获得新的眼睛：学习如何让不可见之物变得可见

模糊裂开的那一刻#

任何试过学一门外语的人，都会认出那一刻。他们已经听了好几个星期，声音确实在那里——一道连绵不断、无疑属于人类的声流——但抵达时是质地，不是内容。温暖。在场。不透明。他们能听出里面有事情在发生，能听出那些有节奏的起落，听出它如何聚拢到一个点上，然后松开。但它就是不肯裂成片段。它整个是一团。

然后某一天，毫无征兆，也没有任何仪式，两个原本混作一团的声音，忽然分解成了两个截然不同的东西。舌尖卷起来——那是一个音。舌头抵住牙齿——那是另一个。它们不一样。它们从来就不一样，只是听者的听觉系统一直把它们当成一回事，像我们把十几种橄榄色一律折叠进“绿色”一样，折进同一个未分化的范畴里。而现在，那个差别听得见了，听者再也没法把它听回去，听这件事本身已经被永久改变。

这一刻——一个未分化的表面裂开、显出结构的那一刻——具体到有它自己特有的感受。它不像回忆。它也不像听懂一个逻辑论证。它更接近视觉：原本是噪声的东西，成了信号。被听的那个世界还是同一个世界。变的是它有多少能被登记下来。

更奇怪的是，这件事发生得如此频繁，又发生在如此多的地方。那个花了几个月只听得出“大概是同一个音”的乐手，忽然把半音听成了一道范畴的边——一个有内侧也有外侧的、清晰的台阶。那个盯了一年代码的人，在一段自己已经瞪了一小时的调用栈里，开始看见一种此前不曾察觉的结构——该在那里的函数调用，和不该在那里的那个，忽然被分开了。不是理解，不是推断，也不是演绎，而是看见​。那个东西变得可见了。它一直都在。变的是分辨率。

这就是学习从内部感受起来的样子。而它发生的频率，比我们注意到的高；发生的场所，比我们通常会去看的地方多；靠的是一个不管在哪里运作、都出奇一致的机制。

在词语之前：婴儿能听见什么#

刚才那一刻，最干净的经验案例，并不在我们预期会找到它的地方。它不在成人学习的研究里，不在音乐训练里，也不在技能发展的研究里。它在一个大约七个月大的婴儿的听觉经验里。

音位​（phoneme）是一门语言中能造成意义差别的最小声音单位：​pat 和 bat 之间只差一个音位，即浊唇塞音与清唇塞音之别。不同语言把“人类可能发出的声音”这个连续空间，切成不同的音位清单，把有意义的范畴边界画在不同的位置上。在声音这一层，英语之所以区别于印地语、汉语、伊博语，一部分就在于那些线画在哪儿。

1980 年代初，在不列颠哥伦比亚大学工作的 Janet Werker 和 Richard Tees 问了一个简单的问题：婴儿能不能听见自己母语并不作出的区分？他们用印地语的卷舌音／齿音对立，去测试正在学英语的七个月大婴儿——一个音是舌尖卷起抵住上腭发出的，另一个是舌头平贴上齿发出的。成年英语母语者无法可靠地分辨这两个音。而对印地语母语者来说，这个区分具有音位意义，他们听起来的清楚程度，就像我们听 p 和 b 一样。那些在英语家庭里长大、短暂的一生中一个印地语单词都没听过的婴儿，正确地分辨出了这组对立。¹

这就是那个让人愣住的发现。婴儿并没有额外学到任何东西。她经验更少，接触更少，听觉史更短。然而她听得见成年人听不见的区分。

到十到十二个月大时，同一个婴儿就再也听不见它了。¹ 那个在七个月时还可以被知觉到的印地语区分，变得无法企及。这个学英语的婴儿，她的听觉系统已经重组过，去匹配周围语言的范畴结构：在英语里承载意义的那些区分变锐利了；不承载意义的那些——印地语的这组对立、汉语的声调区分、南部非洲语言里的搭嘴音——则一起糊成了一个未分化的范畴，叫做“外语的声音”。

Werker 和 Tees 在措辞上很谨慎，而这份谨慎很重要：他们把这称为重组​，而不是丧失。人很容易只把它叫作丧失——在某种意义上，它确实是。婴儿失去了一项辨别能力。七月里还听得见的那个印地语区分，到十二月就听不见了。这是真实的减损。

但画面的另一半是获得。母语的音位边界，也就是那些承载英语意义的区分，变锐利了。婴儿的系统所做的，是把辨别能力从“在她的语言环境里不承载意义的对立”上撤走，重新分配到“承载意义的对立”上。这不是往容器里灌东西。这是一次再分配——一个知觉系统的分辨率被重新校准，朝着要紧的区分靠拢，从不要紧的区分上撤离。学习在这里不是加法，而是校准。¹

这场再分配的丧失那一侧，早在孩子学会第一个词之前就开始了——在她拥有任何语义内容、任何概念、任何陈述性知识之前——这正是它在理论上如此重要的原因。知觉重组并不是“知道了关于这门语言的一些事情”的后果。它先于那个。它是内容到来之前，基础设施先被调好。

这个故事的结尾有一点安慰，尽管只是部分的安慰。这扇窗不会永久封死。成年二语学习者在非母语音位对立上的困难有充分记录；试图听见一个自己的语言从没教过自己去听的区分，正是成年之后习得另一门语言的一部分含义。但高变异性语音训练——系统性地接触多位说话人、在许多不同语音环境中发出同一组目标对立——确实能显著改善成年人对此前无法企及的对立的辨别，而且收益可以泛化到新的说话人和新的刺激上。² 这项能力没有被废除；它只是变得更不情愿。系统的可塑性仍在，只是条件更硬了些。

婴儿这个案例，以它干净的发育形态显示出：学习是“什么能被登记下来”这件事的重组。变的是接收器本身——是它的敏感度结构，而不是它的内容。七个月大的婴儿，和十二个月大的婴儿，辨别能力不同。十二个月大的婴儿，和成年人，辨别能力也不同。在每个阶段，知觉上可能的事情都变了。而这个模式——经验重组“什么能被登记下来”——应该也能在非婴儿的学习者身上、在非语言的领域里被看到。

听得见时间的耳朵：音乐与被训练过的身体#

一位受过十一年训练的乐手，能在某个阈限上分辨音程，而未经训练的耳朵在同一个阈限上，仍然把它们融成一个大致的声音。

纽约大学的 Jean Mary Zarate 和同事招募了二十一名被试——其中十三位是乐手，正式训练年数平均为 11.5 年，八位是训练不足一年的非乐手——请他们分辨不同大小的音程。³ 问题在于：这些乐手只是多学了一些音程的名字，还是训练在音高空间本身的知觉结构上造成了一次范畴性的改变？答案是后者。乐手在恰好 100 音分处——一个半音，西方十二平均律里最小的一步——表现出显著更好的辨别力。非乐手则需要大于 125 音分的音程，才能可靠地觉察出差别。作者的表述很精确：半音也许代表着一道由音乐训练所诱发的、加在声学加工上的音程极限。在未经训练的耳朵还只有一条渐变斜坡的地方，被训练过的耳朵有了一道阈限。

这和婴儿显示出的是同一个结构，只不过运作在成年期。这些乐手并没有更低的纯音听阈——也就是说，不是在“绝对敏感度更高”这个笨拙的意义上耳朵更尖。他们拥有的，是一道被训练出来的范畴边界，是音高空间里的某个位置，知觉在那里从“接近”啪地跳到“不同”。阈限之下，声音融为一体。到了阈限处，它们分开。被训练过的知觉系统，拥有一个未经训练的系统所没有的特征。

音乐训练重塑的不只是音高知觉的范畴结构，还有听觉系统本身的时间分辨率。Prawin Kumar 和同事测量了间隔察觉阈限——两个音之间最小的、能被察觉为“间隔”而非“连续声”的静默时长——测量对象是十五位受过训练的声乐乐手和十五位非乐手。⁴ 乐手能察觉 1.81 毫秒的间隔，而非乐手需要 2.47 毫秒。差距是 27%，以毫秒计。另外三项时间指标——时长辨别、脉冲串时长辨别、频率差别阈限——也都显示出乐手的显著优势。被训练过的耳朵，是在更细的颗粒上读取时间。这不只是说乐手对所听到的东西拥有更多范畴；他们是在分辨那些未经训练的听觉系统当作一个连续瞬间来处理的时间事件。

乐手在间隔察觉上的优势，很容易被误认成某种更平常的东西。改善的并不是识别速度——不是“练得多了，察觉间隔更快了”。改善的是能不能登记下这个区分​：1.8 毫秒的间隔与 2.5 毫秒的间隔，未经训练的系统根本分不开。那个间隔一直都在。而听觉器官一旦被训练过，现在能看见它了。

如果耳朵可以这样被重新调音，身体也可以。

西安大略大学的 Jeremy Wong、Elizabeth Wilson 和 Paul Gribble 请被试用右手做够取动作——四百次够取，分布在十分钟里——然后测试他们的本体感觉敏锐度：闭上眼睛，报告自己的手在空间中位于何处的能力。⁵ 本体感觉敏锐度，也就是对肢体位置的内部感觉，在够取练习之后提高了约 11%（不确定度从 10.53 毫米降到 9.43 毫米）。也就是说：练过够取的被试，对自己的手在哪里，知道得更精确了。这个效应具有空间特异性——改善只出现在练习发生的那片工作空间里，在 25 厘米之外的位置上则没有。而且它要求主动参与：那些肢体被以完全相同的运动学轨迹被动带动的被试，没有表现出任何本体感觉上的改善。

想想这个练习的样子：一位被试闭着眼，被要求在食指越过一个记忆中的目标位置时给出指示。十分钟够取之前，误差大约一厘米。之后，稍小一些——但只在够取实际发生的那个特定空间位置上，也只在活动过的那条手臂上。身体关于自身的内部模型，颗粒变细了。不是笼统地变细。是在那个地方变细。

三个领域，全都不是视觉的。音位系统重组了它的范畴边界。听觉系统学会了以更小的增量读取时间。本体感觉系统在它练习过的那片空间里获得了精度。任何担心“分辨率”只是把一个视觉隐喻硬投射到别的领域上的读者，现在已经看到训练对时间性的耳朵和够取的手臂做了什么：两者都与视觉毫无关系，而两者看起来又完全是同一件事。

放射科医生和侍酒师究竟在做什么#

一张胸部 X 光片，在未经训练的眼睛看来，是一片灰色地景，中间亮着。肋骨拉出斜向的弧。心脏是那个显眼的大结构。肺是两侧那两大片更暗的区域。如果那里有一个小结节——左上象限某处一个直径一厘米的苍白圆影——未经训练的眼睛多半找不到它。这不是因为眼睛的光学分辨率不够。结节就在那儿，原则上可见，就像那个印地语音位区分原则上可听一样。但未经训练的视觉系统还没学会该找什么，而在不知道该找什么的时候，看，就只是在质地里瞎搜。

专家放射科医生并不只是读片更快。他们看见了未经训练的眼睛看不见的东西。Steven Waite 和同事在梳理放射学中的知觉专长文献时记录到：主治级放射科医生比新手更快注视到异常，且总注视次数更少。⁶ 在搜寻肺结节时，新手放射科医生的眼睛会被心脏吸引过去——那是胸片上视觉上最显著的结构；专家的眼睛则不会。更值得注意的是：对于高对比度病灶（也就是明显的发现），专家和新手表现相当。专长优势特别集中在低对比度、接近阈限的发现上：那些勉强算在那儿、只有被校准到能登记此类区分的眼睛才看得见的东西。⁶ 专家能察觉到新手根本无法登记的东西。病灶一直都在影像里。变的是什么能被看见。

