Mark I 感知机问世

1958 年夏天，心理学家 Frank Rosenblatt 造出了一台机器，名字很长，叫 Mark I 感知机。它由 400 个光电单元和一个能调整权重的输出单元组成，输入是一块简单的图案，输出是对图案的判断。真正让当时的人震惊的不是它的构造，而是一个动作：它会「学习」。

在那个晶体管才刚刚普及的年代，「学习」这个词属于人和动物。Rosenblatt 的机器却能通过一次次试错自动调整内部权重，学会区分简单的形状。媒体用「电子大脑」来形容它，军方对它寄予厚望，公众第一次直观地感受到：原来机器可以从经验里变聪明。这一幕，比今天任何一场 AI 发布会都更具象征意义。

Mark I 的轰动不是凭空而来。它站在 1943 年 McCulloch-Pitts 神经元模型、1957 年感知机理论论文的肩膀上，把纸上数学变成了看得见摸得着的实物。对研究者来说，它证明了「学习」可以被机器实现；对大众来说，它第一次让「会学习的机器」有了具体的形象。

但神话的背面很快显现。感知机只能处理线性可分的问题，连「异或」都学不会。1969 年，Minsky 与 Papert 在《感知机》一书里严格论证了这些局限，神经网络研究随之被打入冷宫，并成为第一轮 AI 寒冬的导火索之一。这台曾让世界兴奋的机器，最后成了连接主义路线十年低迷的标志。

回看 Mark I 感知机，它的历史意义不在于它能做什么，而在于它示范了什么。它让「机器能学习」从假设变成事实，点燃了一个领域，也因期望过高而迎来寒潮。今天当神经网络重新统治 AI 世界时，1958 年那台光电机器，正是这条长河最初的一滴水。

In the summer of 1958 psychologist Frank Rosenblatt built a machine with a long name: the Mark I perceptron. It had 400 photoelectric input units and an output unit with adjustable weights. In went a simple pattern, out came a judgment about it. What stunned people was not the construction but the behavior: it learned.

In an age when the transistor was barely spreading, the word "learning" belonged to people and animals. Rosenblatt's machine could adjust its internal weights through trial and error and learn to distinguish simple shapes. The press called it an "electronic brain"; the military pinned hopes on it; the public felt for the first time that machines could get smarter from experience. As symbolism, that moment outshines any modern AI launch.

The Mark I did not appear from nowhere. It stood on the 1943 McCulloch–Pitts neuron model and Rosenblatt's own 1957 theory paper, turning paper math into a tangible object. For researchers it proved learning could be implemented in a machine; for the public it gave "machines that learn" a concrete face for the first time.

The myth's reverse side appeared quickly. The perceptron could only handle linearly separable problems—it could not even learn XOR. In 1969 Minsky and Papert's book Perceptrons rigorously proved these limits, and neural-network research was pushed aside, helping trigger the first AI winter. The machine that had excited the world became the symbol of a decade-long retreat from connectionism.

Looking back, the Mark I's historical meaning lies not in what it could do but in what it demonstrated. It turned "machines can learn" from hypothesis into fact, ignited a field, and then drew down a cold front when expectations overshot. Today, as neural networks once again rule the AI world, that 1958 photoelectric machine is the first drop in the river.