一篇启示 - 《神经科学，仍在等待它的牛顿》

在微博上看到果壳网的一条微博《神经科学，仍在等待它的牛顿》：http://www.guokr.com/article/410905/

对这篇文章我觉得写得特别好，所以特此转载如下：

Rob Dobi, via crucialparts.blogspot.com

（文 ／GARY MARCUS）20世纪90年代初，还是MIT研究生的大卫•坡佩尔（David Poeppel）发现了一件惊人的事情。他当时正在研究言语知觉的神经生理学基础，而一种新的技术才刚刚开始流行，那就是正电子发射断层扫描（PET）。 那时候大约有6篇PET研究言语感知方面的论文得到了发表，都是在顶级的科研期刊上，大卫试图整合这些论文，本质上就是把每篇论文中提到的大脑在识别言语 的过程中产生兴奋的那一部分拿来作比较。令人震惊的是，他没有找到任何彼此相一致的地方。这些研究每一个发表出来都广受瞩目，但合起来看它们之间完全没有 一致性，加在一起什么也不是。整件事情就好像有6个证人以6种完全不同的方式描述罪案。

学科发展，意外的流转

这对神经科学来说可是糟糕透顶——如果6个研究得出了6种不同的结果，那谁还要相信神经科学家说的话呢？随之而来的是许多纠结的讨论——是不是PET不可靠，因为它涉及往大脑里注射放射性的示踪剂？还是说这些研究本身不够严谨？似乎没有人知道。

然 后，让人想不到的是，这个学科发达了起来。大脑成像图不但没有遭到弃用，反而变得更加流行。PET技术被更加灵活的功能核磁共振成像（fMRI）所取代， 由此科学家不必使用危险的放射性示踪剂也能研究人脑，还能进行时间更长的实验以采集更多的数据，得出更可靠的结果。渐渐地，实验方法也变得更加谨慎。随着 fMRI仪器变得更加普及，实验方法变得更加标准和完善，研究者们终于开始在实验室里取得了一定程度上的共识。

与 此同时，神经科学开始走向公众，而且声势浩大。工作中的大脑的那些花哨的彩色图片，成了媒体在提及人类思维时必定会举出的例子，从而诱使人们产生了一种错 误的理解。（在《Duped》杂志的一篇特稿里，作者玛格丽特•塔尔博特［Margaret Talbot］描写了耶鲁大学的实验发现，在论文里加入神经学的内容会使人更加相信它们。）1990年时普通人根本连听都没听过的大脑成像技术，成了理解 人类精神生活最负盛名的方法。“神经”这一前缀到哪里都能见到：神经法律、神经经济学、神经政治学。神经伦理家们还想是不是能基于一个人新皮层的大小来改 变他的定罪。

然 后，轰！局面又来了个转变。在几乎可谓绝对霸主的位子上坐了20年以后，少数几个聪明人开始说话了，问：所有这些脑部研究真的想我们想的那样告诉了我们很 多东西吗？去年出版的一本名不见经传的绝妙好书《神经狂热》（Neuromania），便对我们对大脑成像越来越深的迷恋表示了忧虑。雷蒙德•塔利 （Raymond Tallis）的一本书已于今年出版，书中也援引了这一词汇，并论证了类似的观点。在《走出大脑》（Out of our Heads）一书中，哲学家•阿尔瓦•诺埃（Alva Noë，加州大学伯克利分校的哲学教授）写道， “很容易忽视一个事实……用 PET 和 fMRI 得到的并非是大脑在实际行动中的图片。” 实际上，大脑的图像是依靠复杂的数学假设精心重构的结果，而根据今年早些时候的一项研究，在不同类型的计算机上进行分析，这种重构有时会产生稍微不同的结 果。

