English

The idea that the mind is a kind of computer is one of the most consequential ideas of the twentieth century, and it is the foundation of the relationship between computer science and the sciences of mind. When the theory of computation was developed in the 1930s and the first computers were built in the 1940s, it became possible to think about thinking in a new way: perhaps the mind, too, processes information — takes in representations, transforms them according to rules, produces outputs — and perhaps the activity we call thinking is, at bottom, computation. This hypothesis launched cognitive science as a discipline and shaped artificial intelligence, and it remains the central organizing idea, and the central battleground, of the relationship between computer science and the study of mind.

The relationship is genuinely bidirectional and genuinely contested, which makes it different from the relationships in the preceding sections. Computer science gave the sciences of mind a theory — the computational theory of mind, the idea that mental processes are computational — and a set of tools and metaphors for studying cognition. The sciences of mind gave computer science inspiration — neural networks are loosely modeled on the brain, and the ambition of artificial intelligence is to reproduce or exceed human cognitive capacities. But the relationship is contested at its core: whether the mind really is computational, whether artificial neural networks tell us anything about brains, whether intelligence requires a body, whether the impressive capabilities of modern AI systems constitute or illuminate cognition — these are live disputes, and the boundary between the fields is a place of argument as much as agreement.

This section examines three dimensions of the relationship: intelligence (whether and how computation can constitute or explain it), representation (the central concept that both fields are built on and disagree about), and embodiment (the challenge to the view that cognition is abstract computation independent of a body). It connects to artificial intelligence (§5.1–§5.7), which is the engineering attempt to build intelligent systems, but its concern is the relationship to the sciences of mind — what computation tells us about cognition, and what cognition tells us about computation.

Background: The artificial intelligence material (§5.1–§5.7), especially the symbolic and connectionist traditions, provides the computing side; cognitive science, psychology, neuroscience, and philosophy of mind provide the neighboring disciplines.

The Relationship: Computation as a Theory of Mind

Intelligence: The Computational Theory of Mind and Its Critics

The computational theory of mind is the hypothesis that mental processes are computational processes — that thinking is, literally and not just metaphorically, a kind of computation. It is the foundational idea of the relationship, and tracing its fortunes is the best way to understand the relationship’s structure.

The idea emerged from the convergence of several developments around mid-century. The theory of computation showed that a simple mechanical process could perform any computation; the first computers demonstrated that machines could manipulate symbols; and researchers in the new fields of cybernetics and artificial intelligence proposed that the mind might work the same way. The 1956 Dartmouth conference that founded AI was premised on the conjecture that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” Cognitive science emerged in the same period as the interdisciplinary study of mind premised on the computational hypothesis, drawing together psychology, linguistics, computer science, and philosophy around the idea that cognition is information processing.

The classical version — the “symbolic” or “good old-fashioned” computational theory of mind — held that cognition is the rule-governed manipulation of symbolic representations, much as a classical computer manipulates symbols. Newell and Simon’s physical symbol system hypothesis (1976) stated it sharply: a physical symbol system (a system that manipulates symbolic representations) has the necessary and sufficient means for general intelligent action. On this view, the mind is a symbol-processing system, and intelligence is the right kind of symbol processing — a view that motivated the classical AI of §5.1 and the cognitive science that modeled cognition as symbolic computation.

The critics challenged this on several fronts, and the challenges define the contested core of the relationship. Searle’s Chinese Room argument (1980) attacked the claim that symbol manipulation could constitute understanding: a person manipulating Chinese symbols by rules, without understanding Chinese, shows (Searle argued) that symbol manipulation is not sufficient for genuine understanding, however intelligent the behavior it produces. The argument is still debated (the responses, particularly the “systems reply,” are substantial), but it crystallized the question of whether computation could constitute mind or only simulate its behavior. Dreyfus, drawing on phenomenology, argued that human intelligence depends on embodied, contextual, non-rule-governed know-how that symbolic AI could not capture — a critique that looked vindicated by classical AI’s failures and that connects to the embodiment dimension below. These critics did not settle the question, but they established that the computational theory of mind is a contestable hypothesis, not an established fact, and the contest continues.

Representation: The Concept Both Fields Are Built On

Representation is the concept at the center of both computer science and cognitive science, and it is where the two fields are most deeply connected and most deeply in disagreement. A representation is something that stands for something else — a symbol for an object, a pattern for a category, a state for a situation — and the notion that cognition involves representations, internal states that stand for things in the world, is shared across the fields. But what representations are, how they work, and whether cognition really involves them are contested.

The classical view holds that cognitive representations are symbolic and structured, like the data structures of a program: the mind contains symbols that represent things, combined according to rules into structured representations (a representation of “the cat is on the mat” has parts corresponding to the cat, the mat, and the relation), and cognition operates on these structured symbolic representations. This view, congenial to symbolic AI, gives a clear account of how representations can be combined and manipulated, and it explains the systematicity of thought (that anyone who can think “the cat is on the mat” can think “the mat is on the cat”) in terms of structured representations.