专长里的这个知觉成分能多快被习得？Wanchen Chen 和同事以一种很尖锐的方式回答了这个问题。他们训练了 142 名毫无医学背景的被试去识别常规 X 光片上的髋部骨折，系统性地给他们看有骨折和没骨折的影像，并在每次判断后给出反馈。⁷ 持证放射科医生作为专家基准，准确率约为 90%。表现最好的五名新手——142 人里的前五名——在大约 52 分钟的纯知觉训练之后，达到了放射科医生水平的准确率，其间完全没有接受任何医学教育。⁷ 站得住的说法很克制：一小时不会把任何人变成放射科医生。它显示的是，放射学专长中的知觉成分，可以从围绕着它的医学知识中被明确地剥离出来——而这个可剥离的成分，在知觉禀赋最高的那些人身上，可以被极快地习得。

葡萄酒是一个不那么临床的领域，在某些方面也更有教益，因为人们对品酒专长的朴素直觉几乎是完全反的。多数人假定，让一个人成为品酒专家的，要么是更灵的鼻子——字面意义上对气味有更低的察觉阈限——要么是一份更大的风味描述词心理词库。Wendy Parr 和同事直接研究了这个假定，把十一位专家评酒师与十一位新手放在一整套嗅觉任务上作比较。⁸ 专家表现出显著更好的气味再认记忆——他们更可靠地认得出自己此前遇到过的气味。但他们对通用气味的原始嗅觉察觉阈限，与新手相当。两边收到的是同一个信号。专家拥有的，是把新手感知为一团未分化气味模糊的那些香气识别开、区分开的辨别能力。

有一个限定条件让画面更锐利：针对特定的葡萄酒相关化合物做训练——比如与黄油感相关的双乙酰，与木塞污染相关的乙基酚——确实能可测量地降低对这些特定化合物的察觉阈限。⁹ Sébastien Tempere 和同事测试了 201 位葡萄酒专业人士，发现受过正规训练的酿酒师在 5.0 微克／升就能察觉双乙酰，而未受训练的专业人士需要 16.6 微克／升——对于一个被训练过的目标而言，这是一个真实的阈限差异。⁹ 那个更干净利落的标题——“专家闻到的是同样的东西，只是分类方式不同”——是有化合物特异性的：对于他们训练所针对的那些化合物，专家也确实闻得到新手根本察觉不到的东西。

神经层面的证据把这个行为故事钉住了。一项追踪十二名侍酒师学员约十八个月的纵向研究记录到，训练期间嗅球体积出现了可测量的增长，而对照组没有显著变化。¹⁰ 侍酒师的右侧内嗅皮层厚度也增加了。大脑在组织层面，因应嗅觉训练改变了形状。行为上的辨别故事有了一个物理对应物。

两个领域。一位医生在读取肺组织的密度变化。一位侍酒师在读取发酵葡萄的化学复杂性。当我们看的是内部而不是外部时，这两个案例原来是同一件事：训练建立起了登记那些新手无法登记的区分的能力。案例在累积。

论点开始变得古怪的地方：情绪、因果、他人的脸#

到目前为止，论证走过的这些领域尽管彼此不同，却共有一个显眼的特征：它们都在某种常规意义上牵涉感官知觉。音位是声音。半音是声音。放射学是视觉的。葡萄酒是嗅觉的。就连本体感觉，身体对自身的内部感觉，也稳稳地属于感官家族。如果这个论点只是说“知觉训练会重新磨利感官系统”，那它虽然有趣，却是有界的——一个关于耳朵、鼻子和眼睛的故事。

那么，当我们去看那些显然没有现成感官框架可用的领域时，会发生什么？

设想这样两个人的差别：一个人能把忐忑、恐惧、焦虑和不安一一分开；另一个人眼里这些全是一回事——一种单一的、未分化的厌恶状态，体验为“难受”。这个现象学上的差别并不轻微。能作出这些区分的人，知道这份感受在要求什么。对某件具体未来事件的忐忑，要求的是计划。恐惧也许要求回避，也许要求与某件无法改变的事共处。焦虑——弥散的、没有对象的——要求的又是别的东西。作不出这些区分的人，对所有这些只有一种反应，因为在他那里，它们全是同一个东西。

Lisa Feldman Barrett 和同事做过一项为期 14 天的日记式研究，被试每天多次就九个情绪范畴评定自己的情绪体验。¹¹ 问题是：情绪颗粒度上的个体差异——人们在多大程度上对自己的负面情绪作出细颗粒的区分，而不是把它们堆进一个大范畴——能否预测调节行为上的差异？能。更高的负面情绪分化预测了更频繁、也更恰当的调节行为，而且这种关系在高情绪强度时最强——恰恰是在最需要调节、也最容易失败的时候。¹¹ 要紧的是这里的剂量—反应结构：更细的情绪辨别，不只是在容易的轻度情境里、而且在那些真正需要这件工具起作用的急性情境里，都能带来更契合语境的反应。

行为上的后果还不止于调节质量。Todd Kashdan 和同事发现，情绪分化程度更高的人，在被人伤害之后进行攻击性报复的可能性要低 20% 到 50%，这一结论基于 Rachel Pond 和同事的原始研究。¹² Kashdan 自己针对未成年饮酒者的工作——对 106 名被试所做的生态瞬时评估——发现情绪分化能预测压力条件下更少的暴饮。¹² 在高颗粒度个体身上，还记录到面对社交拒绝时更低的脑岛与前扣带回皮层活动。这项知觉能力在神经反应性上留下了痕迹。情感领域里更细的分辨率，在身体、行为和大脑上都产生了可测量的差异。

而且这项辨别能力是可训练的。Ekaterina Vedernikova 和同事做过一项为期 5 天的情绪知识干预：被试学习十二种具体情绪的定义、情境和例子——爱、喜悦、满足、宽慰、愤怒、厌恶、悲伤、孤独、恐惧、焦虑、羞耻、内疚——而对照组花同样的时间学习地理知识。¹³ 相对于对照组，干预组的负面情绪分化显著提高，效应量为中等，且在一个月后的随访中仍然保持。¹³ 把这些区分命名出来，就创造出了辨别它们的能力——这正是婴儿音位案例所显示的同一个结构：范畴一旦建立，就变得可以被知觉。

这个模式在相反的情感方向上同样成立。Michelle Tugade、Barbara Fredrickson 和 Lisa Feldman Barrett 在一项针对 130 名被试、为期 28 天的经验取样研究中发现，对正面情绪状态更细的辨别——更高的正面情绪颗粒度——预测了压力后更快的心血管恢复，以及反应性更低、更审慎的应对方式。¹⁴ 情感领域里的分辨率，无论负面还是正面，都在充当一件调节的工具。

而在这里，当领域既不是身体、也不是耳朵或鼻子，而是因果本身这个抽象范畴时，同样的签名又出现了。Benjamin Rottman、Dedre Gentner 和 Markus Goldwater 请物理科学、心理学和社会学的学生与教师，把对真实世界现象的描述归类。¹⁵ 物理科学家按因果结构归类——他们把捕食者—猎物的种群周期和经济的繁荣—萧条周期归为一类，因为两者都涉及负反馈回路，而不管其中一个属于生物学、另一个属于经济学。心理学和社会学的学生则按领域内容归类——他们把生物学现象和其他生物学现象归在一起，把经济学现象和其他经济学现象归在一起，因为那才是看得见的表面特征。物理科学家能看见横切各个领域的因果架构；新手只能看见领域标签。

因果结构在任何寻常意义上都是不可见的。一个负反馈回路不会在现象的表面上自报家门。它是可被知觉的——而且确实被那些受过训练去知觉它的人所知觉——但只在相关的区分被习得之后。新手知觉到的是内容和领域。专家知觉到的是底下的结构。两者之间的差别，一部分是知识问题；更大的一部分，是眼睛能登记下什么的问题。

同一个签名——某种未分化的表面经由训练裂开、成为可登记的结构——还出现在一个与语言、音乐、身体、葡萄酒、因果结构都没有明显共同之处的领域里。自 Adriaan de Groot 在 1960 年代的工作以来，国际象棋大师一直被格外仔细地研究。William Chase 和 Herbert Simon 在 1973 年确认了那个关键发现：在观看 3 到 4 秒之后，大师能回忆起一盘有意义的棋局中约 93% 的棋子位置；等级选手大约是 51%。¹⁶ 但这项优势的特殊之处，恰恰揭示了它的性质。当同样的棋子被随机摆放——摆成任何真实对局都不会产生的构型——大师的表现并不比初学者更好。¹⁶ 这项专长不是一般性的记忆优越。它是模式辨别：大师把有意义的局面知觉为一组被识别出来的构型，是一批作为整体被激发的组块；而初学者知觉到的是一堆单个的棋子。在新手只知觉到未分化的复杂性的地方知觉到有意义的结构——这正是我们一路追踪的那个分辨率提升签名。国际象棋是一个抽象的组合系统，既没有感官成分，也没有情感内容：同一个签名，出现在了全新的地界上。

在当前证据中，这个模式最远的延伸把我们带到脸上。具体说，是带到与自己不同的社会群体的人的脸上。有充分记录的异族效应——许多人在分辨自己所属群体之外的个体面孔时遇到的困难——通常不被表述为一个知觉分辨率问题，但研究提示，这样看也许有用。Sophie Lebrecht 和同事招募了二十名白人被试，分成两组：一组接受个体化训练——学会分辨八张特定的非裔美国人面孔，把每张脸与一个字母关联起来；另一组则接受训练，按种族对同样这些面孔进行归类。¹⁷ 两组在十天里接触的是相同的刺激，时长也相同。

训练之前，所有被试都表现出内隐种族偏见——在非裔美国人面孔之后对正面词汇的反应时更长，这是自动联结的一项标准测量。个体化训练之后，内隐偏见变得不显著。归类训练之后，它仍然存在。¹⁷ 个体化条件下的相关性很惊人：知觉层面异族效应的下降，预测了内隐偏见的下降（r² = 0.55）。¹⁷ 是辨别训练驱动了下游效应：学会在此前只有范畴归属的地方，登记下个体的变异。

样本很小（每种条件十人），这个发现需要更大规模的重复验证。它应当被当作社会领域里的一个概念验证，而不是已成定论的证据。但它的方向和机制已经足够清楚，足以把社会这个维度带入视野：学会知觉一个社会范畴内部的个体区分——在面孔知觉这一层上的分辨率提升——会改变下游的自动联结。知觉上的改变先行，并产生了社会认知上的改变。

同一个签名，既出现在七个月大婴儿的耳朵里，也出现在一位来自不同社会群体的陌生人的脸上——这不可能是巧合。它出现在棋手对棋盘的读取里、出现在科学家对因果结构的知觉里、出现在忐忑与恐惧与焦虑一齐涌来时仍能把它们分开的能力里，也同样不可能是巧合。这些领域在表面上毫无共同之处：不同的感官，不同的认知操作，不同的训练时间尺度。但结构是相同的。某种未分化的表面，经由训练，裂成了可登记的结构。问题已经转移了——不再是这件事会不会发生​，而是它是什么，又意味着什么​。

给这个机制命名#

我们一直在看的这件事，术语叫知觉学习​，Philip Kellman 和 Patrick Garrigan 把它定义为：由经验诱发的、知觉者从刺激中提取信息的方式上的改变。¹⁸ 这个定义很谨慎，而它的谨慎很重要。知觉学习不同于陈述性学习——知道关于这个领域的事实。它也不同于狭义的技能习得，即运动程序和程序性记忆。它是这样一个过程：经验改变的不是我们知道什么、也不是我们能做什么，而是我们能登记什么——知觉系统有多高的分辨率去从传入信号中提取东西。这个定义在实验室的语域里命名的，正是开篇场景在耳朵的语域里命名的同一件事：一直都在那儿的东西，变成知觉者可以登记的东西的那一刻。

