作者：Matteo Wong

该程序分析了人们听甚至只是回忆三个节目中句子的fMRI扫描：Modern Love、The Moth Radio Hour和The Anthropocene Reviewed。然后，它使用大脑成像数据来重建这些句子的内容。例如，当一个受试者听到“我还没有驾照”时，程序破译了这个人的大脑扫描的结果，并返回“她甚至还没有开始学习开车”——不是逐字逐句的重新创造，而是与原始句子中表达的想法的近似。该程序还能够查看观看短片的人的fMRI数据，并编写剪辑的近似摘要，这表明人工智能不是从大脑扫描中捕获单个单词，而是基本含义。

然而，即使人类认知长期以来一直启发着“智能”计算机程序的设计，我们大脑的内部运作仍然是一个谜。现在，在这种方法的逆转中，科学家们希望通过使用合成神经网络来研究我们的生物神经网络来更多地了解大脑。麻省理工学院的认知科学家Evelina Fedorenko说，这“无疑导致了几年前我们无法想象的进步。”

该人工智能程序明显接近读心术，在社交媒体和传统媒体上引起了轩然大波。但这项工作的这一方面“更像是一个小把戏”，《自然》杂志研究的主要作者、德克萨斯大学奥斯汀分校(UT Austin)的神经科学家亚历山大·胡特(Alexander Huth)告诉我。这些模型相对不精确，且对参与研究的每个人都进行了微调，而大多数大脑扫描技术提供的数据分辨率非常低;我们离一个可以插入任何人的大脑并了解他们在想什么的程序还很远很远。

这项工作的真正价值在于预测在听或想象单词时，大脑的哪些部分会被激活，这可能会让我们更深入地了解由神经元协同工作创造出语言等人类特性的具体方式。

胡特说，成功地构建一个可以重建句子意义的程序，主要是作为“这些模型实际上捕获了很多关于大脑如何处理语言的原理证明”。在这场新生的人工智能革命之前，神经科学家和语言学家依赖于对大脑语言网络的某种广义的口头描述，这种描述不精确，也很难直接与可观察到的大脑活动联系起来。关于不同大脑区域负责语言的哪些方面的假设，甚至是大脑如何学习语言的基本问题，都很难甚至不可能测试。（也许一个区域可以识别声音，另一个区域处理语法，等等。）但现在科学家可以使用人工智能模型来更好地确定这些过程的确切组成。该研究的另一位主要作者、UT Austin的计算机科学家Jerry Tang表示，这些好处可能不仅限于学术问题——例如，帮助某些残疾人。他告诉我：“我们的最终目标是帮助那些失去说话能力的人恢复沟通。”

对于人工智能可以帮助研究大脑的想法，尤其是在研究语言的神经科学家中，一直存在一些阻力。这是因为擅长发现统计模式的神经网络，似乎缺乏人类处理语言的基本要素，比如理解单词的意思。

机器和人类认知之间的差异也是直观的:像GPT-4这样的程序可以写出不错的文章，在标准化测试中表现出色，它通过处理来自书籍和网页的数万亿字节的数据来学习，而儿童学习一门语言的单词量只有它的1%。“老师们告诉我们，人工神经网络和生物神经网络真的不一样，”神经科学家让-拉马米·金(jean - rsami King)跟我谈到他在2000年代末的研究，“在当时看来这只是一个比喻。”如今，在Meta领导大脑和人工智能研究的金是众多反驳这一旧教条的科学家之一。“我们不再认为这是一个隐喻，”他告诉我。“我们认为(人工智能)是大脑如何处理信息的一个非常有用的模型。”

在过去的几年里，科学家已经表明，高级人工智能程序的内部运作为我们的大脑如何处理语言提供了一个有前途的数学模型。当您在ChatGPT或类似程序中键入句子时，其内部神经网络将该输入表示为一组数字。当一个人听到同一个句子时，fMRI扫描可以捕获他们大脑中的神经元的反应，计算机能够将这些扫描解释为基本上是另一组数字。这些过程在许多句子上重复，以创建两个巨大的数据集：一个是机器如何代表语言，另一个是人类。然后，研究人员可以使用一种称为编码模型的算法映射这些数据集之间的关系。一旦完成，编码模型就可以开始推断：人工智能如何响应一个句子也成为预测大脑中神经元如何响应它的基础。