The connectionist view, which rose with neural networks, holds that representations are distributed patterns of activation across many units, not discrete symbols. In a neural network, a concept is represented not by a single symbol but by a pattern across many neurons, and the representation is learned from data rather than given. This view, congenial to the connectionism of §5.2–§5.3 and to the actual structure of the brain (which is a network of neurons, not obviously a symbol processor), explains learning and graceful degradation and similarity in ways the symbolic view struggles with, but it has historically struggled to explain the systematicity and compositionality that the symbolic view handles naturally. The debate between symbolic and connectionist views of representation — fought most sharply in the 1980s (the Fodor-Pylyshyn critique of connectionism, 1988) and reopened by the success of deep learning — is the central theoretical dispute at the boundary of the fields, and it remains unresolved: modern large language models are connectionist systems that exhibit surprising degrees of the systematicity that connectionism was said to lack, which has reopened rather than settled the question.

The interpretability research of §5.6 is, from this perspective, an attempt to answer the representation question empirically for artificial systems: what do the representations in a trained neural network actually look like, are they interpretable, do they correspond to anything like concepts? The finding that features (directions in activation space) can correspond to interpretable concepts, and that they are superposed rather than cleanly separated, is empirical evidence about the nature of distributed representation that bears on the old debate. The relationship between the fields is here at its most active: the question of what representations are, debated philosophically for decades, is now partly an empirical question about artificial neural networks, investigated with the tools of interpretability.

Embodiment: The Challenge to Abstract Computation

The embodiment challenge questions the assumption, shared by the classical computational theory of mind, that cognition is abstract information processing that could in principle happen in any substrate, independent of having a body. The embodied cognition movement argues that cognition is deeply shaped by the body and its interaction with the environment — that thinking is not abstract symbol manipulation in a disembodied mind but is grounded in sensorimotor experience, structured by the body’s form, and extended into the environment and the tools we use.

The challenge has several sources. The failures of classical symbolic AI — its inability to handle the open-ended, context-dependent, common-sense reasoning that humans perform effortlessly — suggested to critics like Dreyfus that something was missing from the disembodied symbol-processing picture, and that the something was the embodied, situated character of human intelligence. Robotics (§6.4) provided concrete evidence: building robots that act in the physical world revealed that much of what seems like it should require abstract reasoning is better handled by tight coupling between perception and action, with the body and environment doing work that the disembodied view attributed to internal computation. Cognitive science research on how abstract concepts are grounded in bodily experience (Lakoff and Johnson’s work on how abstract thought is structured by bodily metaphors) suggested that even abstract cognition is shaped by embodiment.

The embodiment challenge does not necessarily refute the computational theory of mind — one can hold that embodied cognition is still computational, just computation that essentially involves the body and environment rather than being confined to a disembodied symbol processor. But it reshapes the relationship between computation and cognition: if cognition is essentially embodied, then understanding it requires understanding the body-environment-brain system as a whole, not just the abstract information processing, and building genuinely intelligent systems might require giving them bodies and environments rather than just better algorithms. This is a live question for artificial intelligence: whether the path to general intelligence runs through embodiment, or whether the disembodied large language models that have proven so capable show that embodiment is not necessary after all. The success of LLMs — disembodied systems trained only on text that nonetheless exhibit broad capabilities — is evidence that the strongest embodiment claims are too strong, while the persistent gaps in their understanding of the physical and social world they have never experienced are evidence that embodiment may matter after all. The question is open, and it is being investigated empirically as much as argued philosophically.

Brains and Artificial Neural Networks: Correspondence and Its Limits

The relationship to neuroscience specifically centers on the artificial neural network, which was loosely inspired by the brain and is sometimes taken to tell us about it — a correspondence that is real but limited, and whose limits are as important as its existence. Artificial neural networks were inspired by a simplified model of neurons (McCulloch and Pitts, 1943, modeled neurons as threshold logic units), and the architecture — units that compute weighted sums and pass them through nonlinearities, connected in layers — retains a loose resemblance to networks of biological neurons. This inspiration was real and productive: the brain was the existence proof that a network of simple units could produce intelligence, and it motivated the neural network research that became deep learning.

But the correspondence between artificial and biological neural networks is loose, and treating artificial networks as models of the brain is fraught. Biological neurons are vastly more complex than the units of artificial networks (they have intricate dendritic computation, diverse types, temporal dynamics, neuromodulation); the brain’s architecture differs profoundly from artificial networks (it is recurrent, not feedforward; it learns without backpropagation, which is not biologically plausible in its standard form; it is structured by evolution and development in ways artificial networks are not). The artificial neural network is inspired by the brain in the way an airplane is inspired by a bird — capturing some principle (lift, distributed parallel computation) while differing enormously in mechanism. Treating the success of artificial networks as telling us how the brain works, or treating the brain as validating the artificial networks, mostly overstates a loose analogy.