就 在前不久，类似这样的忧虑以及一些神经学博客上发人深省的博文，终于闯入了主流视野，在《纽约时报》的一篇评论文章中亮了相。在文中，纽约时报记者阿利萨 •柯尔特（Alissa Quart）表明了自己的态度，“我举双手赞成抵制这种有时也被称为大脑色情的东西，它对这种还原论的、不严谨的思维方式和我们愿意接受一切看似神经科学 的解释这一事实提出了重要的问题。”

更恰当的应对之道

柯 尔特和那帮不断壮大的神经学批评家队伍只说对了一半：如今这个21世纪初的世界里的确充满了大脑色情，充满了草率的还原论思考和对神经科学解释不体面的欲 望。但正确的解决方法不是把神经科学一刀切掉， 理解神经科学能够告诉我们什么和不能告诉我们什么，才是更恰当的做法。

为 什么我们不应该简简单单地把神经科学全部否决，首先也是最重要的原因是显而易见的：如果我们想要了解人的思维、了解所有的人性生发的地方，我们必须去理解 大脑的生物学。第二个原因，是神经科学已经告诉了我们很多，只是不是我们所想的那样。能上报纸的研究往往是相关性不好、但是讨论了有意思的人类行为的那 种，比如“性高潮时女性大脑的3D图”，以及“玩扑克时你的大脑是这样的”。

但 很多这样的报道都建立在了一个错误的前提之上：在大脑中最亮的神经组织就是参与认知功能的唯一组织。实际上，大脑很少以这样的方式运作。大脑做的大部分事 情都要牵涉到许多不同的组织一起工作。说情感在杏仁核里面，或是决策是前额叶皮质的事，顶多只能算是简写，而且是一个误导性的简写。举例来说，不同的情感 依赖于神经基质的不同组合。理解句子的行为可能涉及布罗卡区（在左脑与语言相关的地方），但同时也会动用颞叶中分析声音信号的脑区和一部分感觉运动皮层， 基底神经节也会变得活跃。（如果是先天失明的人，一些视觉皮层也会发挥一定的作用。）总之，不是只有一处，而是有很多处，其中一些可能没那么活跃，但仍然 有着至关重要的作用；真正重要的，是庞大的神经网络是以何种方式一起工作的。

fMRI 能从大脑图像中选出的最小元素是种叫体素（voxel）的东西。但是，体素比神经元大很多，而且从长远看，了解大脑最好的方式可能不是问哪些特定的体素在 一个给定的过程中最活跃，而是看这些体素里面的许多的神经元是如何共同工作的。也因为这一点，到头来我们可能会发现fMRI并不是研究大脑的最佳工具，尽 管在眼下它用着很方便。fMRI最终可能成为把人类引向显微镜的放大镜，而显微镜才是我们真正需要的。如果人脑中的行为大部分都发生于神经元而不是体素或脑区（通常包含数百或数千的体素）的层面，我们可能需要新的研究方法，比如光遗传学或用于研究单个神经元的自动化机器人引导工具；我自己的猜测是，我们需要对其他动物的大脑有了更多的见解之后，才能完全掌握人类大脑的运作。关于独立神经元组成的阵列是如何与复杂行为相关联的，科学家们还仍然处在努力构建理论的阶段，连大体框架都还没有搭好。神经科学尚未遇见它的牛顿，更不用说爱因斯坦了。

但这并不是放弃的借口。当达尔文写下《物种起源》的时候，没有人知道DNA是做什么的，也没有人想到我们有朝一日还能给它测序。

神经科学如今面临的真正的问题并不在科学——虽然许多方法论上的挑战仍然存在——而在于我们对它的预期。人 脑是一个极度复杂的组合体，在任何时候都有数十亿神经元参与（以及脱离）协作。有一天我们的多数行为都将有神经科学的解释，但这些解释将非常复杂。现在， 关于这些部分是如何关联起来的，我们的理解能力相当有限，有点儿像试图从克利夫兰上空的飞机窗口去看清俄亥俄州的政治动态一样。

