className":1782},[91],[32,1784,1786,1789],{"className":1785},[95],[32,1787,1613],{"className":1788,"style":198},[95,96],[32,1790,1792],{"className":1791},[242],[32,1793,1795,1816],{"className":1794},[246,247],[32,1796,1798,1813],{"className":1797},[251],[32,1799,1801],{"className":1800,"style":256},[255],[32,1802,1804,1807],{"style":1803},"top:-2.55em;margin-left:-0.2222em;margin-right:0.05em;",[32,1805],{"className":1806,"style":264},[263],[32,1808,1810],{"className":1809},[268,269,270,271],[32,1811,1732],{"className":1812},[95,271],[32,1814,285],{"className":1815},[284],[32,1817,1819],{"className":1818},[251],[32,1820,1822],{"className":1821,"style":1441},[255],[32,1823],{},[32,1825,63],{"className":1826},[101],[32,1828],{"className":1829,"style":106},[105],[32,1831,1737],{"className":1832},[507],[32,1834],{"className":1835,"style":106},[105],[32,1837,63],{"className":1838},[101],[32,1840],{"className":1841,"style":106},[105],[32,1843,1845,1848],{"className":1844},[95],[32,1846,1613],{"className":1847,"style":198},[95,96],[32,1849,1851],{"className":1850},[242],[32,1852,1854,1875],{"className":1853},[246,247],[32,1855,1857,1872],{"className":1856},[251],[32,1858,1861],{"className":1859,"style":1860},[255],"height:0.1514em;",[32,1862,1863,1866],{"style":1803},[32,1864],{"className":1865,"style":264},[263],[32,1867,1869],{"className":1868},[268,269,270,271],[32,1870,1746],{"className":1871},[95,96,271],[32,1873,285],{"className":1874},[284],[32,1876,1878],{"className":1877},[251],[32,1879,1881],{"className":1880,"style":1441},[255],[32,1882],{},[32,1884,1749],{"className":1885},[113]," 是内生变量，",[32,1888,1890,1929],{"className":1889,"translate":36},[35],[32,1891,1893],{"className":1892},[40],[42,1894,1895],{"xmlns":44},[46,1896,1897,1926],{},[49,1898,1899,1901,1903,1905,1912,1914,1916,1918,1924],{},[57,1900,1618],{},[52,1902,368],{},[52,1904,315],{"stretchy":54},[143,1906,1907,1910],{},[57,1908,1909],{},"f",[155,1911,1732],{},[52,1913,63],{"separator":62},[52,1915,1737],{},[52,1917,63],{"separator":62},[143,1919,1920,1922],{},[57,1921,1909],{},[57,1923,1746],{},[52,1925,1749],{"stretchy":54},[71,1927,1928],{"encoding":73},"F = \\{f_1, \\dots, f_n\\}",[32,1930,1932,1950],{"className":1931,"ariaHidden":62},[78],[32,1933,1935,1938,1941,1944,1947],{"className":1934},[82],[32,1936],{"className":1937,"style":212},[86],[32,1939,1618],{"className":1940,"style":1675},[95,96],[32,1942],{"className":1943,"style":177},[105],[32,1945,368],{"className":1946},[181],[32,1948],{"className":1949,"style":177},[105],[32,1951,1953,1956,1959,2001,2004,2007,2010,2013,2016,2019,2059],{"className":1952},[82],[32,1954],{"className":1955,"style":87},[86],[32,1957,315],{"className":1958},[91],[32,1960,1962,1966],{"className":1961},[95],[32,1963,1909],{"className":1964,"style":1965},[95,96],"margin-right:0.1076em;",[32,1967,1969],{"className":1968},[242],[32,1970,1972,1993],{"className":1971},[246,247],[32,1973,1975,1990],{"className":1974},[251],[32,1976,1978],{"className":1977,"style":256},[255],[32,1979,1981,1984],{"style":1980},"top:-2.55em;margin-left:-0.1076em;margin-right:0.05em;",[32,1982],{"className":1983,"style":264},[263],[32,1985,1987],{"className":1986},[268,269,270,271],[32,1988,1732],{"className":1989},[95,271],[32,1991,285],{"className":1992},[284],[32,1994,1996],{"className":1995},[251],[32,1997,1999],{"className":1998,"style":1441},[255],[32,2000],{},[32,2002,63],{"className":2003},[101],[32,2005],{"className":2006,"style":106},[105],[32,2008,1737],{"className":2009},[507],[32,2011],{"className":2012,"style":106},[105],[32,2014,63],{"className":2015},[101],[32,2017],{"className":2018,"style":106},[105],[32,2020,2022,2025],{"className":2021},[95],[32,2023,1909],{"className":2024,"style":1965},[95,96],[32,2026,2028],{"className":2027},[242],[32,2029,2031,2051],{"className":2030},[246,247],[32,2032,2034,2048],{"className":2033},[251],[32,2035,2037],{"className":2036,"style":1860},[255],[32,2038,2039,2042],{"style":1980},[32,2040],{"className":2041,"style":264},[263],[32,2043,2045],{"className":2044},[268,269,270,271],[32,2046,1746],{"className":2047},[95,96,271],[32,2049,285],{"className":2050},[284],[32,2052,2054],{"className":2053},[251],[32,2055,2057],{"className":2056,"style":1441},[255],[32,2058],{},[32,2060,1749],{"className":2061},[113]," 是一组函数，满足",[32,2064,2066],{"className":2065,"translate":36},[299],[32,2067,2069,2128],{"className":2068,"translate":36},[35],[32,2070,2072],{"className":2071},[40],[42,2073,2074],{"xmlns":44,"display":308},[46,2075,2076,2125],{},[49,2077,2078,2085,2087,2089,2095,2097,2105,2107,2113,2115,2117,2123],{},[143,2079,2080,2082],{},[57,2081,1613],{},[57,2083,2084],{},"i",[52,2086,131],{},[52,2088,368],{},[143,2090,2091,2093],{},[57,2092,1909],{},[57,2094,2084],{},[52,2096,55],{"stretchy":54},[49,2098,2099,2102],{},[57,2100,2101],{"mathvariant":1083},"P",[57,2103,2104],{"mathvariant":1083},"a",[52,2106,55],{"stretchy":54},[143,2108,2109,2111],{},[57,2110,1613],{},[57,2112,2084],{},[52,2114,69],{"stretchy":54},[52,2116,63],{"separator":62},[143,2118,2119,2121],{},[57,2120,1488],{},[57,2122,2084],{},[52,2124,69],{"stretchy":54},[71,2126,2127],{"encoding":73},"V_i := f_i(\\mathrm{Pa}(V_i), U_i)",[32,2129,2131,2189],{"className":2130,"ariaHidden":62},[78],[32,2132,2134,2138,2179,2182,2186],{"className":2133},[82],[32,2135],{"className":2136,"style":2137},[86],"height:0.8333em;vertical-align:-0.15em;",[32,2139,2141,2144],{"className":2140},[95],[32,2142,1613],{"className":2143,"style":198},[95,96],[32,2145,2147],{"className":2146},[242],[32,2148,2150,2171],{"className":2149},[246,247],[32,2151,2153,2168],{"className":2152},[251],[32,2154,2157],{"className":2155,"style":2156},[255],"height:0.3117em;",[32,2158,2159,2162],{"style":1803},[32,2160],{"className":2161,"style":264},[263],[32,2163,2165],{"className":2164},[268,269,270,271],[32,2166,2084],{"className":2167},[95,96,271],[32,2169,285],{"className":2170},[284],[32,2172,2174],{"className":2173},[251],[32,2175,2177],{"className":2176,"style":1441},[255],[32,2178],{},[32,2180],{"className":2181,"style":177},[105],[32,2183,2185],{"className":2184},[181],":=",[32,2187],{"className":2188,"style":177},[105],[32,2190,2192,2195,2235,2238,2246,2249,2289,2292,2295,2298,2339],{"className":2191},[82],[32,2193],{"className":2194,"style":87},[86],[32,2196,2198,2201],{"className":2197},[95],[32,2199,1909],{"className":2200,"style":1965},[95,96],[32,2202,2204],{"className":2203},[242],[32,2205,2207,2227],{"className":2206},[246,247],[32,2208,2210,2224],{"className":2209},[251],[32,2211,2213],{"className":2212,"style":2156},[255],[32,2214,2215,2218],{"style":1980},[32,2216],{"className":2217,"style":264},[263],[32,2219,2221],{"className":2220},[268,269,270,271],[32,2222,2084],{"className":2223},[95,96,271],[32,2225,285],{"className":2226},[284],[32,2228,2230],{"className":2229},[251],[32,2231,2233],{"className":2232,"style":1441},[255],[32,2234],{},[32,2236,55],{"className":2237},[91],[32,2239,2241],{"className":2240},[95],[32,2242,2245],{"className":2243},[95,2244],"mathrm","Pa",[32,2247,55],{"className":2248},[91],[32,2250,2252,2255],{"className":2251},[95],[32,2253,1613],{"className":2254,"style":198},[95,96],[32,2256,2258],{"className":2257},[242],[32,2259,2261,2281],{"className":2260},[246,247],[32,2262,2264,2278],{"className":2263},[251],[32,2265,2267],{"className":2266,"style":2156},[255],[32,2268,2269,2272],{"style":1803},[32,2270],{"className":2271,"style":264},[263],[32,2273,2275],{"className":2274},[268,269,270,271],[32,2276,2084],{"className":2277},[95,96,271],[32,2279,285],{"className":2280},[284],[32,2282,2284],{"className":2283},[251],[32,2285,2287],{"className":2286,"style":1441},[255],[32,2288],{},[32,2290,69],{"className":2291},[113],[32,2293,63],{"className":2294},[101],[32,2296],{"className":2297,"style":106},[105],[32,2299,2301,2304],{"className":2300},[95],[32,2302,1488],{"className":2303,"style":1530},[95,96],[32,2305,2307],{"className":2306},[242],[32,2308,2310,2331],{"className":2309},[246,247],[32,2311,2313,2328],{"className":2312},[251],[32,2314,2316],{"className":2315,"style":2156},[255],[32,2317,2319,2322],{"style":2318},"top:-2.55em;margin-left:-0.109em;margin-right:0.05em;",[32,2320],{"className":2321,"style":264},[263],[32,2323,2325],{"className":2324},[268,269,270,271],[32,2326,2084],{"className":2327},[95,96,271],[32,2329,285],{"className":2330},[284],[32,2332,2334],{"className":2333},[251],[32,2335,2337],{"className":2336,"style":1441},[255],[32,2338],{},[32,2340,69],{"className":2341},[113],[11,2343,2344,2345,2375,2376,2405,2406,2409,2410,2439,2440,2577,2578,2733,2734,2786,2787,2897,2898,2926,2927,2955,2956,2984,2985,3073,3074,3125,3126,3154],{},"赋值符号 ",[32,2346,2348,2363],{"className":2347,"translate":36},[35],[32,2349,2351],{"className":2350},[40],[42,2352,2353],{"xmlns":44},[46,2354,2355,2361],{},[49,2356,2357,2359],{},[52,2358,131],{},[52,2360,368],{},[71,2362,2185],{"encoding":73},[32,2364,2366],{"className":2365,"ariaHidden":62},[78],[32,2367,2369,2372],{"className":2368},[82],[32,2370],{"className":2371,"style":1261},[86],[32,2373,2185],{"className":2374},[181],"（而非 ",[32,2377,2379,2392],{"className":2378,"translate":36},[35],[32,2380,2382],{"className":2381},[40],[42,2383,2384],{"xmlns":44},[46,2385,2386,2390],{},[49,2387,2388],{},[52,2389,368],{},[71,2391,368],{"encoding":73},[32,2393,2395],{"className":2394,"ariaHidden":62},[78],[32,2396,2398,2402],{"className":2397},[82],[32,2399],{"className":2400,"style":2401},[86],"height:0.3669em;",[32,2403,368],{"className":2404},[181],"）是关键——它表示这是一个",[1364,2407,2408],{},"机制","，不是一个可逆的代数等式。这组机制天然诱导一张有向无环图（DAG）",[32,2411,2413,2427],{"className":2412,"translate":36},[35],[32,2414,2416],{"className":2415},[40],[42,2417,2418],{"xmlns":44},[46,2419,2420,2425],{},[49,2421,2422],{},[57,2423,2424],{},"G",[71,2426,2424],{"encoding":73},[32,2428,2430],{"className":2429,"ariaHidden":62},[78],[32,2431,2433,2436],{"className":2432},[82],[32,2434],{"className":2435,"style":212},[86],[32,2437,2424],{"className":2438},[95,96],"，边 ",[32,2441,2443,2470],{"className":2442,"translate":36},[35],[32,2444,2446],{"className":2445},[40],[42,2447,2448],{"xmlns":44},[46,2449,2450,2467],{},[49,2451,2452,2459,2461],{},[143,2453,2454,2456],{},[57,2455,1613],{},[57,2457,2458],{},"j",[52,2460,141],{},[143,2462,2463,2465],{},[57,2464,1613],{},[57,2466,2084],{},[71,2468,2469],{"encoding":73},"V_j \\to V_i",[32,2471,2473,2531],{"className":2472,"ariaHidden":62},[78],[32,2474,2476,2480,2522,2525,2528],{"className":2475},[82],[32,2477],{"className":2478,"style":2479},[86],"height:0.9694em;vertical-align:-0.2861em;",[32,2481,2483,2486],{"className":2482},[95],[32,2484,1613],{"className":2485,"style":198},[95,96],[32,2487,2489],{"className":2488},[242],[32,2490,2492,2513],{"className":2491},[246,247],[32,2493,2495,2510],{"className":2494},[251],[32,2496,2498],{"className":2497,"style":2156},[255],[32,2499,2500,2503],{"style":1803},[32,2501],{"className":2502,"style":264},[263],[32,2504,2506],{"className":2505},[268,269,270,271],[32,2507,2458],{"className":2508,"style":2509},[95,96,271],"margin-right:0.0572em;",[32,2511,285],{"className":2512},[284],[32,2514,2516],{"className":2515},[251],[32,2517,2520],{"className":2518,"style":2519},[255],"height:0.2861em;",[32,2521],{},[32,2523],{"className":2524,"style":177},[105],[32,2526,141],{"className":2527},[181],[32,2529],{"className":2530,"style":177},[105],[32,2532,2534,2537],{"className":2533},[82],[32,2535],{"className":2536,"style":2137},[86],[32,2538,2540,2543],{"className":2539},[95],[32,2541,1613],{"className":2542,"style":198},[95,96],[32,2544,2546],{"className":2545},[242],[32,2547,2549,2569],{"className":2548},[246,247],[32,2550,2552,2566],{"className":2551},[251],[32,2553,2555],{"className":2554,"style":2156},[255],[32,2556,2557,2560],{"style":1803},[32,2558],{"className":2559,"style":264},[263],[32,2561,2563],{"className":2562},[268,269,270,271],[32,2564,2084],{"className":2565},[95,96,271],[32,2567,285],{"className":2568},[284],[32,2570,2572],{"className":2571},[251],[32,2573,2575],{"className":2574,"style":1441},[255],[32,2576],{}," 存在当且仅当 ",[32,2579,2581,2617],{"className":2580,"translate":36},[35],[32,2582,2584],{"className":2583},[40],[42,2585,2586],{"xmlns":44},[46,2587,2588,2614],{},[49,2589,2590,2596,2598,2604,2606,2612],{},[143,2591,2592,2594],{},[57,2593,1613],{},[57,2595,2458],{},[52,2597,1316],{},[49,2599,2600,2602],{},[57,2601,2101],{"mathvariant":1083},[57,2603,2104],{"mathvariant":1083},[52,2605,55],{"stretchy":54},[143,2607,2608,2610],{},[57,2609,1613],{},[57,2611,2084],{},[52,2613,69],{"stretchy":54},[71,2615,2616],{"encoding":73},"V_j \\in \\mathrm{Pa}(V_i)",[32,2618,2620,2675],{"className":2619,"ariaHidden":62},[78],[32,2621,2623,2626,2666,2669,2672],{"className":2622},[82],[32,2624],{"className":2625,"style":2479},[86],[32,2627,2629,2632],{"className":2628},[95],[32,2630,1613],{"className":2631,"style":198},[95,96],[32,2633,2635],{"className":2634},[242],[32,2636,2638,2658],{"className":2637},[246,247],[32,2639,2641,2655],{"className":2640},[251],[32,2642,2644],{"className":2643,"style":2156},[255],[32,2645,2646,2649],{"style":1803},[32,2647],{"className":2648,"style":264},[263],[32,2650,2652],{"className":2651},[268,269,270,271],[32,2653,2458],{"className":2654,"style":2509},[95,96,271],[32,2656,285],{"className":2657},[284],[32,2659,2661],{"className":2660},[251],[32,2662,2664],{"className":2663,"style":2519},[255],[32,2665],{},[32,2667],{"className":2668,"style":177},[105],[32,2670,1316],{"className":2671},[181],[32,2673],{"className":2674,"style":177},[105],[32,2676,2678,2681,2687,2690,2730],{"className":2677},[82],[32,2679],{"className":2680,"style":87},[86],[32,2682,2684],{"className":2683},[95],[32,2685,2245],{"className":2686},[95,2244],[32,2688,55],{"className":2689},[91],[32,2691,2693,2696],{"className":2692},[95],[32,2694,1613],{"className":2695,"style":198},[95,96],[32,2697,2699],{"className":2698},[242],[32,2700,2702,2722],{"className":2701},[246,247],[32,2703,2705,2719],{"className":2704},[251],[32,2706,2708],{"className":2707,"style":2156},[255],[32,2709,2710,2713],{"style":1803},[32,2711],{"className":2712,"style":264},[263],[32,2714,2716],{"className":2715},[268,269,270,271],[32,2717,2084],{"className":2718},[95,96,271],[32,2720,285],{"className":2721},[284],[32,2723,2725],{"className":2724},[251],[32,2726,2728],{"className":2727,"style":1441},[255],[32,2729],{},[32,2731,69],{"className":2732},[113],"。因果推断的核心操作是 ",[32,2735,2737,2761],{"className":2736,"translate":36},[35],[32,2738,2740],{"className":2739},[40],[42,2741,2742],{"xmlns":44},[46,2743,2744,2758],{},[49,2745,2746,2752,2754,2756],{},[49,2747,2748,2750],{},[57,2749,66],{"mathvariant":1083},[57,2751,1512],{"mathvariant":1083},[52,2753,55],{"stretchy":54},[52,2755,1001],{},[52,2757,69],{"stretchy":54},[71,2759,2760],{"encoding":73},"\\mathrm{do}(\\cdot)",[32,2762,2764],{"className":2763,"ariaHidden":62},[78],[32,2765,2767,2770,2777,2780,2783],{"className":2766},[82],[32,2768],{"className":2769,"style":87},[86],[32,2771,2773],{"className":2772},[95],[32,2774,2776],{"className":2775},[95,2244],"do",[32,2778,55],{"className":2779},[91],[32,2781,1001],{"className":2782},[95],[32,2784,69],{"className":2785},[113]," 算子：",[32,2788,2790,2830],{"className":2789,"translate":36},[35],[32,2791,2793],{"className":2792},[40],[42,2794,2795],{"xmlns":44},[46,2796,2797,2827],{},[49,2798,2799,2801,2803,2806,2809,2815,2817,2819,2821,2823,2825],{},[57,2800,2101],{},[52,2802,55],{"stretchy":54},[57,2804,2805],{},"Y",[52,2807,2808],{},"∣",[49,2810,2811,2813],{},[57,2812,66],{"mathvariant":1083},[57,2814,1512],{"mathvariant":1083},[52,2816,55],{"stretchy":54},[57,2818,59],{},[52,2820,368],{},[57,2822,358],{},[52,2824,69],{"stretchy":54},[52,2826,69],{"stretchy":54},[71,2828,2829],{"encoding":73},"P(Y \\mid \\mathrm{do}(X=x))",[32,2831,2833,2857,2884],{"className":2832,"ariaHidden":62},[78],[32,2834,2836,2839,2842,2845,2848,2851,2854],{"className":2835},[82],[32,2837],{"className":2838,"style":87},[86],[32,2840,2101],{"className":2841,"style":1675},[95,96],[32,2843,55],{"className":2844},[91],[32,2846,2805],{"className":2847,"style":198},[95,96],[32,2849],{"className":2850,"style":177},[105],[32,2852,2808],{"className":2853},[181],[32,2855],{"className":2856,"style":177},[105],[32,2858,2860,2863,2869,2872,2875,2878,2881],{"className":2859},[82],[32,2861],{"className":2862,"style":87},[86],[32,2864,2866],{"className":2865},[95],[32,2867,2776],{"className":2868},[95,2244],[32,2870,55],{"className":2871},[91],[32,2873,59],{"className":2874,"style":97},[95,96],[32,2876],{"className":2877,"style":177},[105],[32,2879,368],{"className":2880},[181],[32,2882],{"className":2883,"style":177},[105],[32,2885,2887,2890,2893],{"className":2886},[82],[32,2888],{"className":2889,"style":87},[86],[32,2891,358],{"className":2892},[95,96],[32,2894,2896],{"className":2895},[113],"))"," 