McCulloch–Pitts 人工神经元

1943 年，芝加哥。一位神经生理学家沃伦·麦卡洛克，和一位几乎自学成才的少年逻辑学家沃尔特·皮茨，合写了一篇题为《神经活动中内在思想的逻辑演算》的论文。两人年龄相差悬殊，性格也截然不同，却在同一件事上达成共识：神经元的运作，可以用数学精确地描述出来。

他们的模型朴素而优雅。每个神经元是一个带阈值的开关：把来自其他神经元的信号加权求和，总和超过某个阈值，神经元就「点火」输出 1，否则保持沉默输出 0。听上去像是一台简化的计算器，但麦卡洛克和皮茨证明了一件事——把足够多的这种单元连成网络，就能模拟与、或、非这类逻辑运算，甚至等同于一台有限的自动机。

这个结论在当时是颠覆性的。图灵提出可计算性概念不过几年，而麦卡洛克和皮茨已经把「神经活动」和「逻辑演算」画上了等号。他们等于在说：大脑里那些会放电的细胞，本质上就是在做数学；而既然数学可以被机器执行，那么机器就有理由拥有「思维」。这正是后来 AI 一切故事的总起点。

当然，这篇论文并没有立刻点燃一场革命。它在当时引发的关注有限，真正的继承者要等十多年后罗森布拉特的感知机。但作为源头，它的地位无可替代——几乎所有人工神经网络的模型，都可以追溯到这篇论文里那个带阈值的开关单元。

回望这段历史，很难不感慨这组搭档的奇特。麦卡洛克是终身学者，皮茨却是天才而坎坷的少年——他早年流离失所，晚年远离学术圈，过早离世。但两人共同写下的那几页纸，为整个人工智能时代铺下了第一块基石。今天当我们讨论神经网络、深度学习时，谈论的正是 1943 年那个简单的数学模型，在漫长岁月里长成的模样。

Chicago, 1943. A neurophysiologist named Warren McCulloch and a teenage logician, largely self-taught, named Walter Pitts co-wrote a paper titled "A Logical Calculus of the Ideas Immanent in Nervous Activity." They were far apart in age and temperament but shared one conviction: the operation of neurons could be described precisely in mathematics.

Their model was plain and elegant. Each neuron is a switch with a threshold: it sums weighted signals from other neurons, and if the total passes the threshold it fires a 1, otherwise it stays silent at 0. It sounds like a simplified calculator, but McCulloch and Pitts proved something more: wire enough of these units into a network and you can simulate logical operations like AND, OR, and NOT—even act as a finite-state machine.

The conclusion was radical for its time. Turing had defined computability only a few years earlier, and McCulloch and Pitts were now equating nervous activity with logical calculus. They were effectively saying: the firing cells of the brain are doing mathematics, and since mathematics can be executed by machines, machines have grounds for "thought." That is the founding premise of the whole AI story.

Of course, the paper did not ignite an immediate revolution. It attracted limited attention, and its true heir—Rosenblatt's perceptron—came more than a decade later. But as an origin it is irreplaceable: nearly every model of artificial neural networks traces back to that threshold switch in this paper.

Looking back, the oddity of the pair is hard to ignore. McCulloch was a lifelong academic; Pitts was a brilliant and troubled youth who had drifted, later withdrew from the field, and died early. Yet the pages they wrote together laid the first stone for the entire artificial intelligence era. When we discuss neural networks and deep learning today, we are talking about the grown-up version of that simple 1943 mathematical model.