这 或许就是为什么今天最好的神经科学家可能是那些最少登上头条新闻的人，比如研究人在理解单个字时脑中复杂动力如何作用的那些科学家。正如大卫•坡佩尔所说 的，我们现在需要的是 “一丝不苟地把一些基本的大脑功能给一层一层分析透彻，而不要听着宏大而实际意义模糊的概念，比如脑基美学（brain-based aesthetics），现在我们连大脑是如何识别一条直线这么简单的事物都还不了解”。

那些对复杂大脑功能的简短的解释往往能成为很好的标题党，但极少是真实的。不过，这并不意味着以后也没有解释，只是意味着我们的大脑没有演化成很容易理解的样子。

原版在这里《 What neuroscience really teaches us and what it doesn't》：http://www.newyorker.com /online/blogs/newsdesk/2012/12/what-neuroscience-really-teaches-us-and-what-it-doesnt.html?mbid=social_retweet

In the early nineteen-nineties, David Poeppel, then a graduate student at M.I.T. (and a classmate of mine)—discovered an astonishing thing. He was studying the neurophysiological basis of speech perception, and a new technique had just come into vogue, called positron emission tomography (PET). About half a dozen PET studies of speech perception had been published, all in top journals, and David tried to synthesize them, essentially by comparing which parts of the brain were said to be active during the processing of speech in each of the studies. What he found, shockingly, was that there was virtually no agreement. Every new study had published with great fanfare, but collectively they were so inconsistent they seemed to add up to nothing. It was like six different witnesses describing a crime in six different ways.

This was terrible news for neuroscience—if six studies led to six different answers, why should anybody believe anything that neuroscientists had to say? Much hand-wringing followed. Was it because PET, which involves injecting a radioactive tracer into the brain, was unreliable? Were the studies themselves somehow sloppy? Nobody seemed to know.

And then, surprisingly, the field prospered. Brain imaging became more, not less, popular. The technique of PET was replaced with the more flexible technique of functional magnetic resonance imaging (fMRI), which allowed scientists to study people’s brains without the use of the risky radioactive tracers, and to conduct longer studies that collected more data and yielded more reliable results. Experimental methods gradually become more careful. As fMRI machines become more widely available, and methods became more standardized and refined, researchers finally started to find a degree of consensus between labs.

Meanwhile, neuroscience started to go public, in a big way. Fancy color pictures of brains in action became a fixture in media accounts of the human mind and lulled people into a false sense of comprehension. (In a feature for the magazine titled “Duped,” Margaret Talbot described research at Yale that showed that inserting neurotalk into a papers made them more convincing.) Brain imaging, which was scarcely on the public’s radar in 1990, became the most prestigious way of understanding human mental life. The prefix “neuro” showed up everywhere: neurolaw, neuroeconomics, neuropolitics. Neuroethicists wondered about whether you could alter someone’s prison sentence based on the size of their neocortex.

And then, boom! After two decades of almost complete dominance, a few bright souls started speaking up, asking: Are all these brain studies really telling us much as we think they are? A terrific but unheralded book published last year, “Neuromania,” worried about our growing obsession with brain imaging. A second book, by Raymond Tallis, published this year, invoked the same term and made similar arguments. In the book “Out of our Heads,” the philosopher Alva Noë wrote, ”It is easy to overlook the fact that images… made by fMRI and PET are not actually pictures of the brain in action.” Instead, brain images are elaborate reconstructions that depend on complex mathematical assumptions that can, as one study earlier this year showed, sometimes yield slightly different results when analyzed on different types of computers.

Last week, worries like these, and those of thoughtful blogs like Neuroskeptic and The Neurocritic, finally hit the mainstream, in the form of a blunt New York Times op-ed, in which the journalist Alissa Quart declared, “I applaud the backlash against what is sometimes called brain porn, which raises important questions about this reductionist, sloppy thinking and our willingness to accept seemingly neuroscientific explanations for, well, nearly everything.”