使用人工智能研究大脑语言网络的新研究似乎每隔几周就会出现一次。麻省理工学院的神经科学家Nancy Kanwisher告诉我，这些模型中的每一个都可以代表“关于大脑中可能发生的事情的计算精确假设”。例如，人工智能可以帮助回答一个悬而未决的问题，即当人类大脑获得一种语言时，它到底打算做什么——不仅仅是一个人正在学习沟通，还有沟通的特定神经机制。这个想法是，如果一个训练有特定目标的计算机模型——例如学习预测序列中的下一个单词或判断句子的语法连贯性——被证明最能预测大脑反应，那么人类的思想可能共享这个目标；也许我们的大脑，如GPT-4，通过确定哪些单词最有可能相互跟随来工作。然后，语言模型的内部运作成为大脑的计算理论。

这些计算方法只有几年的历史，所以有很多分歧和相互竞争的理论。国家心理健康研究所机器学习主任Francisco Pereira告诉我：“没有理由你从语言模型中学到的表征与大脑如何代表一个句子有任何关系。”但这并不意味着一段关系不能存在，有各种方法可以测试它是否存在。与大脑不同，科学家几乎可以无限地分解、检查和操作语言模型——因此，即使人工智能程序不是大脑的完整假设，它们也是研究大脑的强大工具。例如，在麻省理工学院研究大脑和语言的Greta Tuckute告诉我，认知科学家可以尝试预测目标大脑区域的反应，并测试不同类型的句子如何引发不同类型的大脑反应，以弄清楚这些特定的神经元集群做什么，“然后踏入未知的领域。”

目前，人工智能的效用可能不是精确复制那个未知的神经领域，而是设计启发式方法来探索它。麻省理工学院的认知科学家Anna Ivanova在引用一个著名的博尔赫斯寓言时告诉我：“如果你有一张可以再现世界每一个细节的地图，那么这张地图就毫无用处了，因为它和世界一样大。”“所以你需要抽象。”正是通过指定和测试要保留和抛弃什么——在街道、地标和建筑物之间进行选择，然后看看生成的地图有多大用——科学家们才开始在大脑的语言地形中导航。

AI Is Unlocking the Human Brain’s Secrets

Language models similar to ChatGPT have started to transform neuroscience.

By Matteo Wong

If you are willing to lie very still in a giant metal tube for 16 hours and let magnets blast your brain as you listen, rapt, to hit podcasts, a computer just might be able to read your mind. Or at least its crude contours. Researchers from the University of Texas at Austin recently trained an AI model to decipher the gist of a limited range of sentences as individuals listened to them—gesturing toward a near future in which artificial intelligence might give us a deeper understanding of the human mind.

The program analyzed fMRI scans of people listening to, or even just recalling, sentences from three shows: Modern Love, The Moth Radio Hour, and The Anthropocene Reviewed. Then, it used that brain-imaging data to reconstruct the content of those sentences. For example, when one subject heard “I don’t have my driver’s license yet,” the program deciphered the person’s brain scans and returned “She has not even started to learn to drive yet”—not a word-for-word re-creation, but a close approximation of the idea expressed in the original sentence. The program was also able to look at fMRI data of people watching short films and write approximate summaries of the clips, suggesting the AI was capturing not individual words from the brain scans, but underlying meanings.

The findings, published in Nature Neuroscience earlier this month, add to a new field of research that flips the conventional understanding of AI on its head. For decades, researchers have applied concepts from the human brain to the development of intelligent machines. ChatGPT, hyperrealistic-image generators such as Midjourney, and recent voice-cloning programs are built on layers of synthetic “neurons”: a bunch of equations that, somewhat like nerve cells, send outputs to one another to achieve a desired result. Yet even as human cognition has long inspired the design of “intelligent” computer programs, much about the inner workings of our brains has remained a mystery. Now, in a reversal of that approach, scientists are hoping to learn more about the mind by using synthetic neural networks to study our biological ones. It’s “unquestionably leading to advances that we just couldn’t imagine a few years ago,” says Evelina Fedorenko, a cognitive scientist at MIT.