Yet the relationship is becoming more interesting as both fields advance. Computational neuroscience uses the tools and concepts of computer science to model the brain, and the models are tested against neural data. The comparison between representations in artificial neural networks and representations in the brain (measured by neuroscience) has become a research area: in some cases, the representations that deep networks learn for vision resemble, statistically, the representations measured in visual cortex, which is suggestive even if its meaning is debated. Predictive coding and other theories propose that the brain implements something like the computations that machine learning studies. The dialogue between neuroscience and machine learning — each informing the other, with appropriate caution about the limits of the correspondence — is an active frontier, more productive when it respects the differences than when it collapses them into a false identity.

What This Perspective Changes

Seeing the relationship between computer science and the sciences of mind changes how a practitioner understands intelligence, both artificial and natural.

The first change is a properly contested understanding of the computational theory of mind. The practitioner who sees that the idea “the mind is a computer” is a profound and contestable hypothesis — not an established fact, not an empty metaphor, but a serious claim with serious critics — holds it with the appropriate combination of seriousness and skepticism. They can engage with the question of whether AI systems understand or merely simulate understanding, whether intelligence requires more than computation, with the awareness that these are open questions rather than settled ones.

The second change is depth on the representation question, which is central to both AI and cognitive science. The practitioner who understands the symbolic-connectionist debate, and sees that modern AI has reopened it, understands the deepest theoretical question in artificial intelligence and cognitive science: what are representations, and how does cognition use them? This understanding makes the interpretability research of §5.6 legible as an empirical attack on an old philosophical question, and it provides the framework for thinking about what the representations in modern AI systems actually are.

The third change is calibration about the brain-AI correspondence. The practitioner who understands that artificial neural networks are loosely inspired by but profoundly different from brains avoids the two errors: overstating the correspondence (treating deep learning as telling us how the brain works, or the brain as validating deep learning) and dismissing it entirely (missing the genuine and growing dialogue between the fields). The calibrated view — a real but limited correspondence, more productive when it respects the differences — is the basis for engaging with computational neuroscience and the comparison of artificial and biological representations.

The fourth change is engagement with the embodiment question, which bears on the future of AI. The practitioner who understands the embodiment challenge — and the evidence on both sides from the success of disembodied LLMs and their persistent gaps — can think about whether the path to more general intelligence runs through embodiment, which is a live question for the field’s direction. Whether intelligence requires a body is not only a philosophical question but a question about how to build more capable systems, and the practitioner who has engaged with it thinks about the field’s trajectory more deeply.

Resources

Foundational and Bridging Texts

Margaret Boden’s Mind as Machine: A History of Cognitive Science (two volumes, Oxford, 2006) is the comprehensive history of the relationship, tracing how the computational idea of mind developed across psychology, AI, linguistics, and philosophy. It is long and authoritative; for the full story of how computer science and the sciences of mind became entangled, it is the definitive source.

For the computational theory of mind and its philosophical status, the philosophy of mind literature is the source (and connects to §9.9). Fodor’s The Language of Thought (1975) is the classic statement of the symbolic computational theory of mind; the critiques (Searle’s “Minds, Brains, and Programs,” 1980, for the Chinese Room; Dreyfus’s What Computers Can’t Do, 1972, revised as What Computers Still Can’t Do, 1992, for the embodiment-grounded critique) are the essential counterpoints. Reading the symbolic theory and its critics together is the way to understand the contested core.

For an accessible synthesis, Pinker’s How the Mind Works (1997) presents the computational-evolutionary view of mind for a general audience, and Hofstadter’s Gödel, Escher, Bach (1979) explores the relationship between formal systems, computation, and mind with unusual depth and creativity — a book that has drawn many people into thinking seriously about computation and cognition.

Resource Role Type
Boden, Mind as Machine (2 vols.) Comprehensive history of the relationship Reference
Fodor, The Language of Thought Symbolic computational theory of mind Depth
Searle, “Minds, Brains, and Programs” (1980) The Chinese Room; the central critique Depth
Dreyfus, What Computers Still Can’t Do The embodiment-grounded critique Depth
Hofstadter, Gödel, Escher, Bach Formal systems, computation, and mind Depth
Marr, Vision Computational levels of analysis for perception and intelligence Depth

Representation, Connectionism, and the Brain

For the symbolic-connectionist debate, the primary sources are the best entry: the PDP volumes (Rumelhart and McClelland, Parallel Distributed Processing, 1986) launched modern connectionism, and Fodor and Pylyshyn’s “Connectionism and Cognitive Architecture” (1988) is the systematic critique that defined the debate. Reading them together frames the central representational dispute, which the success of deep learning has reopened.

