English
Computer science, cognitive science, and neuroscience ask related questions at different levels. An AI system may be built to perform a task; cognitive science asks how people perceive, learn, remember, and act; neuroscience studies the biological mechanisms involved. A successful model can connect these questions, but success at one level does not automatically answer the others.
Marr’s distinction is a useful starting point: identify the task and why it matters, the representations and algorithms used, and the physical implementation. Two systems can compute the same result by different algorithms, and one algorithm can have different implementations. Comparing a brain with a program requires saying which of these relationships is being proposed.
Background: Classical AI (§5.1), learning (§5.2–§5.3), interpretability (§5.6), and robotics (§6.4). The goal is to understand what evidence would support a claim about cognition, rather than to infer a mind directly from a model’s name.
The Relationship: Computation as a Theory of Mind
Intelligence: The Computational Theory of Mind and Its Critics
The computational theory of mind proposes that at least some cognitive processes can be explained as computations over representations. Classical versions emphasized structured symbols and rules; other versions use distributed representations or dynamical processes. “Computational” therefore does not identify one particular architecture or settle how a representation acquires meaning.
The mid-twentieth-century exchange among logic, psychology, cybernetics, and computing made such explanations concrete. Newell and Simon built programs that searched through possible steps in problem solving and proposed the physical symbol system hypothesis. A program could now express a cognitive proposal precisely enough to produce predictions. Its behavior could be compared with human choices, errors, and solution paths.
Different claims need different tests. Building a system that solves a puzzle establishes a way to solve that task. Explaining human puzzle solving also requires evidence about human behavior, such as which errors occur or how performance changes with task structure. Several algorithms may fit the same accuracy score. A useful cognitive model predicts additional observations that distinguish it from alternatives.
Searle’s Chinese Room argument questions whether following symbol-manipulation rules is sufficient for understanding. Responses include the claim that understanding belongs to the organized system rather than the person executing a step. Dreyfus emphasized situated skills and context that early symbolic models handled poorly. These are arguments about what an explanation must account for; neither fluent output nor the bare fact that a system follows rules resolves the dispute by itself.
Representation: The Concept Both Fields Are Built On
A representation makes some distinctions available to a process. The phrases “the dog chased the cat” and “the cat chased the dog” contain the same words but assign different roles. A bag of word counts loses that distinction; a structured relation can preserve who chased whom. This makes representation a concrete design question: which information must remain available for the task?
Symbolic systems can encode entities, roles, and rules explicitly. Neural systems can learn distributed patterns in which many units jointly support a distinction. These are not mutually exclusive descriptions of every system: a neural model can operate over structured inputs, and a symbolic program can use learned components. Compare the operations a representation supports rather than deciding solely from its label.
A concept being decodable from a model’s state does not show that the model uses it for the task. A classifier might extract information that is correlated with the answer but unnecessary to the original computation. Interventions that change the representation and alter behavior provide another kind of evidence, although their interpretation depends on how the intervention affects the system (§5.6).
The same distinction matters when comparing models with brains. Similar patterns of responses to a set of images can suggest a shared organization, but may also reflect the shared task or stimulus set. Test held-out cases, alternatives, and predictions about errors. Stronger claims about mechanism require evidence beyond a high correlation between two collections of measurements.
Embodiment: The Challenge to Abstract Computation
Embodied approaches ask how a body and its environment help organize cognition and action. In robotics, changing the mechanics of a gripper changes the control problem. A compliant surface can accommodate variation in object shape that a rigid gripper must handle through more precise sensing and control. Some useful behavior comes from the interaction of mechanism, controller, and environment.
This does not automatically refute computational explanation. The body and environment can be included in the modeled system, with feedback determining the next observation and action. The issue is which boundary gives an adequate explanation. A model that assumes fixed inputs may miss how an agent actively changes the information available to it.
Claims about embodiment should be tied to tasks. A text-trained model receives records produced by people with bodily and social experience; a multimodal or robotic system has other inputs and opportunities for interaction. Success on one text task does not establish that direct sensorimotor experience is unnecessary for every task, while a physical body alone does not guarantee understanding. Compare what different learning conditions enable and where they fail.
Brains and Artificial Neural Networks: Correspondence and Its Limits
Artificial neural networks draw on biological inspiration without reproducing all neural mechanisms. McCulloch and Pitts’s 1943 logical neuron abstracted selected features into a threshold unit. Later models developed learning rules and architectures useful for engineering. An abstraction’s value depends on the question it helps answer, not on how literally it resembles a cell.