表示「把 ",[32,2899,2901,2914],{"className":2900,"translate":36},[35],[32,2902,2904],{"className":2903},[40],[42,2905,2906],{"xmlns":44},[46,2907,2908,2912],{},[49,2909,2910],{},[57,2911,59],{},[71,2913,59],{"encoding":73},[32,2915,2917],{"className":2916,"ariaHidden":62},[78],[32,2918,2920,2923],{"className":2919},[82],[32,2921],{"className":2922,"style":212},[86],[32,2924,59],{"className":2925,"style":97},[95,96]," 的赋值机制替换为常数 ",[32,2928,2930,2943],{"className":2929,"translate":36},[35],[32,2931,2933],{"className":2932},[40],[42,2934,2935],{"xmlns":44},[46,2936,2937,2941],{},[49,2938,2939],{},[57,2940,358],{},[71,2942,358],{"encoding":73},[32,2944,2946],{"className":2945,"ariaHidden":62},[78],[32,2947,2949,2952],{"className":2948},[82],[32,2950],{"className":2951,"style":1261},[86],[32,2953,358],{"className":2954},[95,96],"，切断它与所有父节点的连接」之后 ",[32,2957,2959,2972],{"className":2958,"translate":36},[35],[32,2960,2962],{"className":2961},[40],[42,2963,2964],{"xmlns":44},[46,2965,2966,2970],{},[49,2967,2968],{},[57,2969,2805],{},[71,2971,2805],{"encoding":73},[32,2973,2975],{"className":2974,"ariaHidden":62},[78],[32,2976,2978,2981],{"className":2977},[82],[32,2979],{"className":2980,"style":212},[86],[32,2982,2805],{"className":2983,"style":198},[95,96]," 的分布，这与条件概率 ",[32,2986,2988,3016],{"className":2987,"translate":36},[35],[32,2989,2991],{"className":2990},[40],[42,2992,2993],{"xmlns":44},[46,2994,2995,3013],{},[49,2996,2997,2999,3001,3003,3005,3007,3009,3011],{},[57,2998,2101],{},[52,3000,55],{"stretchy":54},[57,3002,2805],{},[52,3004,2808],{},[57,3006,59],{},[52,3008,368],{},[57,3010,358],{},[52,3012,69],{"stretchy":54},[71,3014,3015],{"encoding":73},"P(Y \\mid X=x)",[32,3017,3019,3043,3061],{"className":3018,"ariaHidden":62},[78],[32,3020,3022,3025,3028,3031,3034,3037,3040],{"className":3021},[82],[32,3023],{"className":3024,"style":87},[86],[32,3026,2101],{"className":3027,"style":1675},[95,96],[32,3029,55],{"className":3030},[91],[32,3032,2805],{"className":3033,"style":198},[95,96],[32,3035],{"className":3036,"style":177},[105],[32,3038,2808],{"className":3039},[181],[32,3041],{"className":3042,"style":177},[105],[32,3044,3046,3049,3052,3055,3058],{"className":3045},[82],[32,3047],{"className":3048,"style":212},[86],[32,3050,59],{"className":3051,"style":97},[95,96],[32,3053],{"className":3054,"style":177},[105],[32,3056,368],{"className":3057},[181],[32,3059],{"className":3060,"style":177},[105],[32,3062,3064,3067,3070],{"className":3063},[82],[32,3065],{"className":3066,"style":87},[86],[32,3068,358],{"className":3069},[95,96],[32,3071,69],{"className":3072},[113],"——「观察到 ",[32,3075,3077,3095],{"className":3076,"translate":36},[35],[32,3078,3080],{"className":3079},[40],[42,3081,3082],{"xmlns":44},[46,3083,3084,3092],{},[49,3085,3086,3088,3090],{},[57,3087,59],{},[52,3089,368],{},[57,3091,358],{},[71,3093,3094],{"encoding":73},"X=x",[32,3096,3098,3116],{"className":3097,"ariaHidden":62},[78],[32,3099,3101,3104,3107,3110,3113],{"className":3100},[82],[32,3102],{"className":3103,"style":212},[86],[32,3105,59],{"className":3106,"style":97},[95,96],[32,3108],{"className":3109,"style":177},[105],[32,3111,368],{"className":3112},[181],[32,3114],{"className":3115,"style":177},[105],[32,3117,3119,3122],{"className":3118},[82],[32,3120],{"className":3121,"style":1261},[86],[32,3123,358],{"className":3124},[95,96]," 之后 ",[32,3127,3129,3142],{"className":3128,"translate":36},[35],[32,3130,3132],{"className":3131},[40],[42,3133,3134],{"xmlns":44},[46,3135,3136,3140],{},[49,3137,3138],{},[57,3139,2805],{},[71,3141,2805],{"encoding":73},[32,3143,3145],{"className":3144,"ariaHidden":62},[78],[32,3146,3148,3151],{"className":3147},[82],[32,3149],{"className":3150,"style":212},[86],[32,3152,2805],{"className":3153,"style":198},[95,96]," 的分布」——一般是不相等的。两者的差异，正是相关不等于因果的形式化版本。",[21,3156,3157],{"id":3157},"因果与拓扑",[11,3159,3160],{},"一个自然的问题是：几何遗忘度量得到拓扑，那因果是不是遗忘更多东西之后剩下的更轻的骨架？答案是否定的。",[11,3162,3163,3164,3167,3168,3171,3172,3175,3176,3220],{},"几何到拓扑的「变轻」是",[1364,3165,3166],{},"同一个信息维度上的压缩","——都是对称关系，只是分辨率从「连续的数值」降到了「是否连通」的二值\u002F离散判断。但因果结构引入了一个几何和拓扑都没有的新维度：",[1364,3169,3170],{},"方向性\u002F非对称性","，并且这个方向性原则上",[1364,3173,3174],{},"不能","从纯观测的联合分布 ",[32,3177,3179,3199],{"className":3178,"translate":36},[35],[32,3180,3182],{"className":3181},[40],[42,3183,3184],{"xmlns":44},[46,3185,3186,3196],{},[49,3187,3188,3190,3192,3194],{},[57,3189,2101],{},[52,3191,55],{"stretchy":54},[57,3193,1613],{},[52,3195,69],{"stretchy":54},[71,3197,3198],{"encoding":73},"P(V)",[32,3200,3202],{"className":3201,"ariaHidden":62},[78],[32,3203,3205,3208,3211,3214,3217],{"className":3204},[82],[32,3206],{"className":3207,"style":87},[86],[32,3209,2101],{"className":3210,"style":1675},[95,96],[32,3212,55],{"className":3213},[91],[32,3215,1613],{"className":3216,"style":198},[95,96],[32,3218,69],{"className":3219},[113]," 中单独用几何或拓扑手段识别出来。这就是可识别性问题（identifiability）：",[3222,3223,3224],"blockquote",{},[11,3225,3226,3229,3230,3273,3274,3331,3332,3383,3384,3383,3435,3511,3512,3568],{},[1364,3227,3228],{},"命题（因果不可从观测唯一识别）。"," 对于两个变量 ",[32,3231,3233,3251],{"className":3232,"translate":36},[35],[32,3234,3236],{"className":3235},[40],[42,3237,3238],{"xmlns":44},[46,3239,3240,3248],{},[49,3241,3242,3244,3246],{},[57,3243,59],{},[52,3245,63],{"separator":62},[57,3247,2805],{},[71,3249,3250],{"encoding":73},"X, Y",[32,3252,3254],{"className":3253,"ariaHidden":62},[78],[32,3255,3257,3261,3264,3267,3270],{"className":3256},[82],[32,3258],{"className":3259,"style":3260},[86],"height:0.8778em;vertical-align:-0.1944em;",[32,3262,59],{"className":3263,"style":97},[95,96],[32,3265,63],{"className":3266},[101],[32,3268],{"className":3269,"style":106},[105],[32,3271,2805],{"className":3272,"style":198},[95,96],"，仅凭观测联合分布 ",[32,3275,3277,3301],{"className":3276,"translate":36},[35],[32,3278,3280],{"className":3279},[40],[42,3281,3282],{"xmlns":44},[46,3283,3284,3298],{},[49,3285,3286,3288,3290,3292,3294,3296],{},[57,3287,2101],{},[52,3289,55],{"stretchy":54},[57,3291,59],{},[52,3293,63],{"separator":62},[57,3295,2805],{},[52,3297,69],{"stretchy":54},[71,3299,3300],{"encoding":73},"P(X,Y)",[32,3302,3304],{"className":3303,"ariaHidden":62},[78],[32,3305,3307,3310,3313,3316,3319,3322,3325,3328],{"className":3306},[82],[32,3308],{"className":3309,"style":87},[86],[32,3311,2101],{"className":3312,"style":1675},[95,96],[32,3314,55],{"className":3315},[91],[32,3317,59],{"className":3318,"style":97},[95,96],[32,3320,63],{"className":3321},[101],[32,3323],{"className":3324,"style":106},[105],[32,3326,2805],{"className":3327,"style":198},[95,96],[32,3329,69],{"className":3330},[113],"，一般无法在 ",[32,3333,3335,3353],{"className":3334,"translate":36},[35],[32,3336,3338],{"className":3337},[40],[42,3339,3340],{"xmlns":44},[46,3341,3342,3350],{},[49,3343,3344,3346,3348],{},[57,3345,59],{},[52,3347,141],{},[57,3349,2805],{},[71,3351,3352],{"encoding":73},"X \\to Y",[32,3354,3356,3374],{"className":3355,"ariaHidden":62},[78],[32,3357,3359,3362,3365,3368,3371],{"className":3358},[82],[32,3360],{"className":3361,"style":212},[86],[32,3363,59],{"className":3364,"style":97},[95,96],[32,3366],{"className":3367,"style":177},[105],[32,3369,141],{"className":3370},[181],[32,3372],{"className":3373,"style":177},[105],[32,3375,3377,3380],{"className":3376},[82],[32,3378],{"className":3379,"style":212},[86],[32,3381,2805],{"className":3382,"style":198},[95,96],"、",[32,3385,3387,3405],{"className":3386,"translate":36},[35],[32,3388,3390],{"className":3389},[40],[42,3391,3392],{"xmlns":44},[46,3393,3394,3402],{},[49,3395,3396,3398,3400],{},[57,3397,2805],{},[52,3399,141],{},[57,3401,59],{},[71,3403,3404],{"encoding":73},"Y \\to X",[32,3406,3408,3426],{"className":3407,"ariaHidden":62},[78],[32,3409,3411,3414,3417,3420,3423],{"className":3410},[82],[32,3412],{"className":3413,"style":212},[86],[32,3415,2805],{"className":3416,"style":198},[95,96],[32,3418],{"className":3419,"style":177},[105],[32,3421,141],{"className":3422},[181],[32,3424],{"className":3425,"style":177},[105],[32,3427,3429,3432],{"className":3428},[82],[32,3430],{"className":3431,"style":212},[86],[32,3433,59],{"className":3434,"style":97},[95,96],[32,3436,3438,3462],{"className":3437,"translate":36},[35],[32,3439,3441],{"className":3440},[40],[42,3442,3443],{"xmlns":44},[46,3444,3445,3459],{},[49,3446,3447,3449,3452,3455,3457],{},[57,3448,59],{},[52,3450,3451],{},"←",[57,3453,3454],{},"Z",[52,3456,141],{},[57,3458,2805],{},[71,3460,3461],{"encoding":73},"X \\leftarrow Z \\to Y",[32,3463,3465,3483,3502],{"className":3464,"ariaHidden":62},[78],[32,3466,3468,3471,3474,3477,3480],{"className":3467},[82],[32,3469],{"className":3470,"style":212},[86],[32,3472,59],{"className":3473,"style":97},[95,96],[32,3475],{"className":3476,"style":177},[105],[32,3478,3451],{"className":3479},[181],[32,3481],{"className":3482,"style":177},[105],[32,3484,3486,3489,3493,3496,3499],{"className":3485},[82],[32,3487],{"className":3488,"style":212},[86],[32,3490,3454],{"className":3491,"style":3492},[95,96],"margin-right:0.0715em;",[32,3494],{"className":3495,"style":177},[105],[32,3497,141],{"className":3498},[181],[32,3500],{"className":3501,"style":177},[105],[32,3503,3505,3508],{"className":3504},[82],[32,3506],{"className":3507,"style":212},[86],[32,3509,2805],{"className":3510,"style":198},[95,96],"（混杂）等因果结构之间做出唯一判定——它们可以诱导出完全相同的 ",[32,3513,3515,3538],{"className":3514,"translate":36},[35],[32,3516,3518],{"className":3517},[40],[42,3519,3520],{"xmlns":44},[46,3521,3522,3536],{},[49,3523,3524,3526,3528,3530,3532,3534],{},[57,3525,2101],{},[52,3527,55],{"stretchy":54},[57,3529,59],{},[52,3531,63],{"separator":62},[57,3533,2805],{},[52,3535,69],{"stretchy":54},[71,3537,3300],{"encoding":73},[32,3539,3541],{"className":3540,"ariaHidden":62},[78],[32,3542,3544,3547,3550,3553,3556,3559,3562,3565],{"className":3543},[82],[32,3545],{"className":3546,"style":87},[86],[32,3548,2101],{"className":3549,"style":1675},[95,96],[32,3551,55],{"className":3552},[91],[32,3554,59],{"className":3555,"style":97},[95,96],[32,3557,63],{"className":3558},[101],[32,3560],{"className":3561,"style":106},[105],[32,3563,2805],{"className":3564,"style":198},[95,96],[32,3566,69],{"className":3567},[113],"。",[11,3570,3571,3572,3606,3607,1234,3635,3663],{},"要打破这种不可识别性，需要额外的信息源：随机对照实验（真正的 ",[32,3573,3575,3591],{"className":3574,"translate":36},[35],[32,3576,3578],{"className":3577},[40],[42,3579,3580],{"xmlns":44},[46,3581,3582,3588],{},[49,3583,3584,3586],{},[57,3585,66],{"mathvariant":1083},[57,3587,1512],{"mathvariant":1083},[71,3589,3590],{"encoding":73},"\\mathrm{do}",[32,3592,3594],{"className":3593,"ariaHidden":62},[78],[32,3595,3597,3600],{"className":3596},[82],[32,3598],{"className":3599,"style":170},[86],[32,3601,3603],{"className":3602},[95],[32,3604,2776],{"className":3605},[95,2244],"）、时间先后顺序、结构假设（如可加噪声模型、非高斯性，如 LiNGAM 方法利用非高斯噪声打破对称性）。这说明因果结构不是几何\u002F拓扑那条压缩链上的下一站，而是一个正交的轴——即便你拥有关于系统的完整几何信息（所有点对的精确距离）和完整拓扑信息（所有连通关系），你依然可能无法回答「如果我强行改变 ",[32,3608,3610,3623],{"className":3609,"translate":36},[35],[32,3611,3613],{"className":3612},[40],[42,3614,3615],{"xmlns":44},[46,3616,3617,3621],{},[49,3618,3619],{},[57,3620,59],{},[71,3622,59],{"encoding":73},[32,3624,3626],{"className":3625,"ariaHidden":62},[78],[32,3627,3629,3632],{"className":3628},[82],[32,3630],{"className":3631,"style":212},[86],[32,3633,59],{"className":3634,"style":97},[95,96],[32,3636,3638,3651],{"className":3637,"translate":36},[35],[32,3639,3641],{"className":3640},[40],[42,3642,3643],{"xmlns":44},[46,3644,3645,3649],{},[49,3646,3647],{},[57,3648,2805],{},[71,3650,2805],{"encoding":73},[32,3652,3654],{"className":3653,"ariaHidden":62},[78],[32,3655,3657,3660],{"className":3656},[82],[32,3658],{"className":3659,"style":212},[86],[32,3661,2805],{"className":3662,"style":198},[95,96]," 会怎样」这个问题。这正是 Pearl 提出因果之梯（Ladder of Causation）的意义所在：",[3665,3666,3667,3686],"table",{},[3668,3669,3670],"thead",{},[3671,3672,3673,3677,3680,3683],"tr",{},[3674,3675,3676],"th",{},"层级",[3674,3678,3679],{},"问题类型",[3674,3681,3682],{},"例子",[3674,3684,3685],{},"所需信息",[3687,3688,3689,3769,3867],"tbody",{},[3671,3690,3691,3695,3698,3766],{},[3692,3693,3694],"td",{},"1. 关联",[3692,3696,3697],{},"看见（seeing）",[3692,3699,3700],{},[32,3701,3703,3727],{"className":3702,"translate":36},[35],[32,3704,3706],{"className":3705},[40],[42,3707,3708],{"xmlns":44},[46,3709,3710,3724],{},[49,3711,3712,3714,3716,3718,3720,3722],{},[57,3713,2101],{},[52,3715,55],{"stretchy":54},[57,3717,2805],{},[52,3719,2808],{},[57,3721,59],{},[52,3723,69],{"stretchy":54},[71,3725,3726],{"encoding":73},"P(Y\\mid X)",[32,3728,3730,3754],{"className":3729,"ariaHidden":62},[78],[32,3731,3733,3736,3739,3742,3745,3748,3751],{"className":3732},[82],[32,3734],{"className":3735,"style":87},[86],[32,3737,2101],{"className":3738,"style":1675},[95,96],[32,3740,55],{"className":3741},[91],[32,3743,2805],{"className":3744,"style":198},[95,96],[32,3746],{"className":3747,"style":177},[105],[32,3749,2808],{"className":3750},[181],[32,3752],{"className":3753,"style":177},[105],[32,3755,3757,3760,3763],{"className":3756},[82],[32,3758],{"className":3759,"style":87},[86],[32,3761,59],{"className":3762,"style":97},[95,96],[32,3764,69],{"className":3765},[113],[3692,3767,3768],{},"观测数据即可，几何\u002F拓扑视角覆盖此层",[3671,3770,3771,3774,3777,3864],{},[3692,3772,3773],{},"2. 干预",[3692,3775,3776],{},"行动（doing）",[3692,3778,3779],{},[32,3780,3782,3816],{"className":3781,"translate":36},[35],[32,3783,3785],{"className":3784},[40],[42,3786,3787],{"xmlns":44},[46,3788,3789,3813],{},[49,3790,3791,3793,3795,3797,3799,3805,3807,3809,3811],{},[57,3792,2101],{},[52,3794,55],{"stretchy":54},[57,3796,2805],{},[52,3798,2808],{},[49,3800,3801,3803],{},[57,3802,66],{"mathvariant":1083},[57,3804,1512],{"mathvariant":1083},[52,3806,55],{"stretchy":54},[57,3808,59],{},[52,3810,69],{"stretchy":54},[52,3812,69],{"stretchy":54},[71,3814,3815],{"encoding":73},"P(Y\\mid \\mathrm{do}(X))",[32,3817,3819,3843],{"className":3818,"ariaHidden":62},[78],[32,3820,3822,3825,3828,3831,3834,3837,3840],{"className":3821},[82],[32,3823],{"className":3824,"style":87},[86],[32,3826,2101],{"className":3827,"style":1675},[95,96],[32,3829,55],{"className":3830},[91],[32,3832,2805],{"className":3833,"style":198},[95,96],[32,3835],{"className":3836,"style":177},[105],[32,3838,2808],{"className":3839},[181],[32,3841],{"className":3842,"style":177},[105],[32,3844,3846,3849,3855,3858,3861],{"className":3845},[82],[32,3847],{"className":3848,"style":87},[86],[32,3850,3852],{"className":3851},[95],[32,3853,2776],{"className":3854},[95,2244],[32,3856,55],{"className":3857},[91],[32,3859,59],{"className":3860,"style":97},[95,96],[32,3862,2896],{"className":3863},[113],[3692,3865,3866],{},"需要实验或因果图假设",[3671,3868,3869,3872,3875,3936],{},[3692,3870,3871],{},"3. 反事实",[3692,3873,3874],{},"想象（imagining）",[3692,3876,3877,3878,3906,3907,3935],{},"若 ",[32,3879,3881,3894],{"className":3880,"translate":36},[35],[32,3882,3884],{"className":3883},[40],[42,3885,3886],{"xmlns":44},[46,3887,3888,3892],{},[49,3889,3890],{},[57,3891,59],{},[71,3893,59],{"encoding":73},[32,3895,3897],{"className":3896,"ariaHidden":62},[78],[32,3898,3900,3903],{"className":3899},[82],[32,3901],{"className":3902,"style":212},[86],[32,3904,59],{"className":3905,"style":97},[95,96]," 不同，",[32,3908,3910,3923],{"className":3909,"translate":36},[35],[32,3911,3913],{"className":3912},[40],[42,3914,3915],{"xmlns":44},[46,3916,3917,3921],{},[49,3918,3919],{},[57,3920,2805],{},[71,3922,2805],{"encoding":73},[32,3924,3926],{"className":3925,"ariaHidden":62},[78],[32,3927,3929,3932],{"className":3928},[82],[32,3930],{"className":3931,"style":212},[86],[32,3933,2805],{"className":3934,"style":198},[95,96]," 会怎样",[3692,3937,3938],{},"需要完整 SCM，包括外生噪声的联合分布",[11,3940,3941],{},"几何和拓扑视角，本质上被限制在第一层；只有引入因果结构，才能爬上第二、第三层。这也是为什么「因果关系需要额外信息才能识别」这句话不是技术细节，而是三种视角在存在论意义上的分界线。",[21,3943,3944],{"id":3944},"机器学习视角",[11,3946,3947],{},"回到机器学习的语境，三种视角之间的差异直接决定了模型在分布偏移下的表现。