Kellman 和 Garrigan 指出了两个标志性效应。¹⁸ 第一个是发现：学会究竟该知觉什么——哪些特征或关系承载信息，哪些变异有意义、哪些是噪声。这正是音位收窄所描述的：婴儿发现哪些声学变异在英语里承载意义。这也正是开篇场景所描述的：模糊裂成两个音位的那一刻。发现，就是某种质地变成对象的那一刻。

第二个是流畅：一旦被发现，相关模式的提取就更快、注意负荷更低。这正是放射科医生的眼动追踪所显示的——更快注视到病灶，总注视次数更少，注意力不被无关的显著性劫走。专家不需要有意识地搜索；模式是作为寻常的“看”的一部分被提取出来的。发现只发生一次；流畅是它此后从外部看起来的样子。

Robert Goldstone 在 1998 年梳理知觉学习文献时，指认出实现这些效应的四种成分操作。¹⁹

分化把此前无法区分的刺激分开——这是我们一直称为分辨率提升的那个核心机制。​单元化把多个元素合并成一个可察觉的单位，国际象棋大师把一个兵形知觉为一个组块而不是五枚单独的棋子，靠的就是它。​注意权重分配把注意力导向诊断上相关的特征、导离不相关的特征——这就是为什么专家放射科医生的眼睛不会被心脏吸引过去。​刺激印刻为频繁遇到的刺激建立起专门化的内部探测器——这个过程的物理对应物，也许就体现在侍酒师嗅球体积的增长上。

这四种操作是跨领域通用的：Goldstone 记录到它们在认知心理学、心理物理学、神经科学和发展心理学中运作，涉及的领域差异之大，涵盖阅读、面孔识别和科学分类。¹⁹ 这个机制并不是某一种感觉通道或某一类专长的专属属性。

数学课堂所展示的，也许是这个机制运作起来最引人注目的一次演示。Kellman 和同事为代数开发了知觉学习模块（PLM）——练的不是解方程，而是看见代数变换的结构，练的是识别哪一种变换动作适用于哪一类表达式。²⁰ 三十名学生完成了两次 35 到 40 分钟的 PLM 训练后，把解方程的时间从约 28 秒降到约 12 秒——降幅 57%——而在整个干预过程中，他们一道方程都没解过。²⁰ 他们练的是看。看的能力提高了。而解题这件他们没有练过的事，作为后果也提高了。收益在两周后的随访中仍然保持。

这就是机制统一性在感官之外的样子。同一种结构性操作——训练对相关特征的提取——在抽象的数学材料上产生的标志性效应（发现、流畅），与它在声学音位和放射影像上产生的完全相同。

接下来是测量的问题。如果这是一个跨领域通用的机制，我们怎么在“听觉音位辨别”和“葡萄酒中的嗅觉识别”这样差异巨大的领域之间比较它的效应？答案来自一个不是由认知科学家、而是由雷达工程师发展出来的框架，1966 年由 David Green 和 John Swets 引入心理物理学。²¹ 信号检测论给出一个叫做 d 撇——d′——的量，它刻画的是“信号存在时”与“只有噪声时”两种内部反应分布之间的标准化距离。d′ 为零意味着信号与噪声无法区分；d′ 越高，意味着两者相距越远、辨别越容易、系统越敏感。这个量与感觉通道无关：d′ 可以为一位读片的放射科医生、一位察觉音中间隔的被试、一位识别香气的葡萄酒专家，或一个在两类情绪体验之间作分辨的人计算出来。²¹

John Swets 在 1988 年发表于《Science》的一篇里程碑论文中证明，同一套相对操作特征框架适用于医学影像、材料检测、天气预报、信息检索、测谎和能力测验——把所有这些诊断系统放到了一把共同的、易于解读的尺子上。²² 这是一个测量层面的主张，而这个区分很重要。d′ 是那把同时适用于身高、铅笔长度和国土面积的英寸尺；它对三者都适用，并不因此让身高和铅笔长度成为同一类东西。Swets 确立的是：只要存在信号与噪声，就能计算 d′——这意味着我们可以把一位放射科医生、一位侍酒师和一位哀伤辅导师的辨别敏感度放在同一条形式化的轴上比较，而不必断言他们跑的是同一套认知程序。正是这种测量上的形式统一性，让本文有资格把辨别能力称作一项跨领域通用的属性，而不是一个被套用到非视觉领域上的视觉隐喻。它不是隐喻。它是一把尺。

机制这一锚点与测量这一锚点在此处分开，而这个区分很重要。Swets 告诉我们，我们可以跨领域比较敏感度。Kellman 和 Garrigan 告诉我们，为什么同一件事会不断发生：由经验诱发的信息提取方式的改变，带着同样的“发现—流畅”签名运作，在国际象棋、数学、语言、放射学和情感中产生分辨率提升。一个是测量的统一，另一个是机制的统一。两者都是真的。它们不是同一个主张。

受过认知科学训练的读者会问：这不就是模式识别吗？关于专长的标准解释认为，专家建立起庞大的模式库——组块、模板、图式——新刺激对照它们被更快、更可靠地匹配。国际象棋大师在长时记忆中携带着 5 万到 10 万个局面模式。放射科医生携带着各种病理的签名。模式识别这套解释有充分记录，而且是正确的。

模式识别是对行为输出的计算层描述。分辨率则是对“当那个区分终于变得可登记时，知觉基底发生了什么改变”的现象学描述——这里的现象学，取的是“这段经验从内部感受起来是什么样”的意思。它们是对同一过程的两个描述层次，而不是互相竞争的解释。本文之所以认真对待现象学这一层，是因为那份被感受到的经验，恰好命名了那些行为测量正在测量的东西。

音位这个案例最清楚地为这一步赢得了正当性。一位模式识别理论家会说，学英语的婴儿不过是在修剪未被使用的模板——印地语的那组对立因为得不到强化而脱落。但这个解释无法说明，为什么那个印地语区分变得在知觉上无法企及​，而不只是不再被使用：婴儿并不是把印地语模板留在原处然后忽略它；她们重组的是“什么能被登记下来”这件事本身，以至于那个区分对模板系统而言已经不再可得。辨别是从知觉者身上消失了，而不只是被分类器绕过了。分辨率这套说法能抓住这场重组；模板增删那套说法抓不住。两个层次都需要。

而现象学描述与行为描述并不对立。分辨率提高了的证据，​就是辨别行为：那只注视到病灶的眼睛，那只听得见半音的耳朵，那只知道自己身在何处的手。我们在九个领域里所做的，是找出同一个辨别性的行为签名，然后追问它从内部是什么。

为什么侍酒师读不了 X 光片#

看过同一个机制在九个领域里运作之后，自然的问题是：这种效应会迁移吗？如果音高辨别的训练磨利了听觉系统，那个被磨利的系统在分辨言语声音时会更好吗？如果多年的品酒训练精细化了专家的嗅觉辨别，这份精细化能用到检测燃气泄漏或诊断疾病上吗？

答案是，除极少数例外，不能。

Waite 等人的综述说得很明确：“放射科医生在训练过程中发展出来的知觉技能，局限于特定的放射影像知觉任务。事实上，放射科医生在执行非放射学搜寻任务时，并不比非放射科医生更好。”⁶ 那位学会了以多年才练出的精度在胸片上找到肺结节的专家放射科医生，面对一项自然场景搜寻任务时，表现和一个从没读过 X 光片的人一模一样。在 X 光片上被训练出来的知觉机器，不会泛化到别的视觉搜寻任务上。

本体感觉那项训练研究以更收敛的形式显示了同样的特异性：本体感觉敏锐度在练习发生的那个空间位置上提高了 11%，而在 25 厘米之外则毫无改善。⁵ 身体学会的是知道自己在那片工作空间里的位置，而不是笼统地知道。换个任务位置，改善就消失。换到另一个感觉系统，消失得更彻底。

音乐带来的时间分辨率——察觉 1.81 毫秒间隔的能力——并不赋予一种通用的时间加工优势。受过训练的乐手在非音乐情境中察觉短间隔的能力，并不比未受训练的听者更好。⁴

分辨率论点的边界就在这里。学习建立的是领域特异的分辨率，而不是通用的认知锐度。侍酒师读不了 X 光片。国际象棋大师预测股价的本事，并不比一个初学下棋的人更强。机制是统一的——同一种由经验诱发的信息提取方式的改变，在所有这些领域里运作——但这个机制的每一个实例都是局部的。分辨率是朝着训练发生的那个具体领域和任务校准的。这种局部性是这个机制的实际形状，而不是它的局限。

校准这个词很要紧。校准并不比泛化低一等；在某些方面，它是更精细的成就。一件校准到自身任务的仪器，其敏感度恰好配得上它所测量的东西——不至于粗到登记不下相关的区分，也不至于细到把无关的噪声也放大出来。侍酒师的嗅觉分辨率之所以调准在葡萄酒上，是因为品酒训练把它调准在了那里。放射科医生的视觉分辨率之所以配得上肺片，是因为肺片训练了它。每一个都是被自己所工作的领域塑造出来的仪器。这种校准不是偶然。它正是训练所产生的东西。限制与成就是同一件事。

而这也正好揭示了校准的代价。如果分辨率是朝着某个特定领域校准的，那它就可能失准​——被调准到某个任务上的方式，会在任务改变时、或者当高于任务所需的分辨率被施加到它身上时，产生特有的失败。

合适的分辨率，而不是最高的分辨率#

2008 年，牛津的 Merim Bilalić 和同事做了一项国际象棋研究，它的开头看起来平平无奇，像是一个直白的专长效应实验。²³ 他们给国际象棋专家——等级从候补大师到国际大师不等的棋手——看一些局面，其中包含两种可能的解法：一个是专家会迅速认出的、熟悉的五步将杀模式，另一个是不那么熟悉的三步将杀，而后者实际上更简单、也更高效。问题是：专家在找到那个好解法之后，还能不能找到更好的那个。

他们不能。当局面中存在一个熟悉的解法时，专家的表现急剧下降——下降幅度约为三个棋力标准差。²³ 国际大师与俱乐部级棋手之间的差距，是“在竞技赛场里泡了多年的人”与“下得不错但只是消遣”的人之间的差距——而一个熟悉却次优的模式的存在，把这个差距抹平了。这并不是因为那个三步解法有多隐晦。它更简单。它就摆在那儿。但五步模式先被激发了，而被激活的图式劫走了注意力，把它导离了包含更简单答案的那些棋盘区域。

配套的眼动追踪研究把这个机制变得可见。²³ 那些已经找到熟悉的五步解法的专家——而且在被问及时，报告说自己仍在寻找更好的解法——眼睛却继续注视着与五步将杀相关的格子。他们的眼睛并没有在搜索棋盘的其余部分。他们是在绕着第一个答案打转，被那个已经激活的图式牵着走，同时嘴上宣称自己对更好的东西持开放态度。从内部看，这感觉就像在搜索。扫视棋盘、留意、考量，这些主观体验是在场的，也是真诚的。缺席的是登记那些不属于活跃模式的格子的能力——它们就在那儿，视觉上可得，但图式已经替他们判定了它们不相关。专家并不是没有在看。专家是没能看见自己正在看的东西。

那个在有意义的局面上产出 93% 回忆率的国际象棋专长，与产出这种系统性盲视的，是同一个专长。收益与代价共用同一个机制：高分辨率的专家图式在它受过训练的地界上又快又准地激活，而在这样做的同时，它把注意力导向符合图式的东西，导离不符合的东西。分辨率提升与 Einstellung——心理学中指称这类定势效应的术语——是同一项能力，只是从棋盘上两个不同的位置去看。