For cognitive science broadly, the standard texts — such as Cognitive Science: An Introduction to the Science of the Mind (Bermúdez) — provide the disciplinary context. For the relationship between deep learning and neuroscience specifically, the contemporary literature comparing artificial and biological representations (work on whether deep network representations resemble those in cortex) is in research papers; the reviews in journals like Nature Neuroscience and the work of researchers bridging the fields are the current sources.

For embodiment, Lakoff and Johnson’s Philosophy in the Flesh (1999) and Metaphors We Live By (1980) develop the view that abstract cognition is grounded in bodily experience, and Andy Clark’s Being There: Putting Brain, Body, and World Together Again (1997) and Supersizing the Mind (2008) are the philosophical treatments of embodied and extended cognition.

Resource Role Type
Rumelhart & McClelland, Parallel Distributed Processing Connectionism’s foundational text Depth
Fodor & Pylyshyn, “Connectionism and Cognitive Architecture” (1988) The systematicity critique of connectionism Depth
Lakoff & Johnson, Philosophy in the Flesh Embodied grounding of abstract cognition Depth
Clark, Being There / Supersizing the Mind Embodied and extended cognition Depth
Computational neuroscience reviews (current) Artificial-biological representation comparison Reference
Neuromatch Academy Computational Neuroscience (free) Modern computational neuroscience course and tutorials Practice
Cognitive Computational Neuroscience proceedings (free/current) Current AI-cognition-neuroscience research venue Reference
Allen Brain Map / Allen Institute open data (free) Large-scale open neuroscience data for computational work Practice

Traps

Trap Why it misleads Better response
Treating “the mind is a computer” as established fact or empty metaphor The computational theory of mind is neither a proven fact nor an empty metaphor; it is a profound, serious, and contested hypothesis. Treating it as established (the mind simply is a computer, question closed) ignores the serious critiques; treating it as mere metaphor (a loose analogy with no real content) ignores how seriously and literally it is meant and how much it has shaped both fields. Hold it as the contested hypothesis it is: a serious claim that mental processes are literally computational, with substantial arguments for and substantial critics against (Searle, Dreyfus). Engage with the arguments on both sides. The question of whether AI systems understand or merely simulate understanding is a live form of this dispute, not a settled matter.
Overstating the brain-AI correspondence Artificial neural networks are loosely inspired by the brain, and the loose inspiration is often inflated into a claim that deep learning reveals how the brain works or that the brain validates deep learning. The correspondence is real but limited: biological neurons and brain architecture differ profoundly from artificial networks, and treating the analogy as identity overstates it badly. Respect the limits of the correspondence. Artificial neural networks relate to brains as airplanes to birds — sharing some principle while differing enormously in mechanism. The genuine and growing dialogue between neuroscience and machine learning is more productive when it respects the differences than when it collapses them. Treat claims that deep learning explains the brain, or vice versa, with calibrated skepticism.
Assuming the representation question is settled The symbolic-connectionist debate about the nature of representation is sometimes treated as having been settled — by the victory of symbolic AI in an earlier era, or by the victory of connectionism in the deep learning era. It is not settled: deep learning’s success reopened it, with modern systems showing connectionist representations that exhibit more of the systematicity the symbolic side claimed than connectionism was supposed to allow. Treat the representation question as open and actively investigated. The interpretability research of §5.6 is attacking it empirically for artificial systems; the comparison with biological representation is attacking it from neuroscience; and the theoretical debate continues. What representations are, and how cognition uses them, is among the deepest open questions at the boundary of the fields.
Settling the embodiment question prematurely in either direction Both strong positions on embodiment are overconfident given the evidence. The strong embodiment claim (intelligence requires a body) is challenged by disembodied LLMs that exhibit broad capabilities; the strong disembodiment claim (embodiment is irrelevant, as LLMs show) is challenged by the persistent gaps in those systems’ understanding of the physical and social world they have never experienced. Hold the embodiment question open, with attention to the evidence on both sides. The success of disembodied LLMs is real evidence that the strongest embodiment claims are too strong; the gaps in their grounded understanding are real evidence that embodiment may matter. Whether the path to more general intelligence runs through embodiment is a live empirical question about the field’s direction, not a settled matter to be assumed either way.