Biological neurons have temporal dynamics, varied cell types, and complex interactions within circuits. Artificial networks also vary: some are feedforward, some recurrent, and some explicitly model spikes or other dynamics. It is inaccurate to contrast “a recurrent brain” with “all artificial networks are feedforward.” Standard backpropagation specifies a mathematical learning procedure; whether and how biological circuits realize comparable credit assignment is a separate research question.
Computational neuroscience uses models to make predictions about neural recordings and behavior. Machine learning can supply candidate representations and analysis tools, while neuroscience suggests architectures, learning constraints, and experiments. A model that predicts some observations may still omit essential mechanisms for another task. State the level and scope of the proposed explanation before interpreting success as an account of the brain.
What This Perspective Changes
When an AI result is presented as evidence about minds, identify the task, representation, learning conditions, and measurements. Ask which alternative mechanisms would produce the same result and what observation could distinguish them.
Try explaining a familiar task, such as adding two numbers, at Marr’s three levels. Then imagine two algorithms and two physical implementations. The exercise makes clear why agreement on an answer, an algorithm, and a biological mechanism are different achievements.
Resources
Foundational and Bridging Texts
Boden’s Mind as Machine traces the interaction among AI, psychology, linguistics, and neuroscience. Its discussion of levels of explanation helps place Marr’s Vision in context. Fodor develops a symbolic account of thought; Searle and Dreyfus challenge different assumptions of computational accounts. Read each argument with its target and premises in view. Hofstadter connects formal systems and self-reference with questions about mind, but is not a substitute for empirical cognitive science.
| Resource | Role | Type |
|---|---|---|
| Boden, Mind as Machine (2 vols.) | History of AI and cognitive-science research traditions | Reference |
| Fodor, The Language of Thought | A symbolic theory of mental representation | Depth |
| Searle, “Minds, Brains, and Programs” (1980) | An argument about computation and understanding | Depth |
| Dreyfus, What Computers Still Can’t Do | Critiques involving situated and embodied activity | Depth |
| Hofstadter, Gödel, Escher, Bach | Formal systems, self-reference, and questions about mind | Depth |
| Marr, Vision | Levels of explanation in visual processing | Depth |
Representation, Connectionism, and the Brain
Parallel Distributed Processing and the Fodor–Pylyshyn paper present different accounts of representation and systematicity. Clark and Lakoff–Johnson examine the contributions of bodies and environments. For empirical work, use computational-neuroscience tutorials and inspect an actual dataset: identify the recorded quantity, task, model, and comparison. The Neuromatch materials, Cognitive Computational Neuroscience proceedings, and Allen Institute data serve different stages of that exercise.