国际象棋 Einstellung 在临床上的表亲，是偶发瘤（incidentaloma）。随着影像技术分辨率的提高——更高场强的 MRI、多层螺旋 CT、更细颗粒的乳腺摄影——机器如今能检出那些一直存在于组织中、却低于早期技术检出阈限的结构性变异。John O'Sullivan 和同事在《BMJ》上做了一项跨影像模态的偶发瘤患病率伞式综述：45% 的胸部 CT 会产出临床上无关的发现；CT 结肠镜为 38%；心脏 MRI 为 34%；脑与脊柱 MRI 为 22%。²⁴ 这些发现是真实的——那些结构性变异确实在那儿——但它们在临床上无关紧要。它们引发患者焦虑，并层层触发不必要的干预。影像仪器如今在辨别那些出于功能目的本不该被辨别的区分。

国际象棋的案例和偶发瘤的案例并不完全是同一类东西——一个是专家推理中的认知失败，另一个是机器造成的假象——但它们共有一条结构性原理。未朝任务校准的分辨率会产生特有的失败。国际象棋专家的分辨率是朝着找出棋类模式校准的；当任务要求不要顺着第一个被识别出来的模式走时，这份专家分辨率就成了负担。影像机器的分辨率是朝着找出结构性变异校准的；当临床任务要求区分重要的与无关的结构性变异时，这份分辨率产出的就是看起来像信号的噪声。

Erik Dane 在《Academy of Management Review》的一篇论文中提出认知固化这个概念，用来命名这一模式在认知型专家身上的版本：领域图式高度稳定，这让它们在处理标准问题时高效，却在新颖问题需要不同框架时难以调整。²⁵ 随着专长加深，图式变得更稳定——更快，更可靠，也更难被推翻。那位看不见简单解法的专家并不是没在思考。他们正在自己那个已成为主要工具的图式里以极高的效率思考，而恰恰是这份效率，阻止了那一步横向的移动。

放射学差错文献给这个模式添上了数量上的分量。日常工作中的专家放射科医生，在约 3% 到 5% 的检查中会犯实时差错；而对病例的回顾性复审——根据后来的诊断结果回头再看影像——显示，事后可见的病灶中约有 30% 在实时判读时被漏掉了。²⁶ 所有放射学差错中，有六成到八成是知觉性的而非认知性的：差错不是从看见的发现中作出了错误推断，而是漏掉了本就在那里的发现。²⁶ 在后来被确诊肺癌的患者中，回顾性复审此前那些“正常”的胸片，发现癌症在高达 90% 的病例中是可见的。高分辨率并不等于无差错。校准——而不是单靠分辨率——才是检验的标准。

专长发展中那份被感受到的经验——本文开篇所写的那种现象学，模糊裂成特征的那一刻——在环境不规则、反馈被污染时，可以在辨别力并未真正改善的情况下发生。Kahneman 和 Gary Klein 在他们 2009 年关于直觉专长成立条件的论文中确立：真正的专长需要两样东西，一个规则到足以被预测的环境，以及长期清晰而诚实的反馈。²⁷ 在国际象棋、放射学、语言、葡萄酒、音乐——也就是本文收集证据的那些领域——环境足够规则，反馈足够清晰，分辨率会随训练可靠地提升。而在长周期的政治预测、金融市场预测、以及某些结果被延迟或被混淆的临床诊断中——分辨率提升的那份现象学可以发展出来，而底下的辨别能力却没有。被感受到的经验不是一个可靠的标记。算数的是朝着领域结构的校准，而不是那份“我知道”的感觉有多笃定。

这些都没有推翻主要论证。分辨率可能失准、最高分辨率并不是目标、专家的地图可能变成约束——这些是从论点中推出的后果，而不是对它的反驳。一个能预测自身病理的论点更稳健，而不是更脆弱。诚实版本的主张是：学习是分辨率的提升，朝着训练发生的那个任务校准，运作在足够规则的环境中，并有足够清晰的反馈来真正改善辨别力。这是这个论点诚实的形状，而它因为所让出的东西而更强。

边界：这个论点预测了什么，又尚未证明什么#

前面那九个领域——音位、音乐、时间、医学、嗅觉、本体感觉、情绪、因果、社会——构成了证据实际支持的部分。若说分辨率提升是每一个领域中学习的机制，那就说过头了，而本文并没有这样说。

这个论点的预测既有意思，也仍然开放。亚里士多德意义上的 phronesis​——实践智慧，即知觉一个情境在道德上要求什么的能力——如果论点延伸到那里，就会预测道德识别遵循同样的分辨率结构：登记那些没有经验的道德知觉者无法区分的、道德上显著的特征的能力。如果审美专长显示出同样的结构，它会预测传统内部的结构性辨别——爵士乐手对一次和声替换的耳朵，而新手只听见了一次和弦变化；艺术评论者对构图张力的知觉，而美术馆访客体验到的只是笼统的品质感或一丝说不清的不适。这个论点预测——但并未证明——政治知觉是在没有经验的观察者只看见个体行动的地方看见结构性权力的能力，而元认知技能则是知觉自身不确定性地图上的种种区分的能力，是“笃定的无知”与“经过校准的不确定”之间的差别，经由练习而变得可见。

这些都是开放的问题：论点可以在这些地方被检验，可能在这些地方被证伪，也在这些地方仍然真正悬而未决。它们没有一个具备音位收窄研究或放射学训练研究那样的经验地位。它们是预测，不是证据。

如果专长主要是一个知觉现象——如果训练建立起来的不是更大的事实库存，而是一个分辨率更高的知觉系统——那么围绕结构化辨别练习来组织的教学法，值得比它通常得到的更多关注。那是一条留给别人去追的线索。

还有什么正在进入焦点#

同一个机制还有一个更大的版本。

想想 1870 和 1880 年代，细菌理论对医学做了什么。传染性因子与非传染性物质之间的区分，在经验上一直在场——一直在数据里，在疾病传播的模式里，在不同卫生条件下的不同结局里。Koch 和 Pasteur 之前的医学无法登记它。概念上的和仪器上的装置都还不在。临床表现是可见的；因果结构是不可见的。细菌理论提供的，主要并不是新事实，尽管它确实产出了很多新事实。它提供的是因果维度上的一种新分辨率：一个此前的框架无法使之可被知觉的区分，忽然变得无法忽视。

反馈回路这个概念，为理解生物系统、经济系统和社会系统做了类似的事。负反馈——一个系统偏离设定点会产生纠正力的那种属性——一直存在于种群动力学中，存在于市场价格中，存在于激素调节中，存在于恒温器的设计中。Rottman 关于专家科学家的发现，不只是关于个别科学家分类行为的事实：它是关于一个科学传统对其从业者的知觉做了什么的事实。¹⁵ 一门学科在成熟的过程中，教会它的成员去知觉新手看不见的因果结构。个体的分辨率升级，乘以一个研究者共同体、乘以数十年，从内部看就是一次科学范式转移。

婴儿的音位收窄，并不是一件局限于生命头一年的发育奇观。它是一个在学习发生的每一个尺度上反复出现的结构的最早、最干净的实例。一个婴儿的听觉系统重组它的范畴边界，去匹配周围的语言。一门学科重组它的概念边界，去匹配那些已被证明在经验上多产的因果结构。一种文化获得了知觉前几代人无法登记的区分的能力——并且在这个过程中，失去了知觉此前某些东西的能力。

诸如机会成本、熵、自然选择、相对风险这样的概念升级，是文明尺度上的分辨率提升。不是发现了一个新东西，而是有了一个新区分，它一旦变得可被知觉，就让此前不可见的结构变得无法忽视。人类探究的集体装置，就是一台用来获得这类新视力的机器。婴儿和学科跑的是同一个过程。

在自然科学中，在社会科学中，在我们持续尝试理解如何共同生活的过程中——此刻还处在集体辨别阈限之下的，有哪些东西只需要再多一点分辨率，就会变得无可否认？数据里在场着哪些因果结构，经验里在场着哪些道德区分，会被未来的某一代人视为像疾病的细菌理论一样显而易见——而在有人造出看见它的装置之前，又像那个理论一样不可见？我们不知道，因为我们在阈限之下。此时此刻，我们所处的位置，正是那个听着印地语的、学英语的七个月大婴儿——全套区分仍然可得，尚未被收窄成我们叫得出名字的那些。

延伸阅读#

• Kellman, P.J. "Perceptual Learning." 收于 Stevens' Handbook of Experimental Psychology​（第 3 版），第 1 卷。Wiley，2002。——关于知觉学习研究广度的背景；为“机制”一节提供了依据，并帮助把发现与流畅框定为两个不同的功能范畴。
• Werker, J.F. & Tees, R.C. (2005). Speech perception as a window for understanding plasticity and commitment in language acquisition. Developmental Psychobiology, 46​, 233-251.——对此后数十年间音位收窄研究的综述；是 E01–E02 那组展示的相关背景；未直接引用，以便让第 2 节聚焦于 1984 年的原始研究。
• Maurer, D. & Werker, J.F. (2013). Perceptual narrowing during infancy: A comparison of language and faces. Developmental Psychobiology, 56​, 154-178.——把音位收窄的框架扩展到面孔知觉；是 E27（Lebrecht）那则社会领域展示的相关背景；备用证据。
• Barrett, L.F. How Emotions Are Made: The Secret Life of the Brain​. Houghton Mifflin Harcourt，2017。——关于情绪颗粒度与建构情绪理论的背景；为第 5 节的情绪领域展示提供了依据；也提供了概念背景，说明为什么情绪辨别被框定为知觉性的而非认知性的。
• Ericsson, K.A. & Pool, R. Peak: Secrets from the New Science of Expertise​. Houghton Mifflin Harcourt，2016。——关于专长习得的主流通俗论述；本文明确没有引用它，因为它那套行为／刻意练习的框架，是本文所要补充的（而不是与之竞争的）。它在背景上塑造了本文该处理什么、以及该把什么留给其他论述。
• Horowitz, A. On Looking: Eleven Walks with Expert Eyes​. Scribner，2013。——现有最接近的通俗论述；它展示了这个现象，却没有为机制提供理论；是本文那道空隙的背景框架。
• Kahneman, D. Thinking, Fast and Slow​. Farrar, Straus and Giroux，2011，第 21–22 章（直觉对公式；专家直觉：我们何时可以信任它？）。——第 8 节吸收 Kahneman／Klein 有效性条件的背景；也是读者占主导地位的认知框架。

Gaining New Eyes — How Learning Makes the Invisible Visible

The Moment the Blur Breaks#

Anyone who has tried to learn a language will recognize the moment. They have been listening for weeks, and the sounds are there — a continuous stream of something unmistakably human — but arriving as texture, not content. Warm. Present. Opaque. They can hear that things are happening in it. They can hear the rhythmic falls and rises, the way it gathers to a point and then releases. But it is not separating into pieces. It is all one thing.

And then one day, suddenly and without ceremony, two sounds that had been a single blurred phenomenon resolve into two distinct objects. The tongue curling back — that is one sound. The tongue at the teeth — that is another. They are not the same. They were never the same, but the listener's auditory system had been treating them as if they were, folding them into a single undifferentiated category the way we fold a dozen shades of olive into simply green​. And now the difference is audible, and the listener cannot unhear it, and the listening has permanently changed.

This moment — the moment when an undifferentiated surface breaks into structure — is so specific that it has a characteristic feeling. It is not like remembering. It is not like understanding a logical argument. It is closer to vision: something that was noise has become signal. The world being listened to is the same world. What has changed is what can be registered of it.

The stranger thing is how often this happens, and in how many places. The musician who spent months hearing roughly the same note suddenly hears a semitone as a categorical edge, a distinct step with an inside and an outside. The person who has been looking at code for a year begins to see, in a stack trace they've stared at for an hour, a structure they hadn't perceived before — the function call that belongs and the one that doesn't, suddenly differentiated. Not understood, not inferred, not deduced: seen​. The thing became visible. It was always there. The resolution changed.