中文

心智是一种计算机这一思想，是二十世纪最有影响力的思想之一，也是计算机科学与心智科学之间关系的基础。1930 年代计算理论被发展出来，1940 年代第一批计算机被建造出来之后，人们开始能够以一种新方式思考“思考”本身：也许心智也在处理信息——接收表征，按照规则变换它们，产生输出——也许我们称为思考的活动，在底层上就是计算。这个假设催生了认知科学这门学科，并塑造了人工智能；它至今仍然是计算机科学与心智研究之间关系的核心组织思想，也是核心争议场域。

这种关系是真正双向的，也是真正有争议的，这使它不同于前几节中的关系。计算机科学给了心智科学一种理论——心智的计算理论，也就是心理过程是计算过程这一思想——以及一组研究认知的工具和隐喻。心智科学则给了计算机科学灵感——神经网络在宽泛意义上以大脑为模型，而人工智能的抱负就是复制或超越人类认知能力。但这种关系在核心处存在争议：心智是否真的就是计算性的？人工神经网络是否告诉了我们关于大脑的任何东西？智能是否需要身体？现代 AI 系统令人印象深刻的能力是否构成或照亮了认知？这些都是仍在进行的争论；两个领域的边界不仅是共识之地，也是争论之地。

本节考察这种关系的三个维度：智能（计算是否以及如何能够构成或解释智能）、表征（两个领域都建立其上、同时又存在分歧的核心概念），以及具身性（对“认知是独立于身体的抽象计算”这一观点的挑战）。它连接到人工智能（§5.1–§5.7），也就是构建智能系统的工程尝试；但这里关心的是它与心智科学的关系——计算告诉我们关于认知的什么，认知又告诉我们关于计算的什么。

背景知识：人工智能材料（§5.1–§5.7），尤其是符号主义和连接主义传统，提供计算一侧；认知科学、心理学、神经科学和心智哲学提供相邻学科背景。

这种关系：计算作为心智理论

智能：心智的计算理论及其批评者

心智的计算理论是一种假设：心理过程就是计算过程——思考在字面意义上，而不只是隐喻意义上，是一种计算。它是这种关系的基础思想；追踪它的命运，是理解这种关系结构的最佳方式。

这一思想出现在二十世纪中期若干发展汇合之处。计算理论表明，一个简单的机械过程可以执行任何计算；第一批计算机证明机器可以操纵符号；而控制论和人工智能这些新领域的研究者提出，心智也许以同样方式工作。1956 年创立 AI 的 Dartmouth 会议，正是建立在这样一个猜想之上：“学习的每一个方面，或者智能的任何其他特征，原则上都可以被如此精确地描述，以至于机器可以被制造出来模拟它。”认知科学也在同一时期兴起，作为一门以计算假设为前提的跨学科心智研究，把心理学、语言学、计算机科学和哲学围绕“认知是信息处理”这一思想聚合起来。

经典版本——“符号主义”或 “good old-fashioned” 的心智计算理论——认为认知是对符号表征的规则化操纵，很像经典计算机操纵符号。Newell 和 Simon 的物理符号系统假说（1976）把这一点说得很尖锐：一个物理符号系统（即操纵符号表征的系统）拥有进行一般智能行动的必要且充分手段。在这种观点下，心智是一个符号处理系统，智能则是正确类型的符号处理——这一观点推动了 §5.1 中的经典 AI，也推动了把认知建模为符号计算的认知科学。

批评者从多个方向挑战这一观点，而这些挑战定义了这种关系中有争议的核心。Searle 的中文房间论证（1980）攻击的是“符号操纵可以构成理解”这一主张：一个人按照规则操纵中文符号，却不理解中文，这表明——按照 Searle 的说法——无论产生的行为多么智能，符号操纵都不足以构成真正理解。这个论证至今仍在争论中（回应，尤其是“系统回应”，也很有分量），但它清晰凝结了一个问题：计算能否构成心智，还是只能模拟心智的行为？Dreyfus 借助现象学提出，人类智能依赖具身的、情境化的、非规则支配的 know-how，而符号 AI 无法捕捉这些东西——这一批评似乎被经典 AI 的失败所验证，并且连接到下文的具身性维度。这些批评者并没有解决问题，但他们确立了这样一点：心智的计算理论是一种可争辩的假设，而不是已被确立的事实；这种争论仍在继续。

表征：两个领域都建立其上的概念

表征是计算机科学和认知科学共同位于中心位置的概念，也是二者连接最深、分歧也最深的地方。一个表征，就是某个代表其他东西的东西——代表对象的符号，代表类别的模式，代表情境的状态。认知涉及表征，也就是内部状态代表世界中的事物，这一观念在两个领域中都被共享。但表征是什么、如何工作，以及认知是否真的涉及表征，仍然存在争议。

经典观点认为，认知表征是符号性的、结构化的，就像程序中的数据结构：心智包含代表事物的符号，这些符号按照规则组合成结构化表征（“猫在垫子上”的表征具有分别对应猫、垫子和关系的部分），认知则作用于这些结构化符号表征。这个观点与符号 AI 相容，它清楚解释了表征如何能够被组合和操纵，也用结构化表征解释了思维的系统性（任何能想“猫在垫子上”的人，也能想“垫子在猫上”）。