| Resource | Role | Type |
|---|---|---|
| Rumelhart & McClelland, Parallel Distributed Processing | Models using distributed representations | Depth |
| Fodor & Pylyshyn, “Connectionism and Cognitive Architecture” (1988) | Representation and systematicity in cognitive models | Depth |
| Lakoff & Johnson, Philosophy in the Flesh | An embodied account of concepts and reasoning | Depth |
| Clark, Being There / Supersizing the Mind | Embodied and extended accounts of cognition | Depth |
| Computational neuroscience reviews (current) | Comparisons of model and biological observations | Reference |
| Neuromatch Academy Computational Neuroscience (free) | Computational-neuroscience exercises and tutorials | Practice |
| Cognitive Computational Neuroscience proceedings (free/current) | Research on computational accounts of cognition | Reference |
| Allen Brain Map / Allen Institute open data (free) | Neuroscience observations for computational analysis | Practice |
Traps
| Trap | Why it misleads | Better response |
|---|---|---|
| Equating task performance with a cognitive explanation | Different mechanisms can produce the same output. | Compare representations, learning conditions, errors, and additional predictions. |
| Treating neural inspiration as biological identity | A useful abstraction can differ substantially from a biological circuit. | State which level and observations the model explains. |
| Assuming decodable information must be causally used | Correlation can support a probe without driving the original task. | Compare alternative explanations and carefully designed interventions. |
| Using embodiment as an all-or-nothing slogan | The contribution of interaction depends on task and learning conditions. | Specify the body-environment feedback and test its role. |
中文
计算机科学、认知科学和神经科学在不同层次提出相关问题。AI 系统可能以完成任务为目标；认知科学研究人怎样感知、学习、记忆与行动；神经科学研究其中的生物机制。成功模型可以联系这些问题，但一个层次上的成功不会自动回答其他层次。
Marr 的区分提供了一个起点：明确任务及其意义、所用表示与算法，以及物理实现。两个系统可以用不同算法得到相同结果，同一算法也可以有不同实现。比较大脑与程序时，需要说明提出的是哪一种关系。
背景知识：经典 AI（§5.1）、学习（§5.2–§5.3）、可解释性（§5.6）与机器人学（§6.4）。目标是理解什么证据能够支持认知方面的结论，而不是直接从模型名称推断心智。
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
心智计算理论提出，至少某些认知过程可以解释为对表示进行的计算。经典版本强调结构化符号与规则；其他版本则使用分布式表示或动态过程。因此，“计算性的”并不指定某一种架构，也不能解决表示怎样获得意义的问题。
二十世纪中期，逻辑、心理学、控制论与计算之间的交流，使这类解释变得具体。Newell 与 Simon 构建程序，在问题求解的可能步骤中搜索，并提出物理符号系统假说。程序可以把认知设想表达得足够精确，从而产生预测；再将其行为与人的选择、错误和解题路径比较。
不同结论需要不同检验。构建一个能解谜的系统，确立了一种完成任务的方法。解释人怎样解谜，还需要人类行为证据，例如出现哪些错误，或任务结构变化后表现怎样改变。多个算法可能得到相同准确率。有用的认知模型应预测更多观察，以便与其他解释区分。
Searle 的中文房间论证质疑，遵循符号操作规则是否足以产生理解。回应之一认为，理解应归于有组织的整个系统，而不是执行某一步的人。Dreyfus 强调了早期符号模型难以处理的情境技能与背景。这些论证讨论解释必须说明什么；流畅输出本身，或系统遵循规则这一事实，都不能单独解决争论。
表征：两个领域都建立其上的概念
表示让某些区别能够被处理过程使用。“狗追猫”和“猫追狗”包含相同词语，却分配了不同角色。只统计词频的表示会丢失这个区别；结构化关系则可以保留谁追谁。由此，表示成为具体设计问题：任务需要哪些信息始终可用？