This is what learning feels like from the inside. And it happens more often than we notice, in more places than we typically look, by a mechanism that turns out to be surprisingly consistent wherever it operates.

Before the Word: What the Infant Can Hear#

The cleanest empirical case for what just happened in that moment is not where we would expect to find it. It is not in a study of adult learning, or musical training, or the development of skill. It is in the auditory experience of an infant who is approximately seven months old.

A phoneme is the smallest unit of sound that makes a meaningful difference in a language: the difference between pat and bat is a single phoneme, the voiced versus unvoiced labial stop. Different languages carve up the continuous space of possible human sounds into different phoneme inventories, drawing their meaningful categorical boundaries in different places. What distinguishes English from Hindi from Mandarin from Igbo, at the level of sound, is partly a matter of where those lines are drawn.

Janet Werker and Richard Tees, working at the University of British Columbia in the early 1980s, asked a simple question: can infants hear distinctions that their language doesn't make? They tested English-learning seven-month-old infants on the Hindi retroflex/dental distinction — the difference between a sound made with the tongue curled back against the palate, and a sound made with the tongue placed flat against the upper teeth. Adult English speakers cannot reliably distinguish these two sounds. Hindi speakers, for whom the distinction is phonemically meaningful, hear them as clearly different as we hear p and b​. The infants, raised in English-speaking homes, having heard not a single word of Hindi in their brief lives, correctly discriminated the contrast.¹

This is the finding that stops people. The infant has not learned anything additional. She has had less experience, less exposure, fewer months of auditory history. And yet she hears a distinction the adult cannot.

By ten to twelve months, the same infant could no longer hear it.¹ The Hindi distinction, which was perceptually accessible at seven months, had become inaccessible. The English-learning infant's auditory system had reorganized itself to match the category structure of the ambient language: the distinctions that carry meaning in English had sharpened; the distinctions that do not — the Hindi contrast, the Mandarin tonal distinctions, the clicks of southern African languages — had blurred together into an undifferentiated category of foreign sound​.

Werker and Tees were careful in their framing, and that care matters: they described this as reorganization​, not loss. It is tempting to call it only loss — and in one sense, it is. The infant has lost a discrimination capacity. The Hindi distinction that was audible in July is not audible in December. That is a real diminishment.

But the other half of the picture is gain. The native phoneme boundaries, the distinctions that carry English meaning, have sharpened. What the infant's system has done is redistribute its discrimination capacity away from contrasts that don't carry meaning in her linguistic environment and toward contrasts that do. This is not the filling of a vessel. It is a reallocation — the resolution of a perceptual system being recalibrated toward the distinctions that matter, and away from those that don't. Learning, here, is not addition. It is calibration.¹

The fact that this redistribution begins, on its loss side, before a child has learned a single word — before she has any semantic content, any concepts, any declarative knowledge — is what makes it so theoretically important. The perceptual reorganization is not a consequence of knowing things about the language. It is prior to that. It is the infrastructure being tuned before the content arrives.

There is a reassurance at the end of this story, though it is a partial one. The window does not seal permanently. Adult second-language learners face well-documented difficulty with non-native phoneme contrasts; the experience of trying to hear a distinction the language hasn't taught us to hear is part of what it means to acquire another language past childhood. But high-variability phonetic training — systematic exposure to multiple talkers producing the same target contrast across many different phonetic environments — produces significant improvement in discrimination of contrasts that were previously inaccessible to adult ears, with gains that generalize to novel speakers and stimuli.² The capacity is not abolished; it becomes more reluctant. The system's plasticity persists, if under somewhat stiffer conditions.

What the infant case shows, in its clean developmental form, is that learning is the reorganization of what can be registered. The receiver itself has changed — its sensitivity structure, not its contents. The seven-month-old has different discrimination capacity than the twelve-month-old. The twelve-month-old has different discrimination capacity than the adult. At each stage, what is perceptually possible has changed. And this pattern — experience reorganizing what can be registered — should be visible in learners who are not infants, and in domains that are not language.

The Ear That Hears Time: Music and the Trained Body#

A musician with eleven years of training discriminates pitch intervals at a threshold where an untrained ear still fuses them into a single approximate sound.

Jean Mary Zarate and colleagues at NYU recruited twenty-one subjects — thirteen of them musicians with a mean of 11.5 years of formal training, eight non-musicians with less than a year — and asked them to discriminate pitch intervals of different sizes.³ The question was whether the musicians had simply learned more interval names, or whether training had produced a categorical change in the perceptual structure of the pitch space itself. The answer was the latter. Musicians showed significantly improved discrimination at exactly 100 cents — one semitone, the smallest step in the Western chromatic scale. Non-musicians required intervals larger than 125 cents before they could reliably perceive a difference. The authors put it precisely: the semitone may represent a musical training-induced intervallic limit to acoustic processing. The trained ear has a threshold where the untrained ear has a gradient.

This is the same structure the infant showed, operating in adulthood. The musicians did not have lower pure-tone audiometric thresholds — sharper hearing in the blunt sense of greater absolute sensitivity. What they had was a trained categorical boundary, a place in pitch space where perception snaps from close to different​. Below the threshold, sounds fuse. At the threshold, they separate. The trained perceptual system has a feature the untrained one does not.

Musical training reshapes not only the categorical structure of pitch perception but the temporal resolution of the auditory system itself. Prawin Kumar and colleagues measured gap detection thresholds — the smallest silent gap between two tones that is perceptible as a gap rather than a continuous sound — in fifteen trained vocal musicians and fifteen non-musicians.⁴ Musicians detected gaps of 1.81 milliseconds where non-musicians required 2.47 milliseconds. The difference is 27%, measured in milliseconds. Three other temporal measures — duration discrimination, pulse-train duration discrimination, and frequency differential limen — all showed significant musician advantages. The trained ear is reading time at finer grain. It is not simply that musicians have more categories for what they hear; they are resolving temporal events that the untrained auditory system is treating as one continuous moment.

The musician's advantage in gap detection is easy to mistake for something more ordinary. What improved was not speed of identification — not practice at listening for gaps that have gotten faster to detect. What improved was the capacity to register a distinction at all: a 1.8 ms gap versus a 2.5 ms gap, which the untrained system cannot separate. The gap was always there. The auditory apparatus, once trained, can now see it.

And if the ear can be retuned in this way, so can the body.

Jeremy Wong, Elizabeth Wilson, and Paul Gribble at the University of Western Ontario asked subjects to perform a reaching movement with their right hand — four hundred reaching movements, spread across ten minutes — and then tested their proprioceptive acuity: their ability, with eyes closed, to report where their hand was in space.⁵ Proprioceptive acuity, the internal sense of limb position, improved by approximately 11% after the reaching practice (from 10.53 mm uncertainty to 9.43 mm uncertainty). That is: the subjects who had practiced reaching knew, with greater precision, where their hand was. The effect was spatially specific — the improvement was present only in the workspace where practice occurred, not at a location 25 centimeters away. And it required active engagement: subjects whose limbs were moved passively through identical kinematics showed no proprioceptive improvement.

Consider the exercise: a subject, eyes closed, asked to indicate when the index finger crosses a remembered target position. Before ten minutes of reaching, the error is about a centimeter. After, slightly less — but only in the specific spatial location where reaching happened, and only in the arm that was active. The body's internal model of itself has become finer-grained. Not in general. In that place.

Three domains, all non-visual. The phonemic system reorganizes its categorical boundaries. The auditory system learns to read time at smaller increments. The proprioceptive system gains precision in the space where it practiced. Any reader who worried that "resolution" was just a visual metaphor projected onto other domains has now seen what training does to the temporal ear and the reaching arm: neither has anything to do with vision, and both look exactly like the same thing.

What the Radiologist and the Sommelier Are Doing#

A chest X-ray, to an untrained eye, is a gray landscape with a bright center. The ribs make diagonal arcs. The heart is the obvious large structure. The lungs are the large darker areas on either side. If there is a small nodule — a pale circular shadow a centimeter in diameter somewhere in the upper left quadrant — the untrained eye is unlikely to find it. Not because the eye doesn't have the optical resolution. The nodule is there, visible in principle, just as the Hindi phoneme distinction was audible in principle. But the untrained visual system has not yet learned what to look for, and without knowing what to look for, looking is searching through texture.

Expert radiologists do not simply read films faster. They see what the untrained eye cannot. Steven Waite and colleagues, reviewing the perceptual expertise literature in radiology, documented that attending radiologists fixate on abnormalities faster and produce fewer total fixations than novices.⁶ The novice radiologist's eye is drawn to the heart — the most visually salient structure on a chest X-ray — when searching for lung nodules; the expert's eye is not. More strikingly: for high-contrast lesions (obvious findings), experts and novices perform comparably. The expertise advantage concentrates specifically on low-contrast, near-threshold findings: findings barely there, visible only to an eye calibrated to register such distinctions.⁶ Experts detect what novices cannot register at all. The lesion was always in the image. What changed was what could be seen.

The question of how fast this perceptual component of expertise can be acquired is one that Wanchen Chen and colleagues answered in a pointed way. They trained 142 medically naïve participants to identify hip fractures in conventional radiographs, exposing them systematically to images with and without fractures and giving them feedback after each identification.⁷ Board-certified radiologists served as the expert benchmark, achieving approximately 90% accuracy. The top five performing novices — the top five out of 142 — matched radiologist-level accuracy after approximately 52 minutes of purely perceptual training, involving no medical education whatsoever.⁷ The defensible claim is careful: an hour does not make anyone a radiologist. What it shows is that the perceptual component of radiological expertise is demonstrably separable from the medical knowledge that surrounds it — and that this separable component can be acquired remarkably quickly by those with the highest perceptual aptitude.

Wine is a less clinical domain and in some ways a more instructive one, because the naive intuition about wine expertise is almost exactly backwards. Most people assume that what makes a wine expert is either a more sensitive nose — literally a lower detection threshold for odors — or a larger mental lexicon of flavor descriptors. Wendy Parr and colleagues studied this assumption directly, comparing eleven expert wine judges with eleven novices across a battery of olfactory tasks.⁸ The experts showed significantly better odor recognition memory — they were reliably better at recognizing odors they had previously encountered. But their raw olfactory detection thresholds for generic odors were comparable to the novices'. The same signal was arriving in both cases. What the experts had was the discriminative capacity to recognize and differentiate scents that the novices perceived as one undifferentiated odorous blur.

One qualification sharpens the picture: training on specific wine-relevant compounds — diacetyl, associated with a buttery quality; ethylphenols, associated with cork taint — does measurably lower detection thresholds for those specific compounds.⁹ Sébastien Tempere and colleagues testing 201 wine professionals found that formally trained enologists detected diacetyl at 5.0 micrograms per liter where untrained professionals required 16.6 — a real threshold difference for a trained target.⁹ The cleaner headline — "experts smell the same things but classify them differently" — is compound-specific: for the compounds their training targets, experts also smell what novices cannot detect at all.

The neural evidence ties the behavioral story down. A longitudinal study following twelve sommelier trainees over approximately eighteen months documented measurable increases in olfactory bulb volume during the training period, with no significant change in controls.¹⁰ Right entorhinal cortex thickness also increased in the sommeliers. The brain, at the tissue level, changed shape in response to olfactory training. The behavioral discrimination story has a physical correlate.

Two domains. A physician reading the density variations of lung tissue. A sommelier reading the chemical complexity of fermented grape. Both cases turn out to be the same thing when we look at the inside rather than the outside: training builds the capacity to register distinctions the novice cannot. The case accumulates.

