Rosenblatt 提出感知机

康奈尔航空实验室里，一块 20×20 的光电管阵列正对着输入。

四百个像素信号经过可调权重——最初以电位器等形式实现——送入分类单元：对了留下，错了就拧。这不是后来服务器机房里的抽象张量；光照、阵列分辨率与电位器调节都是实验参数。史密森尼美国国家历史博物馆仍收藏相关装置记录。机器按一条极短的规则运转：每个输入乘一个权重，加权求和，与阈值比较——越过阈值归一类，否则归另一类。答案错了，就按样本指示的方向改权重：猜成正类而实际是负类时，减小对应输入的权重；反方向错误则增大。下一次面对相似输入，分界线移动一点。

弗兰克·罗森布拉特把这条回路写进 1958 年《心理学评论》的论文 The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain。标题谈的是大脑，核心装置却可以写得非常短：错误修正权重。他的兴趣越过分类技巧，延伸到物理世界的信息如何被感知、以何种形式留在记忆中、记忆又如何影响识别和行为。论文试图在生物物理与心理学之间架桥。后来进入机器学习教材的，主要是桥上最便于计算的那一段。

这条规则的意义在于，分类标准不必全部由程序员预先写成“如果—那么”。样本本身开始塑造参数。对线性可分的数据——只要确有一条直线或超平面能把两类分开——更新过程可以证明会找到解。这就是感知机收敛定理。它给出的不是“总会学好任何任务”，而是“在模型能表达答案的前提下，错误驱动的更新有终点”。数学规则、训练样本和 Mark I 硬件在这里接上了同一条回路。

公共报道一度走得比实验更远。纽约时报等媒体在 1950 年代末对“能学习的机器”的报道，把实验室装置写成更接近科幻的承诺：语言、决策乃至更广智能。这些承诺超出了单层阈值单元实际给出的证据。模型学到的仍是一道线性边界。若问题需要把平面上交错的区域分开——后来反复出现的 XOR 就是最容易转述的例子——它就无能为力。训练再久、样本再多也不会出现那条边界，因为模型能够表达的形状里根本没有它。收敛定理与表达限制来自同一份简洁：它为什么可靠，也说明它为什么不够。

学习规则本身也值得拆开看。权重更新只在预测错误时发生；若当前权重已能正确分类某个样本，该样本不再推动参数。对线性可分问题，可以证明错误次数有上界；对不可分问题，算法可能在边界两侧来回摆动，工程上因此需要停止条件或学习率安排。因此，感知机不宜只被放在“早期成功”或“后来失败”中的一边。它第一次把学习写成一个能执行、能证明、也能明确反驳的过程。

1969 年明斯基与佩珀特的《Perceptrons》把几何边界写进系统数学；1980 年代中期前后的反向传播与多层网络，则试图在保留“误差更新参数”这一判断方式的同时，扩大模型能表达的几何。多层网络后来越过了其中一些限制，却保留了这项早期工作的提问顺序：先说明模型能表示什么，再说明数据如何改变它。康奈尔后来的纪念文章有时把罗森布拉特的工作称为“早了六十年”——夸张可以留给标题。技术账本里更稳妥的说法是：他留下了一条可检验的学习回路，以及一份同样可检验的边界清单。没有这份清单，后来的多层网络也不知道自己越过的是哪条线。

At Cornell Aeronautical Laboratory, a 20-by-20 array of photocells faced the input.

Four hundred pixel signals traveled through adjustable weights—initially realized with devices such as potentiometers—into classification units. Correct answers stayed put; wrong ones got a twist. This was not an abstract tensor in a later server room. Illumination, array resolution, and potentiometer adjustment were laboratory parameters. The Smithsonian National Museum of American History still preserves related hardware records. The machine ran a short rule: multiply each input by a weight, sum, compare with a threshold. One side of the threshold is one class; the other side is another. When the answer is wrong, alter the weights in the direction the example suggests—decrease weights on active inputs after a false positive, increase them after a false negative. The next similar input meets a slightly shifted boundary.

Frank Rosenblatt wrote that loop into a 1958 Psychological Review paper, The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain. The title spoke of the brain; the core device fits in a few lines of algebra: error changes weight. His ambition reached past classification into perception, memory, and how information from a physical world might alter later behavior. What machine-learning textbooks retained most readily was the tractable middle of that larger bridge.

That update replaced a familiar kind of programming. Instead of enumerating every visible condition under which an image belonged to one category, a developer supplied examples, and those examples moved the decision boundary. If two classes are linearly separable—if some line, plane, or higher-dimensional hyperplane really can divide them—the learning process can be proved to converge on a solution. The theorem does not promise success on arbitrary tasks; it promises a finish line when the model can represent the answer at all. A mathematical model, labeled experience, and the Mark I hardware now occupied one loop. The hardware made the learning rule something that could fail under real sensors and real noise, not only under abstract vectors on a page.

Public reporting sometimes traveled farther than the experiments. Press coverage in the late 1950s of a “machine that can learn” often moved laboratory hardware toward science-fiction promises—language, decision-making, broader intelligence—that the single-layer threshold unit had not earned. The model still learned only a linear boundary. XOR arranges its classes so that no one line can separate them. More training cannot repair this failure. More examples cannot discover a boundary that the model has no means to represent. The same simplicity that permits a convergence theorem also places some answers outside reach.

The learning rule itself rewards close inspection. Weights update only when the prediction is wrong; a correctly classified example leaves parameters alone. For linearly separable problems, the number of mistakes can be proved bounded. For inseparable problems, the algorithm may oscillate across the boundary, so engineering practice needs a stopping rule or a learning-rate schedule. It is therefore misleading to file the perceptron under either triumph or defeat. It made “learning” into a process that could run, be proved, and be disproved with unusual precision.

Minsky and Papert’s 1969 Perceptrons wrote the geometric limits into systematic mathematics. Multilayer networks and backpropagation, clarified for connectionist models in the mid-1980s, kept the error-driven update while enlarging what could be represented. Multilayer nets would later enlarge that geometry; the perceptron’s compact lesson survived them. Every claim that a system learns contains two separate questions: how experience alters it, and what kinds of answer the system was capable of expressing in the first place. Cornell’s later commemorations sometimes call Rosenblatt’s work “sixty years too soon.” For the technical ledger, a quieter formulation is enough: he left an executable learning loop and an equally executable list of boundaries. Without that list, later multilayer networks would not know which line they had crossed.