Quart and the growing chorus of neuro-critics are half right: our early-twenty-first-century world truly is filled with brain porn, with sloppy reductionist thinking and an unseemly lust for neuroscientific explanations. But the right solution is not to abandon neuroscience altogether, it’s to better understand what neuroscience can and cannot tell us, and why.

The first and foremost reason why we shouldn’t simply disown neuroscience altogether is an obvious one: if we want to understand our minds, from which all of human nature springs, we must come to grips with the brain’s biology. The second is that neuroscience has already told us lot, just not the sort of things we may think it has. What gets play in the daily newspaper is usually a study that shows some modest correlation between a sexy aspect of human behavior, with headlines like “FEMALE BRAIN MAPPED IN 3D DURING ORGASM” and “THIS IS YOUR BRAIN ON POKER”

But a lot of those reports are based on a false premise: that neural tissue that lights up most in the brain is the only tissue involved in some cognitive function. The brain, though, rarely works that way. Most of the interesting things that the brain does involve many different pieces of tissue working together. Saying that emotion is in the amygdala, or that decision-making is the prefrontal cortex, is at best a shorthand, and a misleading one at that. Different emotions, for example, rely on different combinations of neural substrates. The act of comprehending a sentence likely involves Broca’s area (the language-related spot on the left side of the brain that they may have told you about in college), but it also draws on the parts of the brain in the temporal lobe that analyze acoustic signals, and part of sensorimotor cortex and the basal ganglia become active as well. (In congenitally blind people, some of the visual cortex also plays a role.) It’s not one spot, it’s many, some of which may be less active but still vital, and what really matters is how vast networks of neural tissue work together.

The smallest element of a brain image that an fMRI can pick out is something called a voxel. But voxels are much larger than neurons, and, in the long run, the best way to understand the brain is probably not by asking which particular voxels are most active in a given process. It will instead come from asking how the many neurons work together within those voxels. And for that, fMRI may turn not out not to be the best technique, despite its current convenience. It may ultimately serve instead as the magnifying glass that leads us to the microscope we really need. If most of the action in the brain lies at the level of neurons rather than voxels or brain regions (which themselves often contain hundreds or thousands of voxels), we may need new methods, like optogenetics or automated, robotically guided tools for studying individual neurons; my own best guess is that we will need many more insights from animal brains before we can fully grasp what happens in human brains. Scientists are also still struggling to construct theories about how arrays of individual neurons relate complex behaviors, even in principle. Neuroscience has yet find its Newton, let alone its Einstein.

But that’s no excuse for giving up. When Darwin wrote “The Origin of Species,” nobody knew what DNA was for, and nobody imagined that we would eventually be sequencing it.

The real problem with neuroscience today isn’t with the science—though plenty of methodological challenges still remain—it’s with the expectations. The brain is an incredibly complex ensemble, with billions of neurons coming into—and out of—play at any given moment. There will eventually be neuroscientific explanations for much of what we do; but those explanations will turn out to be incredibly complicated. For now, our ability to understand how all those parts relate is quite limited, sort of like trying to understand the political dynamics of Ohio from an airplane window above Cleveland.

Which may be why the best neuroscientists today may be among those who get the fewest headlines, like researchers studying the complex dynamics that enter into understanding a single word. As Poeppel says, what we need now is “the meticulous dissection of some elementary brain functions, not ambitious but vague notions like brain-based aesthetics, when we still don’t understand how the brain recognizes something as basic as a straight line.”

The sort of short, simple explanations of complex brain functions that often make for good headlines rarely turn out to be true. But that doesn’t mean that there aren’t explanations to be had, it just means that evolution didn’t evolve our brains to be easily understood.