Where the Thesis Gets Strange: Emotion, Causation, Other Faces#

The argument so far has moved through domains that, however different, share an obvious feature: they all involve sensory perception in some conventional sense. Phonemes are sounds. Semitones are sounds. Radiology is visual. Wine is olfactory. Even proprioception, the body's internal sense of itself, belongs comfortably to the sensory family. If the thesis were only that perceptual training resharpens sensory systems, it would be interesting but bounded — a story about the ear and the nose and the eye.

What happens when we look at domains where the sensory framing is not obviously available?

Consider the difference between a person who can distinguish apprehension from dread from anxiety from unease, and a person for whom all of these are one thing: a single aversive undifferentiated state experienced as bad​. The phenomenological difference is not trivial. The person who can make these distinctions knows what the feeling is calling for. Apprehension about a specific future event calls for planning. Dread may call for avoidance or for sitting with something that cannot be changed. Anxiety — diffuse, object-less — calls for something different again. The person who cannot make these distinctions has only one response to all of them, because they are all, to them, the same thing.

Lisa Feldman Barrett and colleagues ran a 14-day diary protocol in which participants rated their emotional experience across nine emotion categories every day, multiple times a day.¹¹ The question was whether individual differences in emotional granularity — the degree to which people made fine-grained distinctions among their negative emotions rather than lumping them into one large category — predicted differences in regulatory behavior. It did. Higher negative emotion differentiation predicted more frequent and appropriate regulatory behavior, with the relationship strongest at high emotional intensity — precisely when regulation is most needed, and most likely to fail.¹¹ The dose-response structure is what matters: finer emotional discrimination enables more context-appropriate response not just in mild situations where it is easy, but in the acute ones where the instrument needs to work.

The behavioral consequences reach beyond regulation quality. Todd Kashdan and colleagues found that people with higher emotion differentiation were 20 to 50% less likely to retaliate aggressively against someone who had hurt them, based on primary research by Rachel Pond and colleagues.¹² Kashdan's own work with underage drinkers — ecological momentary assessment across 106 participants — found that emotion differentiation predicted less binge drinking under stress conditions.¹² Lower insula and anterior cingulate cortex activity in response to social rejection has been documented in high-granularity individuals. The perceptual capacity leaves a trace in neural reactivity. Finer resolution in the affective domain produces measurable differences in body, behavior, and brain.

And this discrimination capacity turns out to be trainable. Ekaterina Vedernikova and colleagues ran a 5-day emotion knowledge intervention in which participants learned definitions, situational contexts, and examples for twelve specific emotions — love, joy, satisfaction, relief, anger, disgust, sadness, loneliness, fear, anxiety, shame, guilt — while a control group spent the same time learning geography facts.¹³ Negative emotion differentiation increased significantly in the intervention group relative to the control, with a medium effect size, and the effect was maintained at one-month follow-up.¹³ Naming the distinctions created the capacity to discriminate them — the same structure the infant phoneme case showed: the category, once established, becomes perceivable.

The pattern holds in the opposite affective direction. Michelle Tugade, Barbara Fredrickson, and Lisa Feldman Barrett found that finer discrimination of positive emotional states — higher positive emotional granularity — predicted faster cardiovascular recovery from stress and less reactive, more deliberate coping in a 28-day experience-sampling study of 130 participants.¹⁴ Resolution in the affective domain, negative and positive both, functions as a regulatory instrument.

And here, where the domain is not the body or the ear or the nose but the abstract category of causation itself​, the signature appears again. Benjamin Rottman, Dedre Gentner, and Markus Goldwater asked students and faculty in the physical sciences, psychology, and sociology to sort descriptions of real-world phenomena into categories.¹⁵ The physical scientists sorted by causal structure — they grouped a predator-prey population cycle with an economic boom-bust cycle because both involve negative feedback loops, regardless of whether one concerns biology and the other economics. Psychology and sociology students sorted by domain content — they grouped the biological phenomenon with other biological phenomena and the economic phenomenon with other economic phenomena, because that is the visible surface feature. The physical scientists could see the causal architecture cutting across domains; the novices could only see the domain labels.

Causal structure is not visible in any ordinary sense. A negative feedback loop does not announce itself on the surface of the phenomenon. It is perceivable — and perceived, by those with the training to perceive it — but only after the relevant distinction has been acquired. The novice perceives content and domain. The expert perceives the structure underneath. The difference between them is partly a matter of knowledge; the larger part is a matter of what the eye can register.

The same signature — something that was undifferentiated surface breaks, through training, into registrable structure — appears in a domain that shares nothing obvious with language, or music, or bodies, or wine, or causal structure. Chess grandmasters have been studied with particular care since Adriaan de Groot's work in the 1960s. William Chase and Herbert Simon confirmed the key finding in 1973: grandmasters recall roughly 93% of piece positions from a meaningful game board after 3 to 4 seconds of exposure; class-level players recall around 51%.¹⁶ But the advantage is specific in a way that reveals its character. When the same pieces are arranged randomly — in configurations that no actual game would produce — the grandmasters perform no better than the beginners.¹⁶ The expertise is not a general memory superiority. It is pattern discrimination: the grandmaster perceives the meaningful position as a set of recognized configurations, chunks that fire as units, where the beginner perceives a collection of individual pieces. This perception of meaningful structure where the novice perceives undifferentiated complexity is exactly the resolution-increase signature we have been following. Chess is an abstract combinatorial system, with no sensory component and no affective content: the same signature, in entirely new territory.

The furthest extension of this pattern in the current evidence takes us to faces. Specifically, to the faces of people from a social group different from one's own. The well-documented other-race effect — the difficulty many people experience in distinguishing individual faces from a social group other than their own — is not typically framed as a perceptual resolution problem, but the research suggests it might usefully be. Sophie Lebrecht and colleagues recruited twenty Caucasian participants and divided them into two groups: one trained to individuate — to discriminate between — eight specific African American faces, learning to associate each face with a letter; the other trained to categorize the same faces by race.¹⁷ Both groups were exposed to the same stimuli for the same amount of time over ten days.

Before training, all participants showed implicit racial bias — longer response times to positive words following African American faces, a standard measure of automatic association. After individuation training, implicit bias became non-significant. After categorization training, it remained.¹⁷ The correlation in the individuation condition was striking: reduction in the perceptual other-race effect predicted reduction in implicit bias (r² = 0.55).¹⁷ The discrimination training drove the downstream effect: learning to register individual variation where before there had been only category membership.

The sample is small (ten participants per condition), and the finding requires replication at larger scale. It should be held as a proof-of-concept in the social domain rather than as settled evidence. But its direction and mechanism are clear enough to bring the social dimension into view: learning to perceive individual distinctions within a social category — resolution increase at the level of face perception — changes automatic associations downstream. The perceptual change precedes and produces the social-cognitive one.

The same signature in the ear of the seven-month-old and in the face of a stranger from a different social group — this cannot be coincidence. Neither can its appearance in the chess master's reading of a board, in the scientist's perception of causal structure, in the capacity to distinguish apprehension from dread from anxiety when all of them arrive at once. The domains share nothing on the surface: different senses, different cognitive operations, different timescales of training. But the structure is identical. Something that was undifferentiated surface breaks, through training, into registrable structure. The question has shifted — no longer does this happen but what is it, and what does it mean​.

Naming the Mechanism#

The term for what we have been watching is perceptual learning​, defined by Philip Kellman and Patrick Garrigan as experience-induced changes in the way perceivers extract information from stimulation.¹⁸ The definition is careful, and its care matters. Perceptual learning is distinct from declarative learning — knowing facts about the domain. It is also distinct from skill acquisition in the narrow sense of motor programs and procedural memory. It is the process by which experience changes not what we know or what we can do, but what we can register — what the perceptual system has the resolution to extract from the incoming signal. What this definition names, in the register of the laboratory, is the same thing the opening scene named in the register of the ear: the moment when what was there all along becomes something the perceiver can register.

Kellman and Garrigan identify two signature effects.¹⁸ The first is discovery: learning what to perceive at all — which features or relations carry information, which variations are meaningful and which are noise. This is what the phonemic narrowing described: the infant discovering which acoustic variations carry meaning in English. It is also what the opening scene described: the moment the blur broke into two phonemes. Discovery is the moment when something that was texture becomes object.

The second is fluency: once discovered, the relevant pattern is extracted faster, with less attentional load. This is what the radiologist's eye-tracking showed — faster fixation on lesions, fewer total fixations, attention not captured by irrelevant salience. The expert does not need to consciously search; the pattern is extracted as part of ordinary seeing. Discovery happens once; fluency is what it looks like afterward, from outside.

Robert Goldstone, reviewing the perceptual learning literature in 1998, identified four component operations through which these effects are achieved.¹⁹

Differentiation separates stimuli that were previously indistinguishable — this is the central mechanism for what we've been calling resolution increase. Unitization merges multiple elements into a single detectable unit, which is what happens when a chess grandmaster perceives a pawn structure as one chunk rather than five individual pieces. Attention weighting directs attention toward features that are diagnostically relevant and away from features that are not — this is why the expert radiologist's eye is not drawn to the heart. Stimulus imprinting builds specialized internal detectors for frequently encountered stimuli — a process whose physical correlate may be visible in the sommelier's olfactory bulb volume increase.

These four operations are domain-general: Goldstone documents them operating across cognitive psychology, psychophysics, neuroscience, and development, in domains as different as reading, face recognition, and scientific categorization.¹⁹ The mechanism is not a property of any one sensory modality or any one kind of expertise.

What the mathematics classroom shows is perhaps the most striking demonstration of the mechanism in action. Kellman and colleagues developed Perceptual Learning Modules for algebra — not practice at solving equations, but practice at seeing the structure of algebraic transformations, at recognizing which transformational move was applicable to which type of expression.²⁰ Thirty students who completed two 35-to-40-minute PLM sessions reduced their equation-solving time from approximately 28 seconds to approximately 12 seconds — a 57% reduction — without having solved a single equation during the intervention.²⁰ They practiced seeing. The seeing improved. The solving, which they did not practice, improved as a consequence. Gains were maintained at two-week follow-up.

This is what mechanism unity looks like outside the senses. The same structural operation — training the extraction of relevant features — produces the same signature effects (discovery, fluency) in abstract mathematical material as it does in acoustic phonemes and radiology films.

The measurement question follows. If this is a domain-general mechanism, how do we compare its effects across domains as different as auditory phoneme discrimination and olfactory recognition in wine? The answer comes from a framework developed not by cognitive scientists but by radar engineers, adapted into psychophysics by David Green and John Swets in 1966.²¹ Signal detection theory produces a measure called d-prime — d' — which captures the standardized distance between the internal response distribution when signal is present and when only noise is present. A d' of zero means signal and noise are indistinguishable; a higher d' means they are farther apart, discrimination is easier, and the system is more sensitive. The measure is modality-agnostic: d' can be computed for a radiologist reading a film, a subject detecting a gap in a tone, a wine expert recognizing a scent, or a person discriminating between two categories of emotional experience.²¹

John Swets, in a landmark 1988 paper in Science​, demonstrated that the same relative operating characteristic framework applies to medical imaging, materials testing, weather forecasting, information retrieval, polygraph lie detection, and aptitude testing — placing all these diagnostic systems on a common, easily interpreted scale.²² This is a measurement-level claim, and the distinction matters. D-prime is the inch that applies to heights and pencil lengths and country sizes; its applicability to all of them does not make heights and pencil lengths the same kind of thing. What Swets established is that wherever there is signal and noise, d-prime can be computed — and that means we can compare the discrimination sensitivity of a radiologist and a sommelier and a grief counselor on a single formal axis without asserting that they are running the same cognitive program. The formal unity of the measurement is what earns the article's right to call discrimination capacity a domain-general property rather than a visual metaphor applied to non-visual domains. It is not a metaphor. It is a measure.