可以按「随分布变化保持不变的程度」给三者排一个序：",[32,3949,3951],{"className":3950,"translate":36},[299],[32,3952,3954,3988],{"className":3953,"translate":36},[35],[32,3955,3957],{"className":3956},[40],[42,3958,3959],{"xmlns":44,"display":308},[46,3960,3961,3985],{},[49,3962,3963,3966,3968,3971,3973,3976,3978,3980,3982],{},[372,3964,3965],{},"Geometry",[372,3967,374],{},[52,3969,3970],{},"≺",[372,3972,374],{},[372,3974,3975],{},"Topology Invariance",[372,3977,374],{},[52,3979,3970],{},[372,3981,374],{},[372,3983,3984],{},"Causality",[71,3986,3987],{"encoding":73},"\\text{Geometry} \\;\\prec\\; \\text{Topology Invariance} \\;\\prec\\; \\text{Causality}",[32,3989,3991,4019,4047],{"className":3990,"ariaHidden":62},[78],[32,3992,3994,3997,4004,4007,4010,4013,4016],{"className":3993},[82],[32,3995],{"className":3996,"style":3260},[86],[32,3998,4001],{"className":3999},[95,4000],"text",[32,4002,3965],{"className":4003},[95],[32,4005],{"className":4006,"style":177},[105],[32,4008],{"className":4009,"style":177},[105],[32,4011,3970],{"className":4012},[181],[32,4014],{"className":4015,"style":177},[105],[32,4017],{"className":4018,"style":177},[105],[32,4020,4022,4026,4032,4035,4038,4041,4044],{"className":4021},[82],[32,4023],{"className":4024,"style":4025},[86],"height:0.8889em;vertical-align:-0.1944em;",[32,4027,4029],{"className":4028},[95,4000],[32,4030,3975],{"className":4031},[95],[32,4033],{"className":4034,"style":177},[105],[32,4036],{"className":4037,"style":177},[105],[32,4039,3970],{"className":4040},[181],[32,4042],{"className":4043,"style":177},[105],[32,4045],{"className":4046,"style":177},[105],[32,4048,4050,4053],{"className":4049},[82],[32,4051],{"className":4052,"style":4025},[86],[32,4054,4056],{"className":4055},[95,4000],[32,4057,3984],{"className":4058},[95],[4060,4061,4062,4069,4119],"ul",{},[4063,4064,4065,4068],"li",{},[1364,4066,4067],{},"几何层面的失败","：训练集和测试集哪怕只是特征做了不同的归一化，相似度排序就可能整体偏移——这是最脆弱的一层，任何非本质的表示变化都会渗透进度量数值本身。",[4063,4070,4071,4074,4075,4118],{},[1364,4072,4073],{},"拓扑层面的稳健性","：持久同调等方法对噪声、尺度变化天然不敏感，因为它们本来就是在滤除度量细节之后提取「形状」。但拓扑结构依然是观测分布 ",[32,4076,4078,4097],{"className":4077,"translate":36},[35],[32,4079,4081],{"className":4080},[40],[42,4082,4083],{"xmlns":44},[46,4084,4085,4095],{},[49,4086,4087,4089,4091,4093],{},[57,4088,2101],{},[52,4090,55],{"stretchy":54},[57,4092,1613],{},[52,4094,69],{"stretchy":54},[71,4096,3198],{"encoding":73},[32,4098,4100],{"className":4099,"ariaHidden":62},[78],[32,4101,4103,4106,4109,4112,4115],{"className":4102},[82],[32,4104],{"className":4105,"style":87},[86],[32,4107,2101],{"className":4108,"style":1675},[95,96],[32,4110,55],{"className":4111},[91],[32,4113,1613],{"className":4114,"style":198},[95,96],[32,4116,69],{"className":4117},[113]," 的函数——一旦分布本身发生结构性改变（不只是尺度变化，而是变量间依赖关系变了），拓扑不变量也会随之改变，因为它仍然停留在因果之梯的第一层。",[4063,4120,4121,4124,4125,4175,4176,4316,4317,4361,4362,4446],{},[1364,4122,4123],{},"因果层面的稳健性","：如果 ",[32,4126,4128,4145],{"className":4127,"translate":36},[35],[32,4129,4131],{"className":4130},[40],[42,4132,4133],{"xmlns":44},[46,4134,4135,4143],{},[49,4136,4137,4139,4141],{},[57,4138,59],{},[52,4140,141],{},[57,4142,2805],{},[71,4144,3352],{"encoding":73},[32,4146,4148,4166],{"className":4147,"ariaHidden":62},[78],[32,4149,4151,4154,4157,4160,4163],{"className":4150},[82],[32,4152],{"className":4153,"style":212},[86],[32,4155,59],{"className":4156,"style":97},[95,96],[32,4158],{"className":4159,"style":177},[105],[32,4161,141],{"className":4162},[181],[32,4164],{"className":4165,"style":177},[105],[32,4167,4169,4172],{"className":4168},[82],[32,4170],{"className":4171,"style":212},[86],[32,4173,2805],{"className":4174,"style":198},[95,96]," 的机制 ",[32,4177,4179,4211],{"className":4178,"translate":36},[35],[32,4180,4182],{"className":4181},[40],[42,4183,4184],{"xmlns":44},[46,4185,4186,4208],{},[49,4187,4188,4194,4196,4198,4200,4206],{},[143,4189,4190,4192],{},[57,4191,1909],{},[57,4193,2805],{},[52,4195,55],{"stretchy":54},[57,4197,59],{},[52,4199,63],{"separator":62},[143,4201,4202,4204],{},[57,4203,1488],{},[57,4205,2805],{},[52,4207,69],{"stretchy":54},[71,4209,4210],{"encoding":73},"f_Y(X, U_Y)",[32,4212,4214],{"className":4213,"ariaHidden":62},[78],[32,4215,4217,4220,4261,4264,4267,4270,4273,4313],{"className":4216},[82],[32,4218],{"className":4219,"style":87},[86],[32,4221,4223,4226],{"className":4222},[95],[32,4224,1909],{"className":4225,"style":1965},[95,96],[32,4227,4229],{"className":4228},[242],[32,4230,4232,4253],{"className":4231},[246,247],[32,4233,4235,4250],{"className":4234},[251],[32,4236,4239],{"className":4237,"style":4238},[255],"height:0.3283em;",[32,4240,4241,4244],{"style":1980},[32,4242],{"className":4243,"style":264},[263],[32,4245,4247],{"className":4246},[268,269,270,271],[32,4248,2805],{"className":4249,"style":198},[95,96,271],[32,4251,285],{"className":4252},[284],[32,4254,4256],{"className":4255},[251],[32,4257,4259],{"className":4258,"style":1441},[255],[32,4260],{},[32,4262,55],{"className":4263},[91],[32,4265,59],{"className":4266,"style":97},[95,96],[32,4268,63],{"className":4269},[101],[32,4271],{"className":4272,"style":106},[105],[32,4274,4276,4279],{"className":4275},[95],[32,4277,1488],{"className":4278,"style":1530},[95,96],[32,4280,4282],{"className":4281},[242],[32,4283,4285,4305],{"className":4284},[246,247],[32,4286,4288,4302],{"className":4287},[251],[32,4289,4291],{"className":4290,"style":4238},[255],[32,4292,4293,4296],{"style":2318},[32,4294],{"className":4295,"style":264},[263],[32,4297,4299],{"className":4298},[268,269,270,271],[32,4300,2805],{"className":4301,"style":198},[95,96,271],[32,4303,285],{"className":4304},[284],[32,4306,4308],{"className":4307},[251],[32,4309,4311],{"className":4310,"style":1441},[255],[32,4312],{},[32,4314,69],{"className":4315},[113]," 本身不随环境改变（这是很多领域里的物理假设——比如「重力加速度与颜色无关」），那么哪怕 ",[32,4318,4320,4340],{"className":4319,"translate":36},[35],[32,4321,4323],{"className":4322},[40],[42,4324,4325],{"xmlns":44},[46,4326,4327,4337],{},[49,4328,4329,4331,4333,4335],{},[57,4330,2101],{},[52,4332,55],{"stretchy":54},[57,4334,59],{},[52,4336,69],{"stretchy":54},[71,4338,4339],{"encoding":73},"P(X)",[32,4341,4343],{"className":4342,"ariaHidden":62},[78],[32,4344,4346,4349,4352,4355,4358],{"className":4345},[82],[32,4347],{"className":4348,"style":87},[86],[32,4350,2101],{"className":4351,"style":1675},[95,96],[32,4353,55],{"className":4354},[91],[32,4356,59],{"className":4357,"style":97},[95,96],[32,4359,69],{"className":4360},[113]," 的边际分布因为采样环境不同而剧烈变化，",[32,4363,4365,4398],{"className":4364,"translate":36},[35],[32,4366,4368],{"className":4367},[40],[42,4369,4370],{"xmlns":44},[46,4371,4372,4396],{},[49,4373,4374,4376,4378,4380,4382,4388,4390,4392,4394],{},[57,4375,2101],{},[52,4377,55],{"stretchy":54},[57,4379,2805],{},[52,4381,2808],{},[49,4383,4384,4386],{},[57,4385,66],{"mathvariant":1083},[57,4387,1512],{"mathvariant":1083},[52,4389,55],{"stretchy":54},[57,4391,59],{},[52,4393,69],{"stretchy":54},[52,4395,69],{"stretchy":54},[71,4397,3815],{"encoding":73},[32,4399,4401,4425],{"className":4400,"ariaHidden":62},[78],[32,4402,4404,4407,4410,4413,4416,4419,4422],{"className":4403},[82],[32,4405],{"className":4406,"style":87},[86],[32,4408,2101],{"className":4409,"style":1675},[95,96],[32,4411,55],{"className":4412},[91],[32,4414,2805],{"className":4415,"style":198},[95,96],[32,4417],{"className":4418,"style":177},[105],[32,4420,2808],{"className":4421},[181],[32,4423],{"className":4424,"style":177},[105],[32,4426,4428,4431,4437,4440,4443],{"className":4427},[82],[32,4429],{"className":4430,"style":87},[86],[32,4432,4434],{"className":4433},[95],[32,4435,2776],{"className":4436},[95,2244],[32,4438,55],{"className":4439},[91],[32,4441,59],{"className":4442,"style":97},[95,96],[32,4444,2896],{"className":4445},[113]," 这个机制仍然成立，模型依然可以泛化。这正是 Arjovsky 等人提出不变风险最小化方法的出发点：不去拟合在各个环境里都表现最好的相关性，而是寻找在所有环境下都保持不变的预测机制——本质上是把学习目标从「几何\u002F统计上的最优拟合」换成「因果机制上的不变性」。",[11,4448,4449],{},"用一句话概括：几何模型学的是「数据长什么样」，拓扑模型学的是「数据的骨架怎么连」，而因果模型学的是「数据为什么会长成这样、什么力量在背后生成它」。前两者是对现象的描述，后者试图逼近生成现象的机制——这也是为什么协变量偏移能摧毁前两者，却不一定能摧毁后者。",[21,4451,4452],{"id":4452},"递进的先验",[11,4454,4455,4456,4525],{},"如果用范畴论的语言简单勾勒一下三者的关系：存在遗忘函子 ",[32,4457,4459,4489],{"className":4458,"translate":36},[35],[32,4460,4462],{"className":4461},[40],[42,4463,4464],{"xmlns":44},[46,4465,4466,4486],{},[49,4467,4468,4476,4478],{},[49,4469,4470,4472,4474],{},[57,4471,1496],{"mathvariant":1495},[57,4473,1499],{"mathvariant":1495},[57,4475,1502],{"mathvariant":1495},[52,4477,141],{},[49,4479,4480,4482,4484],{},[57,4481,1509],{"mathvariant":1495},[57,4483,1512],{"mathvariant":1495},[57,4485,11],{"mathvariant":1495},[71,4487,4488],{"encoding":73},"\\mathbf{Met} \\to \\mathbf{Top}",[32,4490,4492,4513],{"className":4491,"ariaHidden":62},[78],[32,4493,4495,4498,4504,4507,4510],{"className":4494},[82],[32,4496],{"className":4497,"style":1546},[86],[32,4499,4501],{"className":4500},[95],[32,4502,1554],{"className":4503},[95,1553],[32,4505],{"className":4506,"style":177},[105],[32,4508,141],{"className":4509},[181],[32,4511],{"className":4512,"style":177},[105],[32,4514,4516,4519],{"className":4515},[82],[32,4517],{"className":4518,"style":1570},[86],[32,4520,4522],{"className":4521},[95],[32,4523,1577],{"className":4524},[95,1553],"，把度量空间的态射（等距同构）松弛成拓扑空间的态射（同胚），信息在这一步单调减少，但依然停留在「对称关系」的范畴里。