The AI program’s apparent proximity to mind reading has caused uproar on social and traditional media. But that aspect of the work is “more of a parlor trick,” Alexander Huth, a lead author of the Nature study and a neuroscientist at UT Austin, told me. The models were relatively imprecise and fine-tuned for every individual person who participated in the research, and most brain-scanning techniques provide very low-resolution data; we remain far, far away from a program that can plug into any person’s brain and understand what they’re thinking. The true value of this work lies in predicting which parts of the brain light up while listening to or imagining words, which could yield greater insights into the specific ways our neurons work together to create one of humanity’s defining attributes, language.

Read: The difference between speaking and thinking

Successfully building a program that can reconstruct the meaning of sentences, Huth said, primarily serves as “proof-of-principle that these models actually capture a lot about how the brain processes language.” Prior to this nascent AI revolution, neuroscientists and linguists relied on somewhat generalized verbal descriptions of the brain’s language network that were imprecise and hard to tie directly to observable brain activity. Hypotheses for exactly what aspects of language different brain regions are responsible for—or even the fundamental question of how the brain learns a language—were difficult or even impossible to test. (Perhaps one region recognizes sounds, another deals with syntax, and so on.) But now scientists could use AI models to better pinpoint what, precisely, those processes consist of. The benefits could extend beyond academic concerns—assisting people with certain disabilities, for example, according to Jerry Tang, the study’s other lead author and a computer scientist at UT Austin. “Our ultimate goal is to help restore communication to people who have lost the ability to speak,” he told me.

There has been some resistance to the idea that AI can help study the brain, especially among neuroscientists who study language. That’s because neural networks, which excel at finding statistical patterns, seem to lack basic elements of how humans process language, such as an understanding of what words mean. The difference between machine and human cognition is also intuitive: A program like GPT-4, which can write decent essays and excels at standardized tests, learns by processing terabytes of data from books and webpages, while children pick up a language with a fraction of 1 percent of that amount of words. 

In the past few years, scientists have shown that the inner workings of advanced AI programs offer a promising mathematical model of how our minds process language. When you type a sentence into ChatGPT or a similar program, its internal neural network represents that input as a set of numbers. When a person hears the same sentence, fMRI scans can capture how the neurons in their brain respond, and a computer is able to interpret those scans as basically another set of numbers. These processes repeat on many, many sentences to create two enormous data sets: one of how a machine represents language, and another for a human. Researchers can then map the relationship between these data sets using an algorithm known as an encoding model. Once that’s done, the encoding model can begin to extrapolate: How the AI responds to a sentence becomes the basis for predicting how neurons in the brain will fire in response to it, too.

New research using AI to study the brain’s language network seems to appear every few weeks. Each of these models could represent “a computationally precise hypothesis about what might be going on in the brain,” Nancy Kanwisher, a neuroscientist at MIT, told me. For instance, AI could help answer the open question of what exactly the human brain is aiming to do when it acquires a language—not just that a person is learning to communicate, but the specific neural mechanisms through which communication comes about. The idea is that if a computer model trained with a specific objective—such as learning to predict the next word in a sequence or judge a sentence’s grammatical coherence—proves best at predicting brain responses, then it’s possible the human mind shares that goal; maybe our minds, like GPT-4, work by determining what words are most likely to follow one another. The inner workings of a language model, then, become a computational theory of the brain.

Read: ChatGPT is already obsolete

These computational approaches are only a few years old, so there are many disagreements and competing theories. “There is no reason why the representation you learn from language models has to have anything to do with how the brain represents a sentence,” Francisco Pereira, the director of machine learning for the National Institute of Mental Health, told me. But that doesn’t mean a relationship cannot exist, and there are various ways to test whether it does. Unlike the brain, scientists can take apart, examine, and manipulate language models almost infinitely—so even if AI programs aren’t complete hypotheses of the brain, they are powerful tools for studying it. For instance, cognitive scientists can try to predict the responses of targeted brain regions, and test how different types of sentences elicit different types of brain responses, to figure out what those specific clusters of neurons do “and then step into territory that is unknown,” Greta Tuckute, who studies the brain and language at MIT, told me.

For now, the utility of AI may not be to precisely replicate that unknown neurological territory, but to devise heuristics for exploring it. “If you have a map that reproduces every little detail of the world, the map is useless because it’s the same size as the world,” Anna Ivanova, a cognitive scientist at MIT, told me, invoking a famous Borges parable. “And so you need abstraction.” It is by specifying and testing what to keep and jettison—choosing among streets and landmarks and buildings, then seeing how useful the resulting map is—that scientists are beginning to navigate the brain’s linguistic terrain.