符号系统可以显式编码实体、角色与规则。神经系统可以学习分布式模式，由许多单元共同支持一种区别。这些描述并非在所有系统中互斥：神经模型可以处理结构化输入，符号程序也可以使用学习得到的组件。应比较表示支持哪些操作，而不只根据标签判断。
能够从模型状态中解码某个概念，不代表模型在任务中使用了它。一个分类器可能提取出与答案相关、却并非原计算所必需的信息。改变表示并观察行为变化的干预，提供另一类证据；但解释仍取决于干预怎样影响系统（§5.6）。
比较模型与大脑时，同样需要这个区别。面对一组图像，两者响应模式相似，可能提示共同的组织方式，也可能反映任务或刺激集合相同。应检查留出的案例、替代解释，以及关于错误的预测。关于机制的更强结论，需要超出两组测量高度相关的证据。
具身性：对抽象计算的挑战
具身取向考察身体与环境怎样参与组织认知和行动。在机器人学中，改变夹爪机械结构，就改变了控制问题。柔顺表面可以适应物体形状的变化；刚性夹爪可能需要更精确的感知与控制来处理同样变化。有用行为的一部分来自机构、控制器与环境之间的相互作用。
这不会自动否定计算解释。可以把身体与环境纳入模型，由反馈决定下一次观察与动作。问题在于选择什么系统边界，才能给出充分解释。把输入视为固定的模型，可能漏掉行动者怎样主动改变自己能够获得的信息。
关于具身性的结论应对应具体任务。以文本训练的模型接收了由具有身体和社会经验的人产生的记录；多模态或机器人系统则具有其他输入与交互机会。某个文本任务上的成功，不能证明所有任务都不需要直接感知行动经验；仅有身体也不能保证理解。应比较不同学习条件支持什么，又在哪里失败。
大脑与人工神经网络：对应及其限度
人工神经网络借鉴生物思想，但不重现全部神经机制。McCulloch 与 Pitts 在 1943 年提出的逻辑神经元，把部分特征抽象为阈值单元。后来的模型发展了适合工程任务的学习规则和架构。抽象是否有价值，取决于它帮助回答什么问题，而不是外观有多像细胞。
生物神经元具有时间动态、多种细胞类型，以及回路内部的复杂相互作用。人工网络也各不相同：有前馈网络、循环网络，也有显式模拟脉冲或其他动态的模型。把“循环的大脑”与“所有人工网络都是前馈的”对立起来，并不准确。标准反向传播规定了一种数学学习过程；生物回路是否及怎样实现类似的信用分配，是另一个研究问题。
计算神经科学用模型预测神经记录与行为。机器学习可以提供候选表示和分析工具，神经科学则可以提出架构、学习约束和实验。能预测某些观察的模型，仍可能遗漏另一个任务所需的重要机制。应先说明解释的层次与范围，再判断成功是否说明了大脑怎样工作。
这种视角会改变什么
当某项 AI 结果被当作关于心智的证据时，应明确任务、表示、学习条件和测量。问哪些替代机制也能产生同样结果，以及什么观察能够区分它们。
尝试在 Marr 的三个层次上解释一个熟悉任务，例如两数相加，再设想两种算法与两种物理实现。这个练习能说明：答案一致、算法一致和生物机制一致，是不同的成果。
资源
基础性与桥梁性文本
Boden 的《Mind as Machine》追踪 AI、心理学、语言学与神经科学之间的相互作用；对解释层次的讨论，有助于理解 Marr《Vision》的背景。Fodor 发展符号性的思维解释；Searle 与 Dreyfus 则质疑计算解释中的不同假设。阅读每项论证时，应辨认其对象与前提。Hofstadter 联系形式系统、自指与心智问题，但不能替代经验性的认知科学研究。
| 资源 | 作用 | 类型 |
|---|---|---|
| Boden, Mind as Machine（两卷） | AI 与认知科学研究传统的历史 | 参考 |
| Fodor, The Language of Thought | 心智表示的符号理论 | 深入 |
| Searle, “Minds, Brains, and Programs”（1980） | 关于计算与理解的论证 | 深入 |
| Dreyfus, What Computers Still Can’t Do | 涉及情境与具身活动的批评 | 深入 |
| Hofstadter, Gödel, Escher, Bach | 形式系统、自指与心智问题 | 深入 |
| Marr, Vision | 视觉处理中的解释层次 | 深入 |
表征、连接主义与大脑
《Parallel Distributed Processing》与 Fodor–Pylyshyn 论文，对表示和系统性提出不同解释。Clark 与 Lakoff–Johnson 考察身体和环境的作用。开展经验研究时，可以结合计算神经科学教程，检查一个实际数据集，明确记录量、任务、模型与比较方式。Neuromatch 材料、Cognitive Computational Neuroscience 会议论文，以及 Allen Institute 数据，分别支持练习的不同环节。
| 资源 | 作用 | 类型 |
|---|---|---|
| Rumelhart & McClelland, Parallel Distributed Processing | 使用分布式表示的模型 | 深入 |
| Fodor & Pylyshyn, “Connectionism and Cognitive Architecture”（1988） | 认知模型中的表示与系统性 | 深入 |
| Lakoff & Johnson, Philosophy in the Flesh | 概念与推理的具身解释 | 深入 |
| Clark, Being There / Supersizing the Mind | 具身与延展认知解释 | 深入 |
| Computational neuroscience reviews（当前） | 模型与生物观测的比较 | 参考 |
| Neuromatch Academy Computational Neuroscience（免费） | 计算神经科学练习与教程 | 实践 |
| Cognitive Computational Neuroscience proceedings（免费/当前） | 认知计算解释的研究 | 参考 |
| Allen Brain Map / Allen Institute open data（免费） | 可供计算分析的神经科学观测 | 实践 |
陷阱
| 误区 | 为什么会误导 | 更好的做法 |
|---|---|---|
| 把任务表现等同于认知解释 | 不同机制可能产生同样输出。 | 比较表示、学习条件、错误与额外预测。 |
| 把神经启发当成生物机制相同 | 有用抽象也可能与生物回路相差很大。 | 说明模型解释的层次与观察。 |
| 认为能解码的信息必定被因果性使用 | 相关性可能支持探针，却不驱动原任务。 | 比较替代解释与经过设计的干预。 |
| 把具身性当作非此即彼的口号 | 交互作用取决于任务与学习条件。 | 明确身体与环境的反馈，并检验其作用。 |