因果结构则不在这条链上——它是在对象上额外加了一层有向、非对称的生成机制，这层机制无法通过对已有对称结构做进一步遗忘得到，只能通过引入新的认识论工具（干预、时间、结构假设）来「添加」。",[11,4527,4528],{},"所以更准确的图景不是一条线性的「由重到轻」的链条，而是：",[32,4530,4532],{"className":4531,"translate":36},[299],[32,4533,4535,4637],{"className":4534,"translate":36},[35],[32,4536,4538],{"className":4537},[40],[42,4539,4540],{"xmlns":44,"display":308},[46,4541,4542,4634],{},[317,4543,4546,4566,4583,4601,4616],{"rowspacing":4544,"columnalign":4545,"columnspacing":321},"0.16em","center",[323,4547,4548],{},[326,4549,4550],{},[329,4551,4552],{"scriptlevel":157,"displaystyle":54},[4553,4554,4555,4563],"munder",{},[4553,4556,4557,4560],{},[372,4558,4559],{},"几何",[52,4561,4562],{"stretchy":62},"⏟",[372,4564,4565],{},"对称，度量化，最重",[323,4567,4568],{},[326,4569,4570],{},[329,4571,4572],{"scriptlevel":157,"displaystyle":54},[49,4573,4574,4578],{},[52,4575,4577],{"fence":54,"stretchy":62,"minsize":4576,"maxsize":4576},"1.8em","↓",[329,4579,4580],{"scriptlevel":1732,"displaystyle":54},[372,4581,4582],{},"遗忘度量",[323,4584,4585],{},[326,4586,4587],{},[329,4588,4589],{"scriptlevel":157,"displaystyle":54},[4553,4590,4591,4598],{},[4553,4592,4593,4596],{},[372,4594,4595],{},"拓扑",[52,4597,4562],{"stretchy":62},[372,4599,4600],{},"对称，非度量化，中等",[323,4602,4603],{},[326,4604,4605],{},[329,4606,4607],{"scriptlevel":157,"displaystyle":54},[49,4608,4609,4611],{},[52,4610,4577],{"fence":54,"stretchy":62,"minsize":4576,"maxsize":4576},[329,4612,4613],{"scriptlevel":1732,"displaystyle":54},[372,4614,4615],{},"额外的干预\u002F时序信息",[323,4617,4618],{},[326,4619,4620],{},[329,4621,4622],{"scriptlevel":157,"displaystyle":54},[4553,4623,4624,4631],{},[4553,4625,4626,4629],{},[372,4627,4628],{},"因果",[52,4630,4562],{"stretchy":62},[372,4632,4633],{},"非对称，机制化，最稳健",[71,4635,4636],{"encoding":73},"\\begin{array}{c}\n\\underbrace{\\text{几何}}_{\\text{对称，度量化，最重}} \\\\[8pt]\n\\Big\\downarrow \\scriptstyle \\text{遗忘度量} \\\\[8pt]\n\\underbrace{\\text{拓扑}}_{\\text{对称，非度量化，中等}} \\\\[8pt]\n\\Big\\downarrow \\scriptstyle \\text{额外的干预\u002F时序信息} \\\\[8pt]\n\\underbrace{\\text{因果}}_{\\text{非对称，机制化，最稳健}}\n\\end{array}",[32,4638,4640],{"className":4639,"ariaHidden":62},[78],[32,4641,4643,4647],{"className":4642},[82],[32,4644],{"className":4645,"style":4646},[86],"height:11.121em;vertical-align:-5.3105em;",[32,4648,4650],{"className":4649},[95],[32,4651,4653,4658,5237],{"className":4652},[317],[32,4654],{"className":4655,"style":4657},[4656],"arraycolsep","width:0.5em;",[32,4659,4662],{"className":4660},[4661],"col-align-c",[32,4663,4665,5228],{"className":4664},[246,247],[32,4666,4668,5225],{"className":4667},[251],[32,4669,4672,4816,4902,5023,5104],{"className":4670,"style":4671},[255],"height:5.8105em;",[32,4673,4675,4679],{"style":4674},"top:-8.1215em;",[32,4676],{"className":4677,"style":4678},[263],"height:3.151em;",[32,4680,4682],{"className":4681},[95],[32,4683,4685],{"className":4684},[507,4553],[32,4686,4688,4807],{"className":4687},[246,247],[32,4689,4691,4804],{"className":4690},[251],[32,4692,4694,4713],{"className":4693,"style":212},[255],[32,4695,4697,4700],{"style":4696},"top:-1.6737em;",[32,4698],{"className":4699,"style":715},[263],[32,4701,4703],{"className":4702},[268,269,270,271],[32,4704,4706],{"className":4705},[95,271],[32,4707,4709],{"className":4708},[95,4000,271],[32,4710,4565],{"className":4711},[95,4712,271],"cjk_fallback",[32,4714,4716,4719],{"style":4715},"top:-3em;",[32,4717],{"className":4718,"style":715},[263],[32,4720,4722],{"className":4721},[507,4553],[32,4723,4725,4795],{"className":4724},[246,247],[32,4726,4728,4792],{"className":4727},[251],[32,4729,4731,4778],{"className":4730,"style":212},[255],[32,4732,4736,4739],{"className":4733,"style":4735},[4734],"svg-align","top:-2.352em;",[32,4737],{"className":4738,"style":715},[263],[32,4740,4744,4758,4768],{"className":4741,"style":4743},[4742],"stretchy","height:0.548em;min-width:1.6em;",[32,4745,4749],{"className":4746,"style":4748},[4747],"brace-left","height:0.548em;",[551,4750,4755],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4754},"400em","0.548em","0 0 400000 548","xMinYMin slice",[560,4756],{"d":4757},"M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13\n 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688\n 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7\n-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z",[32,4759,4762],{"className":4760,"style":4748},[4761],"brace-center",[551,4763,4765],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4764},"xMidYMin slice",[560,4766],{"d":4767},"M199572 214\nc100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14\n 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3\n 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0\n-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z",[32,4769,4772],{"className":4770,"style":4748},[4771],"brace-right",[551,4773,4775],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4774},"xMaxYMin slice",[560,4776],{"d":4777},"M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3\n 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237\n-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z",[32,4779,4780,4783],{"style":4715},[32,4781],{"className":4782,"style":715},[263],[32,4784,4786],{"className":4785},[95],[32,4787,4789],{"className":4788},[95,4000],[32,4790,4559],{"className":4791},[95,4712],[32,4793,285],{"className":4794},[284],[32,4796,4798],{"className":4797},[251],[32,4799,4802],{"className":4800,"style":4801},[255],"height:0.648em;",[32,4803],{},[32,4805,285],{"className":4806},[284],[32,4808,4810],{"className":4809},[251],[32,4811,4814],{"className":4812,"style":4813},[255],"height:1.3263em;",[32,4815],{},[32,4817,4819,4822],{"style":4818},"top:-5.6442em;",[32,4820],{"className":4821,"style":4678},[263],[32,4823,4825,4896],{"className":4824},[95],[32,4826,4828],{"className":4827},[95],[32,4829,4831],{"className":4830},[514,515],[32,4832,4834,4887],{"className":4833},[246,247],[32,4835,4837,4884],{"className":4836},[251],[32,4838,4841,4854,4872],{"className":4839,"style":4840},[255],"height:1.151em;",[32,4842,4844,4848],{"style":4843},"top:-1.966em;",[32,4845],{"className":4846,"style":4847},[263],"height:2.616em;",[32,4849,4852],{"className":4850},[536,4851],"delim-size1",[32,4853,4577],{},[32,4855,4857,4860],{"style":4856},"top:-2.558em;",[32,4858],{"className":4859,"style":4847},[263],[32,4861,4863],{"style":4862},"height:0.616em;width:0.6667em;",[551,4864,4869],{"xmlns":553,"width":4865,"height":4866,"style":4867,"viewBox":4868,"preserveAspectRatio":558},"0.6667em","0.616em","width:0.6667em","0 0 666.67 616",[560,4870],{"d":4871},"M312 0 H355 V616 H312z M312 0 H355 V616 