The mechanism anchor and the measurement anchor come apart here, and the distinction is important. Swets tells us we can compare sensitivity across domains. Kellman and Garrigan tell us why the same thing keeps happening: experience-induced changes in information extraction, operating with the same discovery and fluency signature, produce resolution increase across chess and mathematics and language and radiology and emotional affect. One is measurement unity; the other is mechanistic unity. Both are real. They are not the same claim.

A reader with cognitive science training will ask: isn't this just pattern recognition? The standard account of expertise holds that experts build large libraries of patterns — chunks, templates, schemas — against which new stimuli are matched faster and more reliably. Chess grandmasters carry 50,000-100,000 position patterns in long-term memory. Radiologists carry signatures of pathology. The pattern-recognition account is well-documented and correct.

Pattern recognition is the computational description of the behavioral output. Resolution is the phenomenological description — phenomenological, in the sense of "what the experience is like from the inside" — of what changes in the perceptual substrate when the distinction becomes registrable at all. They are two levels of description of the same process, not competing accounts. The article takes the phenomenological level seriously because the felt experience names what the behavioral measurements are measuring.

The phonemic case earns this move most clearly. A pattern-recognition theorist would say that the English-learning infant is simply pruning unused templates — the Hindi contrast falls away because it receives no reinforcement. But that account cannot explain why the Hindi distinction becomes perceptually inaccessible rather than merely unused: infants are not leaving the Hindi template in place and ignoring it; they are reorganizing what can be registered at all, such that the distinction is no longer available to the template system. The discrimination is gone from the perceiver, not merely bypassed by the classifier. Resolution-talk captures this reorganization; template-addition-talk does not. Both levels are needed.

And the phenomenological description is not opposed to the behavioral one. The evidence that resolution has increased is discriminative behavior: the eye that fixates the lesion, the ear that hears the semitone, the hand that knows where it is in space. What we have been doing, across nine domains, is finding the same discriminative behavioral signature and asking what it is from the inside.

Why the Sommelier Cannot Read Radiographs#

The natural question, after seeing the same mechanism operate across nine domains, is whether the effect transfers. If training in musical pitch discrimination sharpens the auditory system, does that sharpened system become better at discriminating speech sounds? If years of wine training refine an expert's olfactory discrimination, can that refinement apply to detecting gas leaks or diagnosing medical conditions?

The answer, with few exceptions, is no.

The Waite et al. review is explicit: "the perceptual skills that radiologists develop over the course of their training are restricted to specific radiologic image perception tasks. Indeed, radiologists are no better at performing nonradiologic search tasks than nonradiologists are."⁶ The expert radiologist, who has learned to find lung nodules in chest films with a precision that took years to develop, looks at a natural scene search task and performs identically to someone who has never read a radiograph. The perceptual machinery trained on radiographs does not generalize to other visual search tasks.

The proprioceptive training study showed the same specificity in a more contained form: proprioceptive acuity improved 11% in the spatial location where practice occurred, and not at all 25 centimeters away.⁵ The body learned to know where it was in that workspace, not in general. Moving the task location erases the improvement. Moving to a different sensory system erases it even more completely.

Musical temporal resolution — the ability to detect 1.81-millisecond gaps — does not confer a general-purpose temporal processing advantage. Trained musicians are not better at detecting brief gaps in non-musical contexts than untrained listeners.⁴

The limit of the resolution thesis is here. Learning builds domain-specific resolution, not general-purpose cognitive sharpness. The sommelier cannot read radiographs. The chess master is no better at predicting stock prices than a beginning chess player. The mechanism is unified — the same experience-induced change in information extraction operates across all these domains — but each instance of the mechanism is local. Resolution is calibrated to the specific domain and task where training occurred. This locality is the mechanism's actual shape, not a limitation of it.

That word calibrated matters. Calibration is not a lesser thing than generalization; in some respects it is a finer achievement. An instrument calibrated to its task has a sensitivity that fits what it is measuring — not too coarse to register the relevant distinctions, not so fine that it amplifies irrelevant noise. The sommelier's olfactory resolution is tuned to wine because wine training tuned it. The radiologist's visual resolution fits lung films because lung films trained it. Each is an instrument shaped by the domain it works in. The calibration is not accidental. It is what the training produced. The limitation and the accomplishment are the same thing.

Which is what reveals calibration's cost. If resolution is calibrated to a specific domain, it can be miscalibrated — tuned to a task in a way that produces characteristic failures when the task changes, or when higher resolution than the task requires is brought to bear on it.

Appropriate Resolution, Not Maximum#

In 2008, Merim Bilalić and colleagues at Oxford conducted a chess study that begins, deceptively, as a straightforward expertise-effect experiment.²³ They presented chess experts — players ranging from Candidate Master to International Master in rating — with positions that contained two possible solutions: a familiar five-move checkmate pattern that the expert would recognize quickly, and a less familiar three-move checkmate that was actually simpler and more efficient. The question was whether experts could find the better solution once they had found the good one.

They could not. When a familiar solution was present in the position, expert performance dropped dramatically — by approximately three standard deviations of skill.²³ The gap between an International Master and a club-level player is the gap between someone who has spent years inside competitive tournament chess and someone who plays well but casually — and the presence of a familiar-but-suboptimal pattern closed it. Not because the three-move solution was obscure. It was simpler. It was right there. But the five-move pattern had fired first, and the activated schema captured attention and directed it away from the board regions that contained the simpler answer.

The eye-tracking companion study makes the mechanism visible.²³ Experts who had found the familiar five-move solution — and who, when asked, reported that they were still searching for a better one — continued fixating squares associated with the five-move checkmate. Their eyes were not searching the rest of the board. They were orbiting the first answer, guided by the schema that had activated, while verbally asserting that they were open to something better. From inside, it felt like searching. The subjective experience of scanning a board, attending and considering, was present and genuine. What was absent was the capacity to register the squares that didn't belong to the active pattern — they were there, available to vision, but the schema had already decided they were not relevant. The expert was not failing to look. The expert was failing to see what they were looking at.

The chess expertise that produced 93% recall on meaningful game positions is the same chess expertise that produced this systematic blindness. The gain and the cost share the same mechanism: the high-resolution expert schema activates fast and accurately on its trained territory, and in doing so directs attention toward what fits the schema and away from what does not. Resolution increase and Einstellung — the psychological term for this kind of set effect — are the same capacity, viewed from two different positions on the board.

The clinical cousin of chess Einstellung is the incidentaloma. As imaging technology has improved in resolution — higher-field MRIs, multislice CT scanners, finer-grain mammography — the machines now detect structural variations that were always present in tissue but below the detection threshold of earlier technology. John O'Sullivan and colleagues conducted a BMJ umbrella review of incidentaloma prevalence across imaging modalities: 45% of chest CT scans produce clinically irrelevant findings; 38% of CT colonoscopies; 34% of cardiac MRIs; 22% of brain and spine MRIs.²⁴ The findings are real — the structural variations are there — but they are clinically irrelevant. They trigger patient anxiety and cascade into unnecessary interventions. The imaging instrument now discriminates distinctions that should not be discriminated for functional purposes.

The chess case and the incidentaloma case are not quite the same kind of thing — one is a cognitive failure in an expert's reasoning, the other is a machine artifact — but they share a structural principle. Resolution uncalibrated to the task produces characteristic failure. The chess expert's resolution is calibrated to find chess patterns; when the task requires not following the first recognized pattern, the expert resolution becomes a liability. The imaging machine's resolution is calibrated to find structural variations; when the clinical task requires distinguishing significant from irrelevant structural variations, the resolution produces noise that looks like signal.

Erik Dane, in an Academy of Management Review paper, proposed the concept of cognitive entrenchment to name the cognitive-expert version of this pattern: a high level of stability in domain schemas that makes those schemas efficient for standard problems but resistant to adaptation when novel problems require different framings.²⁵ As expertise deepens, the schemas become more stable — faster, more reliable, harder to override. The expert who cannot see the simple solution is not failing to think. They are thinking with extreme efficiency in the schema that has become their primary tool, and that efficiency is precisely what prevents the lateral move.

The radiology error literature gives this pattern numerical weight. Expert radiologists in daily practice commit real-time errors in approximately 3 to 5% of studies, and retrospective review of cases — going back through images in light of later diagnoses — reveals that approximately 30% of lesions visible in hindsight were missed in real-time interpretation.²⁶ Sixty to eighty percent of all radiology errors are perceptual rather than cognitive: the error is not wrong inference from a seen finding but missed detection of a finding that was there.²⁶ In patients later diagnosed with lung cancer, retrospective review of prior "normal" chest X-rays found the cancer visible in up to 90% of cases. High resolution does not equal error-free. Calibration — not resolution alone — is the test.

The felt experience of expertise development — the phenomenology the article opened with, the moment the blur breaks into features — can occur without actual improvement in discrimination when the environment is irregular and feedback is corrupted. Kahneman and Gary Klein, in their 2009 paper on conditions for intuitive expertise, established that genuine expertise requires two things: an environment regular enough to be predictable, and clear, honest feedback over time.²⁷ In chess, radiology, language, wine, music — the domains where this article has collected evidence — the environments are regular enough and the feedback clear enough that resolution reliably increases with training. In long-range political prediction, financial market forecasting, some clinical diagnoses where outcomes are delayed or confounded — the phenomenology of resolution increase can develop without the underlying discrimination capacity. The felt experience is not a reliable marker. What counts is calibration to the structure of the domain, not confidence in the feel of the knowing.

None of this overturns the primary case. The fact that resolution can be miscalibrated, that maximum resolution is not the goal, that the expert map can become a constraint — these are consequences that follow from the thesis, not objections to it. A thesis that predicts its own pathology is more robust, not less. The honest version of the claim is: learning is resolution increase, calibrated to the task where training occurred, operating in environments regular enough and with feedback clear enough to actually improve discrimination. This is the honest shape of the thesis, and it is stronger for what it concedes.

Scope: What the Thesis Predicts but Has Not Proven#

The preceding nine domains — phonemic, musical, temporal, medical, olfactory, proprioceptive, emotional, causal, social — constitute what the evidence actually supports. It would be an overstatement to say that resolution increase is the mechanism of learning in every domain, and the article has not said so.

The thesis's predictions are interesting and open. Aristotelian phronesis — practical wisdom, the ability to perceive what a situation morally calls for — if the thesis extends there, would predict that moral discernment follows the same resolution structure: the capacity to register morally salient features that an inexperienced moral perceiver cannot distinguish. If aesthetic expertise shows the same structure, it would predict structural discrimination within a tradition — the jazz musician's ear for a harmonic substitution that novices hear only as a chord change, the art critic's perception of compositional tension that a museum visitor experiences as general quality or vague unease. The thesis predicts — and does not demonstrate — that political perception is the capacity to see structural power where an inexperienced observer sees only individual action, and that meta-cognitive skill is the capacity to perceive distinctions in the map of one's own uncertainty, the difference between confident ignorance and calibrated uncertainty, made visible by practice.

These are open questions: the places where the thesis can be tested, where it might be falsified, where it remains genuinely open. None of them has the empirical standing of the phonemic narrowing study or the radiology training study. They are predictions, not evidence.

If expertise is primarily a perceptual phenomenon — if what training builds is not a larger store of facts but a higher-resolution perceptual system — then pedagogies organized around structured discrimination practice deserve more attention than they typically receive. That is a lead for someone else to follow.

What Else Comes Into Focus#

There is a larger version of the same mechanism.

Consider what germ theory did to medicine in the 1870s and 1880s. The distinction between infectious agents and non-infectious matter was empirically present — always in the data, in the patterns of disease spread, in the differential outcomes of different sanitary conditions. Medicine before Koch and Pasteur could not register it. The conceptual and instrumental apparatus wasn't there. The clinical presentation was visible; the causal structure was invisible. What germ theory provided was not primarily new facts, though it produced many. It provided a new resolution in the causal dimension: a distinction the prior framework could not make perceivable suddenly became unmissable.