Read more: http://www.newyorker.com/online/blogs/newsdesk/2012/12/what-neuroscience-really-teaches-us-and-what-it-doesnt.html#ixzz2FNRUiFvn

文章中提出：”了解大脑最好的方式可能不是问哪些特定的体素在一个给定的过程中最活跃，而是看这些体素里面的许多的神经元是如何共同工作的。“这 一观点对AI及其有利，因为以冯诺依曼机的可并行性，是无法从头到尾按照人类神经元方式组装直到完整系统的，组装起来也根本不具有可运行性。对比神经元计 算系统的体系结构来说，冯诺依曼体系结构基本上是不可并行的。而我们要做的使用电器手段实现等价的计算系统，但是如上所述又不能照抄照搬，因为电子元件与 生化元件基本上不具备任何有用的共同点。所以要实现这一目的，我们只能在神经元层面网上一层至几层来模拟，这句话的意思是打个比方，现在的生物系统是由细 胞元件组成的，不错，他们组成了组织，组织又构成器官，这就好像汇编到高级语言，到软件一样，我们不能模仿他的汇编级别，因为不具备这样的条件，我们要模 仿他的上层，也就是高级语言级别，这个级别的模仿对我们有利。而按照文中的意思，我们所模仿的最低的级别，很肯能就是大脑的秘密所在的级别。也就是说有可 能性做到这个事情：大脑的秘密所在的级别恰好就在我们所模仿的级别。我们要研究的是大脑运作的模式，而不是彻底的从神经元工作方式上的模仿。我现在的看法 是应该比神经网络再往上一些。

文章中提出：“关于独立神经元组成的阵列是如何与复杂行为相关联的，科学家们还仍然处在努力构建理论的阶段，连大体框架都还没有搭好。神经科学尚未遇见它的牛顿，更不用说爱因斯坦了。...” 确实，之前还一直想问刘老师，自己也在人工智能有关的论坛上搜资料，大家普遍的看法都非常消极，说这个任务基本上要很久很久才能实现，当时他们也都信心满 满，等做上之后再看才发现目标实在是太远。我一直不明白他们为什么突然看到了什么样的阻力，不过看完这篇文章我觉得，很有可能就是因为神经科学不给力的缘 故。AI是个交叉学科，瓶颈是短板确定的，而且神经科学原理这么重要，肯定是很重要的短板，所以很有可能神经科学就是这个问题的关键。他们在初入这行的时 候可能也是忽略了这一学科的现状，其实当时我也没考虑这方面，谁会想到生物学范畴里面的脑生物学基础居然如此的不发达，发展了几百年基本上等于没发展。文 章还说：”现在我们连大脑是如何识别一条直线这么简单的事物都还不了解“，这也说明了这个问题。目前神经科学的供给及其缺少，再加上神经科学本身也没有时间的奠基，总是就是非常不给力。

文章中提出：“人 脑是一个极度复杂的组合体，在任何时候都有数十亿神经元参与（以及脱离） 协作。有一天我们的多数行为都将有神经科学的解释，但这些解释将非常复杂。现在，关于这些部分是如何关联起来的，我们的理解能力相当有限，有点儿像试图从 克利夫兰上空的飞机窗口去看清俄亥俄州的政治动态一样。”这一点也是证明，大脑的体系结构就像计算机一样，从基础电路到成型的操作系统，到云的层次，是经过了无数层的封装和加工磊起来的。你从amazon的某个CPU的某个晶体管的通断上也是看不出amazon它买一本书的过程的。

文章中写道：”当达尔文写下《物种起源》的时候，没有人知道DNA是做什么的，也没有人想到我们有朝一日还能给它测序。“。确实，神经科学的不发达不是借口，人类探求大自然的足迹，向来都是深一脚浅一脚的。要了解大自然，了解大脑的秘密，我们人类必须且仅能从上往下走，从外往里走，从顶层往底层走。

而数学确更多是从底层的1+1发展起来的，所以我觉得数学是人类的智慧，而科学，是大自然的智慧，我，在这两者中间行走。