H312z",[32,4873,4875,4878],{"style":4874},"top:-3.166em;",[32,4876],{"className":4877,"style":4847},[263],[32,4879,4881],{"className":4880},[536,4851],[32,4882,4883],{},"⏐",[32,4885,285],{"className":4886},[284],[32,4888,4890],{"className":4889},[251],[32,4891,4894],{"className":4892,"style":4893},[255],"height:0.65em;",[32,4895],{},[32,4897,4899],{"className":4898},[95,4000,271,268,269,270],[32,4900,4582],{"className":4901},[95,4712,271],[32,4903,4905,4908],{"style":4904},"top:-3.6442em;",[32,4906],{"className":4907,"style":4678},[263],[32,4909,4911],{"className":4910},[95],[32,4912,4914],{"className":4913},[507,4553],[32,4915,4917,5015],{"className":4916},[246,247],[32,4918,4920,5012],{"className":4919},[251],[32,4921,4923,4940],{"className":4922,"style":212},[255],[32,4924,4925,4928],{"style":4696},[32,4926],{"className":4927,"style":715},[263],[32,4929,4931],{"className":4930},[268,269,270,271],[32,4932,4934],{"className":4933},[95,271],[32,4935,4937],{"className":4936},[95,4000,271],[32,4938,4600],{"className":4939},[95,4712,271],[32,4941,4942,4945],{"style":4715},[32,4943],{"className":4944,"style":715},[263],[32,4946,4948],{"className":4947},[507,4553],[32,4949,4951,5004],{"className":4950},[246,247],[32,4952,4954,5001],{"className":4953},[251],[32,4955,4957,4987],{"className":4956,"style":212},[255],[32,4958,4960,4963],{"className":4959,"style":4735},[4734],[32,4961],{"className":4962,"style":715},[263],[32,4964,4966,4973,4980],{"className":4965,"style":4743},[4742],[32,4967,4969],{"className":4968,"style":4748},[4747],[551,4970,4971],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4754},[560,4972],{"d":4757},[32,4974,4976],{"className":4975,"style":4748},[4761],[551,4977,4978],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4764},[560,4979],{"d":4767},[32,4981,4983],{"className":4982,"style":4748},[4771],[551,4984,4985],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4774},[560,4986],{"d":4777},[32,4988,4989,4992],{"style":4715},[32,4990],{"className":4991,"style":715},[263],[32,4993,4995],{"className":4994},[95],[32,4996,4998],{"className":4997},[95,4000],[32,4999,4595],{"className":5000},[95,4712],[32,5002,285],{"className":5003},[284],[32,5005,5007],{"className":5006},[251],[32,5008,5010],{"className":5009,"style":4801},[255],[32,5011],{},[32,5013,285],{"className":5014},[284],[32,5016,5018],{"className":5017},[251],[32,5019,5021],{"className":5020,"style":4813},[255],[32,5022],{},[32,5024,5026,5029],{"style":5025},"top:-1.1668em;",[32,5027],{"className":5028,"style":4678},[263],[32,5030,5032,5089],{"className":5031},[95],[32,5033,5035],{"className":5034},[95],[32,5036,5038],{"className":5037},[514,515],[32,5039,5041,5081],{"className":5040},[246,247],[32,5042,5044,5078],{"className":5043},[251],[32,5045,5047,5057,5068],{"className":5046,"style":4840},[255],[32,5048,5049,5052],{"style":4843},[32,5050],{"className":5051,"style":4847},[263],[32,5053,5055],{"className":5054},[536,4851],[32,5056,4577],{},[32,5058,5059,5062],{"style":4856},[32,5060],{"className":5061,"style":4847},[263],[32,5063,5064],{"style":4862},[551,5065,5066],{"xmlns":553,"width":4865,"height":4866,"style":4867,"viewBox":4868,"preserveAspectRatio":558},[560,5067],{"d":4871},[32,5069,5070,5073],{"style":4874},[32,5071],{"className":5072,"style":4847},[263],[32,5074,5076],{"className":5075},[536,4851],[32,5077,4883],{},[32,5079,285],{"className":5080},[284],[32,5082,5084],{"className":5083},[251],[32,5085,5087],{"className":5086,"style":4893},[255],[32,5088],{},[32,5090,5092,5096,5100],{"className":5091},[95,4000,271,268,269,270],[32,5093,5095],{"className":5094},[95,4712,271],"额外的干预",[32,5097,5099],{"className":5098},[95,271],"\u002F",[32,5101,5103],{"className":5102},[95,4712,271],"时序信息",[32,5105,5107,5110],{"style":5106},"top:0.8332em;",[32,5108],{"className":5109,"style":4678},[263],[32,5111,5113],{"className":5112},[95],[32,5114,5116],{"className":5115},[507,4553],[32,5117,5119,5217],{"className":5118},[246,247],[32,5120,5122,5214],{"className":5121},[251],[32,5123,5125,5142],{"className":5124,"style":212},[255],[32,5126,5127,5130],{"style":4696},[32,5128],{"className":5129,"style":715},[263],[32,5131,5133],{"className":5132},[268,269,270,271],[32,5134,5136],{"className":5135},[95,271],[32,5137,5139],{"className":5138},[95,4000,271],[32,5140,4633],{"className":5141},[95,4712,271],[32,5143,5144,5147],{"style":4715},[32,5145],{"className":5146,"style":715},[263],[32,5148,5150],{"className":5149},[507,4553],[32,5151,5153,5206],{"className":5152},[246,247],[32,5154,5156,5203],{"className":5155},[251],[32,5157,5159,5189],{"className":5158,"style":212},[255],[32,5160,5162,5165],{"className":5161,"style":4735},[4734],[32,5163],{"className":5164,"style":715},[263],[32,5166,5168,5175,5182],{"className":5167,"style":4743},[4742],[32,5169,5171],{"className":5170,"style":4748},[4747],[551,5172,5173],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4754},[560,5174],{"d":4757},[32,5176,5178],{"className":5177,"style":4748},[4761],[551,5179,5180],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4764},[560,5181],{"d":4767},[32,5183,5185],{"className":5184,"style":4748},[4771],[551,5186,5187],{"xmlns":553,"width":4751,"height":4752,"viewBox":4753,"preserveAspectRatio":4774},[560,5188],{"d":4777},[32,5190,5191,5194],{"style":4715},[32,5192],{"className":5193,"style":715},[263],[32,5195,5197],{"className":5196},[95],[32,5198,5200],{"className":5199},[95,4000],[32,5201,4628],{"className":5202},[95,4712],[32,5204,285],{"className":5205},[284],[32,5207,5209],{"className":5208},[251],[32,5210,5212],{"className":5211,"style":4801},[255],[32,5213],{},[32,5215,285],{"className":5216},[284],[32,5218,5220],{"className":5219},[251],[32,5221,5223],{"className":5222,"style":4813},[255],[32,5224],{},[32,5226,285],{"className":5227},[284],[32,5229,5231],{"className":5230},[251],[32,5232,5235],{"className":5233,"style":5234},[255],"height:5.3105em;",[32,5236],{},[32,5238],{"className":5239,"style":4657},[4656],[11,5241,5242],{},"前两步是同一维度上的信息压缩，第三步是维度的跃迁——从「关系是否存在\u002F多强」跃迁到「关系为什么存在、能否被操控」。这三层构成了一个智能系统理解世界所需的最小先验地基：先要有度量才能比较，有了比较才能抽象出结构骨架，而只有抽出骨架背后的生成机制，系统才能真正做到举一反三——在没见过的分布上依然可靠地行动，而不只是在训练分布的邻域里插值。",[11,5244,5245],{},"这或许也是当前很多「表示学习」范式的天花板所在：无论嵌入空间做得多精细（几何）、图结构建模得多完备（拓扑），只要模型学到的始终是观测层面的关联，它就永远停留在因果之梯的第一级。因果表征学习（causal representation learning）想做的事情，正是把几何\u002F拓扑意义上学到的表示，进一步「提纯」出背后满足机制不变性的潜变量和因果图——这也是几何、拓扑、因果这三种视角在方法论上真正会师的地方。",{"title":5247,"searchDepth":5248,"depth":5248,"links":5249},"",2,[5250,5251,5252,5253,5254,5255],{"id":23,"depth":5248,"text":24},{"id":1172,"depth":5248,"text":1173},{"id":1581,"depth":5248,"text":1582},{"id":3157,"depth":5248,"text":3157},{"id":3944,"depth":5248,"text":3944},{"id":4452,"depth":5248,"text":4452},false,true,"md",null,{},"2026-08-19","\u002Fblog\u002F2026\u002F2026-08-19-yan-jiu-guan-xi-de-san-zhong-shi-jiao",{"title":6,"description":13},"blog\u002F2026\u002F2026-08-19-yan-jiu-guan-xi-de-san-zhong-shi-jiao","研究「关系」有三种互补视角：几何，拓扑和因果。分别关注测量、连通和约束。它们不是并列的学科，而是看待「存在」的三个层层递进的维度。三者共同构成智能系统理解世界的先验地基。",[5267],"machine-learning","JouSEmv-eU7CnfBnekiG75RAlq4ZoOC_3cvO_1NWTAI",{"id":5270,"title":5271,"body":5272,"description":5276,"draft":5256,"enableComment":5257,"extension":5258,"image":5247,"important":5256,"location":5529,"meta":5530,"navigation":5257,"ogImage":5259,"onday":5531,"path":5532,"seo":5533,"stem":5534,"summary":5535,"tags":5536,"__hash__":5538},"blog\u002Fblog\u002F2026\u002F2026-08-14-demystification-research.md","祛魅科研，每个研究生的必修课",{"type":8,"value":5273,"toc":5521},[5274,5277,5280,5283,5286,5289,5292,5302,5305,5308,5311,5314,5317,5320,5323,5326,5329,5332,5336,5339,5342,5345,5352,5359,5362,5365,5368,5371,5374,5380,5383,5386,5389,5392,5397,5400,5403,5406,5409,5414,5417,5420,5427,5430,5433,5436,5439,5442,5445,5451,5454,5457,5460,5463,5466,5469,5472,5475,5478,5485,5489,5495,5500,5503,5506,5509,5512,5515,5518],[11,5275,5276],{},"从 2025 年开始入学读研，到现在已经有一个年头了，不知不觉已经研二了。在过去的一年的研究生学习中，我发现我的心态、世界观经历了一次完整的、打击性的重塑。如果说之前我的心态是中规中矩，那么现在则是满心失落。自从经历过完整的科研训练，调研课题、找 idea，做实验、处理数据、撰写论文、修改文章、等投稿、写 Rebuttal、拒稿再转投、论文改版等一系列流程完整经历过后，只剩下满心苦涩。一度怀疑，我们搞的这个所谓的「科研」的价值和意义。😇",[11,5278,5279],{},"因为笔者读的研究生是计算机工程类专业，对于其他理工科领域、文科领域，笔者不是很熟悉，所以以下观点体验仅适用于该专业。",[21,5281,5282],{"id":5282},"草台班子",[11,5284,5285],{},"第一个对我的改变就是对学术成果的怀疑。学术圈的科研，其实并没有那么高端，可以说是草台班子。",[11,5287,5288],{},"在读研之前，我一直认为学术论文其实是非常高端的东西，至少是可信度很高的参考。但是读研之后，这个想法被改变了：绝大多数论文，他们的观点、数据、实验，看看就好，不必认真。",[11,5290,5291],{},"一篇论文的结论，往往只有作者自己的机器上成立。审稿人不会复现，编辑不会复现，引用者也不会复现。大家都默认了这个不可验证的游戏潜规则。论文并不是经过验证的知识，而是一个经过美化的故事。",[11,5293,5294,5295,5301],{},"有人专门对被 ICML 收录的 Oral 论文进行了复现，并给出了",[2104,5296,5300],{"href":5297,"rel":5298},"https:\u002F\u002Fx.com\u002FChenhaoTan\u002Fstatus\u002F2079969545629118737",[5299],"nofollow","论文的报告信息","。结果发现，在 105 篇完整的复现论文中，只有 27 篇复现了超过 40% 的声称效果，剩余的几乎无法复现，要么效果不达标。原因包括代码无法运行、文件缺失以及模型已弃用等等。",[11,5303,5304],{},"复现困难确实是机器学习领域长期存在的问题。因为，CUDA 版本、显卡型号、CPU 型号、PyTorch 框架版本，可能都有大量隐蔽的细节差异，如果再组合起来，那么效果就完全无法验证。实验中的很多关键 Hack（比如特定的随机种子、数据预处理的微妙顺序、超参数的精细调节）可能只存在于作者的实验笔记或潜意识里。这些隐性经验很难通过文字传递，导致复现者像在黑暗中摸索。",[11,5306,5307],{},"甚至如果作者的代码里有 Bug，也很可能会造成完全错误的结论。我曾经跑过前人的基线实验，发现一篇论文，他的某个指标又非常虚高。最后才发现，作者对数据集根本没有清洗干净，也就是说，这份成绩是假的。",[11,5309,5310],{},"但问题是，这个论文已经被大量后人引用且作为基线了，后人想要再投稿，那么他的成绩必须要比这个假成绩更高，但是你根本比不过这假成绩，怎么办呢？你只能偷偷摸摸用点魔法手段，万不可认真，否则反而会显得像是你做错了，审稿人不接受。后来的研究者反而成了受害者。这不就是难为后人么？",[11,5312,5313],{},"所以，后人也只能不得不逼着也参与这种数字游戏，偷偷使用手段。结果就是，学术论文逐渐沦为一场“数字通胀”。每一篇新论文都在前人的数字泡沫上再吹一层泡沫，直到某个数据集上的准确率达到 99.9%，然后这个方向就“死掉”了——因为没人能再超越了，而大家都知道那个 99.9% 是假的，但谁都不想捅破。",[11,5315,5316],{},"这已经超出了学术不端的范畴，演变成了一种系统性的「囚徒困境」，很遗憾，这种困境，在学术圈里几乎无解。",[11,5318,5319],{},"无法复现不一定是作者故意有造假意向，而是因为深度学习本身就是一个巨大的黑盒，可验证性、可解释性非常差。