连接主义观点随着神经网络兴起而上升，它认为表征不是离散符号，而是分布在许多单元上的激活模式。在神经网络中，一个概念不是由单个符号表示，而是由许多神经元上的模式表示；而且表征是从数据中学得的，而不是预先给定的。这种观点与 §5.2–§5.3 中的连接主义以及大脑的实际结构相容（大脑是神经元网络，并不明显是符号处理器），它以符号观点难以处理的方式解释学习、优雅退化和相似性；但它在历史上很难解释符号观点自然处理的系统性和组合性。关于表征的符号主义与连接主义之争——在 1980 年代争论最激烈（Fodor-Pylyshyn 对连接主义的批评，1988），并因深度学习的成功而重新开启——是两个领域边界上的核心理论争议，而且仍未解决：现代大语言模型是连接主义系统，却表现出令人惊讶程度的系统性，而系统性曾被认为是连接主义所缺乏的东西；这重新打开了问题，而不是解决了问题。

从这个角度看，§5.6 的可解释性研究，是一种试图对人工系统中的表征问题给出经验答案的努力：训练后的神经网络中，表征到底是什么样子？它们是否可解释？它们是否对应于类似概念的东西？一些发现表明，特征（激活空间中的方向）可以对应可解释概念，而且这些特征是叠加的，而不是清晰分离的。这些结果提供了关于分布式表征性质的经验证据，并且与旧争论相关。这里正是两个领域关系最活跃的地方：表征是什么这个被哲学争论了几十年的问题，如今部分变成了一个关于人工神经网络的经验问题，并由可解释性工具加以研究。

具身性：对抽象计算的挑战

具身性挑战质疑的是经典心智计算理论所共享的假设：认知是抽象信息处理，原则上可以发生在任何基质中，而不依赖身体。具身认知运动认为，认知深受身体及其与环境互动的塑造——思维并不是在一个无身体心智中进行抽象符号操纵，而是扎根于感觉运动经验，受身体形式组织，并延伸到环境和我们使用的工具中。

这一挑战有几个来源。经典符号 AI 的失败——它无法处理人类毫不费力完成的开放式、依赖语境的常识推理——让 Dreyfus 这样的批评者认为，无身体符号处理图景中缺失了某种东西，而这个东西就是人类智能的具身性和情境性。机器人学（§6.4）提供了具体证据：构建在物理世界中行动的机器人表明，许多看起来似乎需要抽象推理的事情，实际上更适合由感知与行动之间的紧密耦合处理，其中身体和环境承担了无身体观点归因于内部计算的工作。关于抽象概念如何扎根于身体经验的认知科学研究（例如 Lakoff 和 Johnson 关于抽象思维如何由身体隐喻结构化的工作）则暗示，即使抽象认知也受到具身性的塑造。

具身性挑战并不必然反驳心智的计算理论——一个人可以认为具身认知仍然是计算性的，只是这种计算本质上涉及身体和环境，而不是被限制在一个无身体的符号处理器中。但它重塑了计算与认知之间的关系：如果认知本质上是具身的，那么理解它就需要把身体—环境—大脑系统作为整体来理解，而不只是理解抽象信息处理；构建真正智能的系统，也许需要给它们身体和环境，而不只是更好的算法。这对人工智能来说是一个正在进行的问题：通向通用智能的路径是否经过具身性？还是那些已被证明如此强大的无身体大语言模型表明，具身性其实并非必要？LLM 的成功——只在文本上训练、没有身体的系统，却表现出广泛能力——证明最强版本的具身性主张过强；而它们在从未经验过的物理和社会世界理解上的持续缺口，则提供了具身性可能仍然重要的证据。这个问题是开放的，而且它既通过经验研究被考察，也通过哲学争论被讨论。

大脑与人工神经网络：对应及其限度

与神经科学的关系，尤其围绕人工神经网络展开。人工神经网络在宽泛意义上受到大脑启发，有时也被认为能告诉我们关于大脑的事情——这种对应是真实存在的，但也是有限的；它的限度和它的存在同样重要。人工神经网络受到简化神经元模型的启发（McCulloch 和 Pitts，1943，把神经元建模为阈值逻辑单元）；其架构——计算加权和并通过非线性函数的单元，按层连接——仍然与生物神经元网络保持松散相似。这种启发是真实且富有生产力的：大脑是一个存在性证明，证明由简单单元组成的网络可以产生智能；它推动了后来成为深度学习的神经网络研究。

但人工神经网络与生物神经网络之间的对应是松散的，把人工网络当作大脑模型是很危险的。生物神经元远比人工网络中的单元复杂（它们具有复杂的树突计算、多样类型、时间动力学和神经调制）；大脑架构与人工网络也有深刻差异（它是循环的，而不是前馈的；它不通过标准形式中那种在生物学上不太可信的反向传播来学习；它受到进化和发育的结构化，而人工网络并非如此）。人工神经网络受到大脑启发的方式，就像飞机受到鸟启发一样——捕捉某些原则（升力，分布式并行计算），但在机制上差异巨大。把人工网络的成功视为告诉我们大脑如何工作，或者把大脑视为验证人工网络，通常都会夸大一个松散类比。