The concept of a feedback loop did something similar for the understanding of biological, economic, and social systems. Negative feedback — the property by which a system's deviation from a set point produces a correcting force — was always present in population dynamics, in market prices, in hormonal regulation, in thermostat design. Rottman's finding about expert scientists is not just a fact about individual scientists' categorization behavior: it is a fact about what a scientific tradition does to the perception of its practitioners.¹⁵ A discipline, as it matures, teaches its members to perceive causal structures that novices cannot see. The individual resolution upgrade, multiplied across a community of researchers over decades, is what a scientific paradigm shift is from the inside.

The infant's phonemic narrowing is not a developmental curiosity localized to the first year of life. It is the earliest, cleanest instance of a structure that recurs at every scale at which learning occurs. An infant's auditory system reorganizing its category boundaries to match the ambient language. A discipline reorganizing its conceptual boundaries to match the causal structures that have proven empirically productive. A culture gaining the capacity to perceive distinctions that previous generations could not register — and, in the process, losing the capacity to perceive some of what came before.

Conceptual upgrades like these — opportunity cost, entropy, natural selection, relative risk — are resolution increases at civilizational scale. Not a new thing discovered, but a new distinction that, once perceivable, makes previously invisible structure unmissable. The collective apparatus of human inquiry is a machine for gaining this kind of new sight. The infant and the discipline are running the same process.

What is currently below the threshold of collective discrimination — in the natural sciences, in the social sciences, in the ongoing attempt to understand how to live together — that requires only a little more resolution to become undeniable? What causal structures are present in the data, what moral distinctions are present in experience, that a future generation will find as obvious as we find the germ theory of disease, and as invisible as that theory was before someone developed the apparatus to see it? We do not know, because we are below the threshold. We are, at this moment, in the position of the English-learning seven-month-old hearing Hindi — with the full complement of distinctions still available, not yet narrowed into what we can name.

References#

1. Werker, J.F. & Tees, R.C. (1984). Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant Behavior and Development, 7​(1), 49-63. https://www.sciencedirect.com/science/article/abs/pii/S0163638384800223 ↩︎ ↩︎ ↩︎

2. Flege, J.E. (1995). Second language speech learning: Theory, findings, and problems. In W. Strange (Ed.), Speech Perception and Linguistic Experience: Issues in Cross-Language Research (pp. 233-277). York Press. Supplemented by: high-variability phonetic training studies reviewed in Springer (2021). https://link.springer.com/article/10.1007/s10936-021-09774-3 ↩︎

3. Zarate, J.M., Ritson, C.R. & Poeppel, D. (2012). Pitch-interval discrimination and musical expertise: Is the semitone a perceptual boundary? The Journal of the Acoustical Society of America, 132​(2), 984-993. DOI: 10.1121/1.4733535. PMC3427364. ↩︎

4. Kumar, P., Sanju, H.K. & Nikhil, J. (2016). Temporal resolution and active auditory discrimination skill in vocal musicians. International Archives of Otorhinolaryngology, 20​(4), 310-314. DOI: 10.1055/s-0035-1570312. PMC5063729. ↩︎ ↩︎

5. Wong, J.D., Wilson, E.T. & Gribble, P.L. (2011). Spatially selective enhancement of proprioceptive acuity following motor learning. Journal of Neurophysiology, 105​(5), 2512-2521. PMC3094168. ↩︎ ↩︎

6. Waite, S., Grigorian, A., Alexander, R.G., Macknik, S.L., Carrasco, M., Heeger, D.J., et al. (2019). Analysis of perceptual expertise in radiology — Current knowledge and a new perspective. Frontiers in Human Neuroscience​. PMC6603246. ↩︎ ↩︎ ↩︎

7. Chen, W., HolcDorf, D., McCusker, M.W., Gaillard, F. & Howe, P.D.L. (2017). Perceptual training to improve hip fracture identification in conventional radiographs. PLOS ONE​. PMC5739398. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0189192 ↩︎ ↩︎

8. Parr, W.V., et al. (2002). Demystifying wine expertise: Olfactory threshold, perceptual skill and semantic memory in expert and novice wine judges. Chemical Senses, 27​(8), 747-755. https://academic.oup.com/chemse/article/27/8/747/387724 ↩︎

9. Tempere, S., Cuzange, E., Malak, J., Bougeant, J.C., de Revel, G. & Sicard, G. (2011). The training level of experts influences their detection thresholds for key wine compounds. Chemosensory Perception, 4​, 99-115. DOI: 10.1007/s12078-011-9090-8. ↩︎ ↩︎

10. Seubert, J., et al. (2022). Olfactory bulb volume and cortical thickness evolve during sommelier training. Human Brain Mapping, 43​(8), 2621-2633. PubMed 35218277. ↩︎

11. Barrett, L.F., Gross, J., Christensen, T.C. & Benvenuto, M. (2001). Knowing what you're feeling and knowing what to do about it: Mapping the relation between emotion differentiation and emotion regulation. Cognition & Emotion, 15​(6), 713-724. https://www.tandfonline.com/doi/abs/10.1080/02699930143000239 ↩︎ ↩︎

12. Kashdan, T.B., Barrett, L.F. & McKnight, P.E. (2015). Unpacking Emotion Differentiation: Transforming Unpleasant Experience by Perceiving Distinctions in Negativity. Current Directions in Psychological Science, 24​(1), 10-16. Primary aggression finding: Pond, R.S., et al. (2012). Emotion differentiation moderates aggressive tendencies in angry people. Emotion, 12​, 326-337. Primary alcohol finding: Kashdan, T.B., et al. (2010). Emotion differentiation as resilience against excessive alcohol use. Psychological Science​. ↩︎ ↩︎

13. Vedernikova, E., Kuppens, P. & Erbas, Y. (2021). From Knowledge to Differentiation: Increasing Emotion Knowledge Through an Intervention Increases Negative Emotion Differentiation. Frontiers in Psychology, 12​, 703757. PMC8662934. ↩︎ ↩︎

14. Tugade, M.M., Fredrickson, B.L. & Barrett, L.F. (2004). Psychological Resilience and Positive Emotional Granularity: Examining the Benefits of Positive Emotions on Coping and Health. Journal of Personality, 72​(6), 1161-1190. PMC1201429. ↩︎

15. Rottman, B.M., Gentner, D. & Goldwater, M.B. (2012). Causal Systems Categories: Differences in Novice and Expert Categorization of Causal Phenomena. Cognitive Science, 36​(5), 919-932. https://onlinelibrary.wiley.com/doi/full/10.1111/j.1551-6709.2012.01253.x ↩︎ ↩︎

16. de Groot, A.D. (1965). Thought and Choice in Chess​. Mouton. Chase, W.G. & Simon, H.A. (1973). Perception in chess. Cognitive Psychology, 4​(1), 55-81. ↩︎ ↩︎

17. Lebrecht, S., Pierce, L.J., Tarr, M.J. & Tanaka, J.W. (2009). Perceptual other-race training reduces implicit racial bias. PLoS ONE, 4​(1), e4215. PMC2627769. DOI: 10.1371/journal.pone.0004215. ↩︎ ↩︎ ↩︎

18. Kellman, P.J. & Garrigan, P. (2009). Perceptual learning and human expertise. Physics of Life Reviews, 6​(2), 53-84. DOI: 10.1016/j.plrev.2008.12.001. PubMed: 20416846. ↩︎ ↩︎

19. Goldstone, R.L. (1998). Perceptual learning. Annual Review of Psychology, 49​, 585-612. PubMed: 9496632. ↩︎ ↩︎

20. Kellman, P.J., Massey, C., et al. (2010). Perceptual Learning Modules in Mathematics: Enhancing Students' Pattern Recognition, Structure Extraction, and Fluency. Topics in Cognitive Science, 2​(2), 285-305. PMC6124488. ↩︎ ↩︎

21. Green, D.M. & Swets, J.A. (1966). Signal Detection Theory and Psychophysics​. Wiley. ↩︎ ↩︎

22. Swets, J.A. (1988). Measuring the accuracy of diagnostic systems. Science, 240​(4857), 1285-1293. PubMed: 3287615. ↩︎

23. Bilalić, M., McLeod, P. & Gobet, F. (2008). Inflexibility of experts — Reality or myth? Quantifying the Einstellung effect in chess masters. Cognitive Psychology, 56​(2), 73-102. PubMed: 17418112. Companion eye-tracking study: Bilalić, M., McLeod, P. & Gobet, F. (2008). Why good thoughts block better ones: The mechanism of the pernicious Einstellung (set) effect. Cognition, 108​(3), 652-661. PubMed: 18565505. ↩︎ ↩︎ ↩︎

24. O'Sullivan, J.W., Muntinga, T., Grigg, S. & Ioannidis, J.P.A. (2018). Prevalence and outcomes of incidental imaging findings: umbrella review. BMJ​. PMC6283350. ↩︎

25. Dane, E. (2010). Reconsidering the Trade-off Between Expertise and Flexibility: A Cognitive Entrenchment Perspective. Academy of Management Review, 35​(4), 579-603. https://journals.aom.org/doi/10.5465/amr.35.4.zok579 ↩︎

26. Bruno, M.A., Walker, E.A. & Abujudeh, H.H. (2015). Understanding and confronting our mistakes: The epidemiology of error in radiology and strategies for error reduction. Radiographics​. https://pmc.ncbi.nlm.nih.gov/articles/PMC3609674/ ↩︎ ↩︎

27. Kahneman, D. & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64​(6), 515-526. PubMed: 19739881. ↩︎

Further Reading#

• Kellman, P.J. "Perceptual Learning." In Stevens' Handbook of Experimental Psychology (3rd ed.), Vol. 1. Wiley, 2002. — Background on the breadth of perceptual learning research; informed the mechanism section and helped frame discovery vs. fluency as distinct functional categories.
• Werker, J.F. & Tees, R.C. (2005). Speech perception as a window for understanding plasticity and commitment in language acquisition. Developmental Psychobiology, 46​, 233-251. — Review of phonemic narrowing across subsequent decades; relevant context for the E01-E02 showcase; not cited directly to keep the §2 focused on the 1984 primary.
• Maurer, D. & Werker, J.F. (2013). Perceptual narrowing during infancy: A comparison of language and faces. Developmental Psychobiology, 56​, 154-178. — Extends phonemic narrowing framework to face perception; relevant background for the E27 (Lebrecht) social exhibit; reserve evidence.
• Barrett, L.F. How Emotions Are Made: The Secret Life of the Brain​. Houghton Mifflin Harcourt, 2017. — Background on emotional granularity and the theory of constructed emotion; informed the §5 emotional domain exhibits; provided conceptual context for why emotional discrimination is framed as perceptual, not cognitive.
• Ericsson, K.A. & Pool, R. Peak: Secrets from the New Science of Expertise​. Houghton Mifflin Harcourt, 2016. — The dominant popular treatment of expertise acquisition; explicitly not cited because its behavioral/deliberate-practice frame is what the article is complementing (not competing with). Background shaping of what to address and what to leave to other treatments.
• Horowitz, A. On Looking: Eleven Walks with Expert Eyes​. Scribner, 2013. — The closest existing popular treatment; showed the phenomenon without theorizing the mechanism; background framing of the article's gap.
• Kahneman, D. Thinking, Fast and Slow​. Farrar, Straus and Giroux, 2011, Ch. 21-22 (Intuitions vs. Formulas; Expert Intuition: When Can We Trust It?). — Background for the §8 Kahneman/Klein validity-conditions absorption; audience's dominant cognitive frame.