很多错误，作者不一定能及时发现，审稿人也不一定能发现，于是，“只要能 work 就能发文”。当审稿人面对一篇充满“惊艳”实验数据的论文时，他既没有算力去复现，也没有理论工具去证伪其内部的混沌过程。于是，审稿的标准退化为：“只要故事逻辑自洽，且结果看起来比 SOTA（当前最优）高，我就信。”这就给出了巨大的浑水摸鱼的空间。",[11,5321,5322],{},"这种无法复现的现象，也会给自己带来很多的麻烦：如果你的文章需要依赖前人的工作，但是前人的工作本身就有问题，那么你的工作就几乎不可能顺利下去。你只能花大把时间来反思、调试，最后还很可能一无所获。",[11,5324,5325],{},"最后被浪费大把的时间和精力。这种消耗是精神凌迟。后续工作往往不得不替它买单。",[11,5327,5328],{},"有时候，我也甚至怀疑作者是故意在开源里缺斤少两，删减关键代码和文件或者埋入 Bug，或者论文被 Accept 后立即下架数据集，极力避免复现。😇要知道，审稿人通常只有 2-3 位，他们在几周内免费审稿，不可能复现你的实验，也无法验证你的原始数据。他们主要检查的是逻辑是否顺畅和方法是否看起来合理，而不是结论是否绝对正确。",[11,5330,5331],{},"学术论文它的首要目标是发表，而不是传世。为了发表，作者必须讲一个完整且自洽的故事。他们会突出最漂亮的数据，而弱化不支持的“噪音”结果。所以，他们会在讨论部分夹带私货，把故事讲得更有吸引力。它也确实是痛苦的。因为这意味着你失去了一个可以无条件信任的知识权威。",[21,5333,5335],{"id":5334},"ai-审稿危机","AI 审稿危机",[11,5337,5338],{},"读研起到现在，我已经投稿了三篇文章，这点我是有亲身体会的。",[11,5340,5341],{},"另一种尴尬是 AI 时代的审稿危机。AI 时代的恶果就是，论文灌水越来越容易。往常，写一篇论文很可能需要一学期、大半年的时间，要辛辛苦苦做实验、分析结果、画图表、斟酌写作。但是现在不一样了，利用自动化学术 Agent 如 AutoResearch，只要输入合适的提示词和模糊的想法再交给 Agent，它两天时间就可以编出一篇像模像样的论文，自动编写程序做实验，自动画图表，一个人一个月就能编写出 10 篇论文，直接投到顶会轰炸。",[11,5343,5344],{},"AAAI 在 2026 年收到了五万多份取号，几乎是 2025 年的两倍。而前年 2024 年才不过九千多份，短短两年时间就增长了五倍！其实不止 AAAI，其他的计算机会议的投稿量也是几乎翻倍增长。假设每篇论文标准配备 3 名审稿人，组委会将需要管理 120,000 份独立审稿意见：",[11,5346,5347],{},[5348,5349],"img",{"alt":5350,"src":5351},"计算机顶会近五年投稿量。自从 2025 年后，大多数投稿量都是翻倍式增长","https:\u002F\u002Fimage-assets.dreams.plus\u002F202608111219721.jpg",[11,5353,5354,5355,5358],{},"这就会导致一个后果：劣币驱逐良币。审稿体系正在加速崩溃。就算你辛辛苦苦完成了心血，设计实验，写好文章，",[1364,5356,5357],{},"审稿人也几乎不会认真看你的文章，而是直接交给 LLM 敷衍。"," 因为审稿人压根无法区分哪篇论文是 AI 写得，哪篇不是。更糟糕的是，学术领域是高度分化的，审稿人极有可能也不熟悉、也不理解你的研究领域。所以，他也只能同样一股脑交给 AI 审稿。",[11,5360,5361],{},"在这种情况下，灌水零成本，中间大量灌水的作者和敷衍的审稿人占多数。认真的审稿人和认真的作者只能被动深受其害。受害的永远只是认真搞学术的你。在这种情况下，审稿工作还能正常进行下去吗？",[11,5363,5364],{},"而 LLM 审稿本身就有很大的问题，AI 并不会理解你的文章，LLM 的幻觉问题众所周知。它会用非常刁钻的角度在你的文章里挑刺，哪怕是没有问题也会制造出一些问题。这是因为 LLM 在训练、SFT 偏好微调的时候，天生就倾向于给你打低分。LLM 并不真正拥有论文作者的研究上下文。它尤其容易犯一种很危险的错误：把“我没理解”转换成“作者的方法存在问题”。",[11,5366,5367],{},"LLM 模型学习到的统计规律是：“审稿”这个动作的语义空间，几乎完全由“批评性词汇”构成。因此，即使一篇论文毫无瑕疵，模型根据概率分布生成的“审稿风格”文本，天然就带有负面倾向。SFT 微调的时候，宽松的论文意见并不受欢迎，为了要尽可能压榨 LLM 能力，在微调的时候会特意设计出非常严苛的训练案例。这种奖惩机制直接塑造了模型的“人格”：它必须“生产批评”来证明自己的价值。哪怕没有真实缺陷，它也会启动“防御性挑刺”模式，利用模式匹配强行构造出看似合理的问题。",[11,5369,5370],{},"我的文章被拒稿过一次，三个 Reviewer 里两人给出的意见，明显就是 AI 复制粘贴的。😅 但是你不能抗议，你还要捏着鼻子忍着恶心，在 rebuttal 里面假模假样地感谢审稿人，再假模假样地写 rebuttal，跪求他赏赐一口饭吃。😅",[11,5372,5373],{},"审稿体系存在一个非常明显的权力不对称。Reviewer 可以随意评价“The novelty is insufficient.”而作者很难说：“你根本没读懂我的论文。”因为作者没有证据。另外，Reviewer 可以洋洋洒洒写出几千字的审稿意见，没有字数限制，而作者给的 Rebuttal 却严格限制在几百字以内。这也是最恶心的一点😅",[11,5375,5376],{},[5348,5377],{"alt":5378,"src":5379},"如图，是几个 LLM 在审稿 ArXiv 计算机学科论文时给出的平均得分（10 分制）","https:\u002F\u002Fimage-assets.dreams.plus\u002F202608111210591.png",[11,5381,5382],{},"试过把 21-23 年，GPT 出来以前的三大顶会里，已经发表收录的论文随机爬下来，投给 AI 审，50 篇里 30 多个 weak reject，10 多个 weak accept，剩下的全部都是 reject。一篇 accept 都没有。后 AI 时代所谓的顶会论文已经变成笑话哩。",[11,5384,5385],{},"审稿人因为反驳文太长而疲惫，作者因为审稿人固执而绝望。当作者知道审稿人是 AI，审稿人知道作者用了 AI 时，作者 - 审稿人 - 编辑三方之间的这场学术对话，就彻底沦为了一场荒诞剧。这就造成了非常滑稽的景象，AI 写 AI 审：",[11,5387,5388],{},"用 AI 写论文、写代码，再用 AI 初审，根据 AI 的意见修改，完成初稿。审稿人拿到初稿，再交给 AI 审稿，用 AI 的意见给出 review 意见，作者拿到 AI 写的意见后再交给 AI 写 rebuttal。审稿人再用 AI 根据 rebuttal 做出决定。堪称学术出版领域正在逼近的赛博朋克式奇观。",[11,5390,5391],{},"在三方中，编辑\u002F主席的地位最为尴尬。当所有文字意见如审稿、反驳都由 AI 生成时，编辑失去了判断学术创新性的任何抓手。他唯一能做的，就是检查流程是否走完：AI 是否提了 3 个问题？作者是否逐条回复？回复长度是否达标？只要格式合规，就可以按下接受键。学术判断的权力，在此刻已完全让渡给了硅基算法。",[11,5393,5394],{},[1364,5395,5396],{},"所以，能否发顶会、顶刊，实际上已经越来越像摸彩票中奖的运气、概率问题，跟你的文章质量、工作效果已经几乎没有什么关系了。",[21,5398,5399],{"id":5399},"手艺人",[11,5401,5402],{},"硕博生这个身份曾经是不少人引以为傲的根本。但是真正体验过他们的生活，也会发现他们本质上和流水线的螺丝工人相差无几，这种生活状态其实是非常压抑的，如果过这种日子，那只能用「熬」来形容。",[11,5404,5405],{},"硕博生本质上也是出卖高强度脑力劳动换取生存资格的劳动者。跟工地上抗水泥的农民工、顶着大太阳湿透衣服的清洁工没有本质区别。",[11,5407,5408],{},"对于理工科研究生，996 是常态，实验室的灯永远亮着。每天睁眼闭眼就是要面对屏幕上密密麻麻的实验数据、代码和仪器。高强度脑力劳动后，带给身心的除了疲劳还是疲劳。日子久了，精气神会被消磨掉，慢慢丧失对这个世界的好奇心和一切欲望，不想谈恋爱，不想出去旅游，哪怕是手里的游戏，日子久了玩起来也没意思。",[3222,5410,5411],{},[11,5412,5413],{},"到周末后，只想在床上躺着，啥也不干。脑一旦被单一的高强度任务长期占据，负责发散思维、感受情绪、产生欲望的脑区就会被持续抑制。只像一个漂浮在数据海洋里的意识，拖着一具沉重而麻木的肉体，犹如冢中枯骨而已。",[11,5415,5416],{},"这是最致命的。当你看清你所做的研究可能只是学术游戏里的一个废棋，对外部世界毫无影响时，熬就变成了一种精神上的凌迟。不禁会问，「我做的这一切，受了那么多折磨，到底有什么意义？」",[11,5418,5419],{},"工人进厂时，起码能自知之明、清醒地知道这是出卖身体，用劳动换生存。但硕博生被社会、被家人、被曾经的自己赋予了天子骄子、知识精英的光环。当现实变成日复一日地跑数据、伺候仪器时，日子久了你会有这样一种感觉：自己并不是一个人，而是一个庞大机器中的一颗零件。",[11,5421,5422,5423,5426],{},"硕士生（Master）也不是大师。博士生也不博学。他的知识广度，甚至还可能不如一个高中生。",[1364,5424,5425],{},"读研后，你的视角只会被限制在高度狭窄且专业的小领域里。"," 在自己的学术孤岛里自娱自乐。",[11,5428,5429],{},"中世纪的经院哲学家热衷于争论「一个针尖上能站几个天使」，而今天的学术圈，大量精力被消耗在维护主流范式上。正如前文所提到的，学术论文本身也是高度固定化的、范式的。有时候你不会感觉自己「在创造一个想法」，而是「完成一个八股文」。",[11,5431,5432],{},"如果要比喻硕博生这个群体的身份，它更像是一个高度程序化的手艺人、螺丝工。搞科研的流程本身就是高度固定化、流水线化的。调研、idea、实验、写作、改稿、Rebuttal。这一圈下来，恭喜你，你已经是一名合格的劳工！",[21,5434,5435],{"id":5435},"学术圈",[11,5437,5438],{},"曾经对学术圈的浪漫想象，至少代表了知识的前沿、先进的生产力和思想文化，具有进步性。我其实对科研领域内的学者，高校里的教授、教师等群体，长期是存在敬仰的，认为他们或多或少都是代表了人类开拓认知知识、征服星辰大海的一批人，在心里也会敬三分。",[11,5440,5441],{},"现在才知道，学术圈其实是高度封闭的。因为知识本身就有很强的入门壁垒，当人类认知突破到一定边界时，工具、术语和范式的复杂度必然形成门槛，学术方向往往会走向高度分化、隔行如隔山的细碎分支，各个分支又会高度壁垒。这就造成了学术圈的高度封闭性。",[11,5443,5444],{},"其实，越是高度封闭的圈子，越容易产生高度固化的权力结构，行事作风越是封建化、越是讲政治。😅其实现在的学术圈其实跟欧洲中世纪的经学院教派之争、西藏喇嘛们的辩经并没有什么区别。那些专家，领域学者，头衔看得是挺唬人，但做的无非在极度封闭的圈子里，用只有内部人能懂的黑话，争论着对外部世界影响甚微的问题，而决定胜负的常常不是真理，而是资历、人脉和对经典的诠释权。",[11,5446,5447,5450],{},[1364,5448,5449],{},"学术圈，其实比大多数人想象的还要小。"," 因为现代学科已经进入高度分化、高度专业化的时代了。如果细分下去，全中国十四亿人里，同一个领域的研究同行很可能不超过百人，甚至十几人。在一个村落里，任何小动作，全落在这几十个低头不见抬头见的人手里。这里没法对事不对人，因为在结构上，所有的事，最终都是人的事。",[11,5452,5453],{},"然而，现在的学术圈是零和博弈，资源是极其有限的。这个在申请基金、文章版面、学术交流等等活动中，只要有人胜出，必然会有人落选。这也就意味着，你的小圈子里，可能到处都在「树敌」。这并不意味着本人有错，而是你的存在，本身就是威胁。",[11,5455,5456],{},"所以遇到同行暗中使绊子，也是常见的事。你的审稿人很可能跟你的导师有竞争或者过节，就直接轻松 Reject 你的心血。即使现在的审稿制度大多是双盲制，在一个领域只有几十上百人的圈子里，根本不存在真正的双向匿名。看研究问题、看方法、看引用的文献，审稿人闭着眼都能猜到这篇稿子出自哪个课题组。",[11,5458,5459],{},"对一个埋头苦干的学生来说，这是最深的打击。你相信公正，相信学术质量至上。然后，一堵由学派、人情、资源争夺构成的墙，悄无声息地挡在你面前，将你的心血轻松驳回。你甚至没有一个明确的敌人去质问抗争，你甚至不知道你的敌人是谁，又得罪了谁，只剩下无尽的无力感和被戏弄的愤怒。这种被暗算的体验会深刻腐蚀对学术共同体的信任。",[11,5461,5462],{},"像中世纪的领主分封土地一样，大牛导师和顶尖实验室把持着顶级期刊的版面、重大项目的经费和学阀圈子的话语权。你想在他的领地上发文章，就得遵循他的范式、引他的文章、甚至拜他的码头。学术圈的游戏规则是，正确不等于接受。投稿像一场赌博，审稿人的口味、当期版面、甚至运气，都比你那篇精心打磨的论文权重更高。",[21,5464,5465],{"id":5465},"事业意义",[11,5467,5468],{},"虽然小时候有「长大要当科学家」这种理想，但是，个人认为「学术」这条路并不适合像我这样的平民子弟。没有充裕的家底和财力作为后盾，吃学术这碗饭，也是一种高风险职业。",[11,5470,5471],{},"我觉得那些在学术圈里搞研究的人，其实也挺可悲的。自己把大量的青春，时间，精力投入到自己的课题里，勉强能讨得经费，靠这个饭碗。因为，选择某个研究领域，在初期往往带有偶然性。但一旦投入，就成了无法回头的豪赌。赌的是这个方向在几十年内不被证伪、不被超越、不被认为是死胡同、不会没落。他用的是几十年的青春和精力做赌注，用最严谨的头脑，从事着一项本质上充满不确定性的高风险事业。",[11,5473,5474],{},"如果有人突然跳出来用新理论、新范式挑战他，或者推翻了他的观点课题，这无异于把他降维打击成了一块废品，弃之如敝履，这不是嘲讽，而是真实的悲剧。在高度职业化的学术圈里，一个人的身份、地位、自尊，都深深扎根于他那一亩三分地的研究课题。",[11,5476,5477],{},"当他的理论被推翻，在外人看来不过是一个观点被证伪。但对他而言，无异于整个学术人格被判处了死刑，在这个圈子里，会被迅速边缘化，从而判了死刑。他毕生构建的意义大厦，瞬间崩塌为一座废墟。这就是一种存在主义危机。",[11,5479,5480,5481,5484],{},"大多数普通人活下去本身就很难。因为 ",[1364,5482,5483],{},"他们的人生，还有其他责任"," 。若压制住七情六欲，寒窗几十年，去碰学术圈，也未免太委屈了。如果一个人 25 岁读博士，30 岁左右博士毕业，然后经历博士后、非升即走、青年项目竞争，他可能在四十岁前都处于高度竞争状态。而同期进入工业界的人，可能已经积累了财富、住房和职业资本。",[21,5486,5488],{"id":5487},"破局功利化读研","破局：功利化读研",[11,5490,5491,5492,3568],{},"在中国，虽然知识分子往往被冠以社会期待的光环，但是这也是一种负担和枷锁。请记住：我们是普通人尤其是出身平民家庭，我们首要目标是生存。在生存生计成为问题的时候，我们 ",[1364,5493,5494],{},"没有义务背负太多期待",[3222,5496,5497],{},[11,5498,5499],{},"沧浪之水清兮，可以濯吾缨；沧浪之水浊兮，可以濯吾足。",[11,5501,5502],{},"请卸下你的「学术羞耻感」。把研究生学历视为一份职业准入资格证，而非学术朝圣。学术职业是一种高风险选择，而不是所有人都必须承担的使命。",[11,5504,5505],{},"如何破局——功利化读研不失为一种出路。研究生必须思考：我读研的目的是为了什么？获得更好的就业门槛，暂时避开竞争激烈的就业市场，获得更多选择权。这完全是一种合理的人生规划。",[11,5507,5508],{},"首要的当然是毕业、混文凭，获得硕博的身份————这也是最现实、也最重要的一条。所以，在读研前期，你的一切目标是，必须以尽可能短的时间，完成最低毕业要求。大量水论文、蹭项目就足矣，不需要尽善尽美。科研嘛，也就那样，随便搞搞就行。",[11,5510,5511],{},"这是功利化读研真正的溢价所在。既然科研只求及格，那你必须把多出来的精力毫无愧疚地投入生存技能的构建。",[11,5513,5514],{},"在当前的环境下，「包装」比「做事」更重要，「数量」比「质量」更重要。不要把自己的全部人生价值绑定在学术成果上。科研如此，创业如此，艺术如此，很多长期主义事业都是如此。我不需要证明自己是英雄，我只需要把自己的人生过好。知识值得敬畏，但人的生命也值得敬畏。学术可以是人生的一部分，但不必成为人生的全部。",[11,5516,5517],{},"希望研究生们，不必神化科研、不必自我内耗、不必绑定学术理想，认清行业真相后，依然可以清醒活着、务实成长。",[11,5519,5520],{},"到最后，这种「虽千万人，吾往矣」，本身就有壮士断腕的秋风式悲凉，不是吗？😮‍💨",{"title":5247,"searchDepth":5248,"depth":5248,"links":5522},[5523,5524,5525,5526,5527,5528],{"id":5282,"depth":5248,"text":5282},{"id":5334,"depth":5248,"text":5335},{"id":5399,"depth":5248,"text":5399},{"id":5435,"depth":5248,"text":5435},{"id":5465,"depth":5248,"text":5465},{"id":5487,"depth":5248,"text":5488},"河南郑州",{},"2026-08-14","\u002Fblog\u002F2026\u002F2026-08-14-demystification-research",{"title":5271,"description":5276},"blog\u002F2026\u002F2026-08-14-demystification-research","自 2025 年入学至今，一年光阴悄然而逝，我已步入研二。回望这一年的研究生生活，我的心态与世界观经历了一场彻底而沉重的重塑——曾经的从容平实，如今已被挥之不去的失落感取代。",[5537],"thoughts","5MCFbs1-_AYzP3xpVDJ1BYVVmL8G250ZC9iJQ5yNkBM",1787072559759]