不过，随着两个领域发展，这种关系正在变得更有意思。计算神经科学使用计算机科学的工具和概念来建模大脑，并把这些模型与神经数据比较。人工神经网络中的表征与神经科学测得的大脑表征之间的比较，已经成为一个研究领域：在某些情况下，深度网络为视觉学习到的表征，在统计上类似于视觉皮层中测得的表征；这很有启发性，尽管其意义仍有争议。预测编码等理论提出，大脑实现了某种类似机器学习所研究计算的东西。神经科学与机器学习之间的对话——二者相互提供信息，同时适当谨慎地对待对应关系的限度——是一个活跃前沿；当它尊重差异，而不是把二者压扁成虚假同一性时，最有生产力。

这种视角会改变什么

看清计算机科学与心智科学之间的关系，会改变实践者理解智能的方式，无论是人工智能还是自然智能。

第一个变化，是对心智计算理论形成恰当的争议性理解。一个实践者如果看见“心智是一台计算机”这一思想是一个深刻且可争辩的假设——不是已确立事实，不是空洞隐喻，而是一个有严肃批评者的严肃主张——就会以恰当的严肃与怀疑来持有它。他们能够进入这样的问题：AI 系统是否理解，还是仅仅模拟理解？智能是否需要不止计算？并且明白这些都是开放问题，而不是已解决问题。

第二个变化，是对表征问题形成深度理解，而表征问题正是 AI 和认知科学的核心。一个理解符号主义—连接主义争论，并看见现代 AI 如何重新开启这一争论的实践者，会理解人工智能和认知科学中最深的理论问题：表征是什么，认知如何使用它们？这种理解会使 §5.6 的可解释性研究变得可读：它是对一个古老哲学问题的经验攻击；同时，它也提供了思考现代 AI 系统中的表征究竟是什么的框架。

第三个变化，是对大脑—AI 对应关系形成校准。一个理解人工神经网络在宽泛意义上受大脑启发、但又与大脑有深刻差异的实践者，会避免两种错误：夸大对应关系（把深度学习当作告诉我们大脑如何工作，或把大脑当作验证深度学习）和完全否定对应关系（错过两个领域之间真实且正在增长的对话）。经过校准的观点——真实但有限的对应，而且在尊重差异时更有生产力——是进入计算神经科学，以及比较人工和生物表征的基础。

第四个变化，是进入具身性问题，而这关系到 AI 的未来。一个理解具身性挑战的实践者——并且看到来自无身体 LLM 成功及其持续缺口的双边证据——能够思考通向更通用智能的路径是否需要经过具身性，而这是这个领域方向上的现实问题。智能是否需要身体，不只是哲学问题，也是关于如何构建更强系统的问题；认真处理过这个问题的实践者，会更深入地思考这个领域的轨迹。

资源

基础性与桥梁性文本

Margaret Boden 的 Mind as Machine: A History of Cognitive Science（两卷本，Oxford，2006）是关于这种关系的综合史，追踪心智的计算思想如何在心理学、AI、语言学和哲学之间发展。它篇幅很长，也具有权威性；如果要完整理解计算机科学与心智科学如何纠缠在一起，它是最权威的来源。

对于心智计算理论及其哲学地位，心智哲学文献是来源（也连接到 §9.9）。Fodor 的 The Language of Thought（1975）是符号主义心智计算理论的经典表述；相应批评则是必要对照：Searle 的 “Minds, Brains, and Programs”（1980）针对中文房间，Dreyfus 的 What Computers Can’t Do（1972，后修订为 What Computers Still Can’t Do，1992）提出以具身性为基础的批评。把符号理论与其批评者放在一起阅读，是理解争议核心的方式。

对于可进入的综合性处理，Pinker 的 How the Mind Works（1997）面向普通读者呈现心智的计算—进化观点；Hofstadter 的 Gödel, Escher, Bach（1979）则以异常深度和创造性探索形式系统、计算与心智之间的关系——这本书曾吸引许多人认真思考计算和认知。

资源 作用 类型
Boden, Mind as Machine（两卷） 这种关系的综合史 参考
Fodor, The Language of Thought 符号主义心智计算理论 深入
Searle, “Minds, Brains, and Programs”（1980） 中文房间；核心批评 深入
Dreyfus, What Computers Still Can’t Do 以具身性为基础的批评 深入
Hofstadter, Gödel, Escher, Bach 形式系统、计算与心智 深入
Marr, Vision 面向感知与智能的计算层次分析 深入

表征、连接主义与大脑

对于符号主义—连接主义之争，一手文献是最好的入口：PDP 系列（Rumelhart 和 McClelland，Parallel Distributed Processing，1986）开启了现代连接主义，而 Fodor 和 Pylyshyn 的 “Connectionism and Cognitive Architecture”（1988）是定义这场争论的系统批评。把二者一起阅读，可以框定核心表征争议，而深度学习的成功正是重新开启了它。

对于广义认知科学，标准教材——如 Bermúdez 的 Cognitive Science: An Introduction to the Science of the Mind——提供学科背景。至于深度学习与神经科学之间的关系，当代关于人工表征和生物表征比较的文献（研究深度网络表征是否类似皮层表征）主要存在于研究论文中；Nature Neuroscience 等期刊中的综述，以及连接两个领域的研究者的工作，是当前来源。

对于具身性，Lakoff 和 Johnson 的 Philosophy in the Flesh（1999）和 Metaphors We Live By（1980）发展了抽象认知扎根于身体经验的观点；Andy Clark 的 Being There: Putting Brain, Body, and World Together Again（1997）和 Supersizing the Mind（2008）则是关于具身认知和延展认知的哲学处理。

资源 作用 类型
Rumelhart & McClelland, Parallel Distributed Processing 连接主义的基础文本 深入
Fodor & Pylyshyn, “Connectionism and Cognitive Architecture”（1988） 对连接主义系统性的批评 深入
Lakoff & Johnson, Philosophy in the Flesh 抽象认知的具身基础 深入
Clark, Being There / Supersizing the Mind 具身认知与延展认知 深入
Computational neuroscience reviews（当前） 人工—生物表征比较 参考
Neuromatch Academy Computational Neuroscience（免费） 现代计算神经科学课程与教程 实践
Cognitive Computational Neuroscience proceedings（免费/当前） 当前 AI—认知—神经科学研究场域 参考
Allen Brain Map / Allen Institute open data（免费） 面向计算工作的开放大规模神经科学数据 实践

陷阱

陷阱 为什么会误导 更好的回应
把“心智是一台计算机”当作已确立事实或空洞隐喻 心智的计算理论既不是已证明事实，也不是空洞隐喻；它是一个深刻、严肃且有争议的假设。把它当成已确立事实（心智就是计算机，问题结束）会忽视严肃批评；把它当成纯粹隐喻（一种没有真实内容的松散类比）则会忽视它被多么严肃、字面地提出，以及它对两个领域产生了多大影响。 把它当作它本来的样子：一个有分量的争议性假设，即心理过程在字面意义上是计算性的；它有强有力的支持论证，也有强有力的批评者（Searle、Dreyfus）。同时处理双方论证。AI 系统是否理解，还是只是在模拟理解，是这场争论的当代表现，而不是已经解决的问题。
夸大大脑—AI 对应关系 人工神经网络受到大脑的松散启发，而这种松散启发常常被夸大成“深度学习揭示了大脑如何工作”或“大脑验证了深度学习”。对应是真实的，但有限：生物神经元和大脑架构与人工网络有深刻差异，把类比当作同一性会严重夸大。 尊重这种对应的限度。人工神经网络与大脑的关系，就像飞机与鸟的关系——共享某些原则，但机制上差异巨大。神经科学与机器学习之间真实且不断增长的对话，在尊重差异时比在抹平差异时更有生产力。对于“深度学习解释大脑”或“反过来”的主张，应保持校准过的怀疑。
以为表征问题已经解决 关于表征性质的符号主义—连接主义争论，有时被视为已经被解决——要么被早期符号 AI 的胜利解决，要么被深度学习时代连接主义的胜利解决。但它并没有解决：深度学习的成功重新开启了这个问题，因为现代系统表现出连接主义表征中原本被认为不可能具有的更多系统性。 把表征问题视为开放且正在被积极研究的问题。§5.6 的可解释性研究正在从人工系统一侧经验性地攻击它；与生物表征的比较则从神经科学一侧攻击它；理论争论也仍在继续。表征是什么，以及认知如何使用表征，是两个领域边界上最深的开放问题之一。
过早在任一方向上解决具身性问题 关于具身性的两个强立场，在现有证据下都过度自信。强具身性主张（智能需要身体）受到无身体 LLM 广泛能力的挑战；强无身体主张（具身性无关紧要，LLM 已经证明这一点）则受到这些系统对其从未经验过的物理和社会世界仍有持续理解缺口的挑战。 保持具身性问题开放，并关注双方证据。无身体 LLM 的成功是真实证据，表明最强具身性主张过强；它们在 grounded understanding 上的缺口也是真实证据，表明具身性可能仍然重要。通向更通用智能的路径是否经过具身性，是关于领域方向的现实经验问题，而不是可以任意预设的已定结论。