《Perceptrons》划出单层网络的边界

关于《Perceptrons》，最流行的版本像一桩旧案：1969 年，马文·明斯基和西摩·佩珀特证明神经网络没有前途，研究经费随之消失，连接主义被迫沉寂。这个故事有书名、有“凶手”、也有明确后果，适合在几句话里讲完。它的问题是，原结论里的限定词几乎全被删掉了。

这本书的副题是“计算几何导论”。作者研究的是一类阈值装置，尤其是单层感知机。输入经过加权求和，再由阈值作出决定；从几何上看，它画出的是一条直线或一个超平面。XOR 的正负样本交错排列，没有一条直线能把它们分开。因此，无论怎样调整权重，单层线性分类器都表达不了 XOR。这里说的是一个明确定义的模型类和一个明确定义的函数，不是“所有神经网络永远失败”。

书中的技术讨论也远不止那张后来反复出现的 XOR 图。明斯基和佩珀特分析不同谓词、连接范围与计算代价，追问局部连接的装置为何难以捕捉某些全局性质。这项工作把“实验没训练好”和“模型形式根本表示不了答案”分开。面对 1950 年代末围绕感知机产生的过度宣传，这种边界并非打压研究，而是要求承诺与证据重新对齐。

真正的误读发生在结论离开数学语境之后。“单层感知机”缩成“神经网络”，“无法表示某些函数”扩大为“整条路线无望”。多层网络可以组合出单层没有的决策边界；但在 1969 年，存在这种表达能力，不等于研究者已经拥有稳定、可复用的训练方法，也不等于算力和数据足以支持它。到 1980 年代中期，反向传播在多层连接主义模型中被系统展示，XOR 才重新显出它原本的含义：一道揭示单层局限的题，而不是禁止增加隐藏层的法令。

连接主义此后的低潮是真实的，这本书的批评也确实影响了研究气候。但资金制度、硬件能力、实验可复现性、学术潮流和早期承诺的透支，同样构成因果。把一整个时期压缩成“两位作者杀死神经网络”，既夸大了一本书控制机构的能力，也替当时尚未解决的训练与工程难题找了一个过分整齐的替罪者。两位作者后来还为扩充版加入关于并行计算和新研究方向的讨论；这不像一纸永久禁令，更像一场仍在继续的技术争论。

《Perceptrons》留下了两种值得保留的边界。第一种属于模型：一条线分不开什么，某种连接结构要为哪些全局性质付出高昂代价。第二种属于叙事：一项精确结论在失去“单层”“特定函数”“特定连接”等词之后，会被扩写成多大的历史判决。读回这本书，不必替它洗清所有影响，只需把那些被删掉的词重新放回句子，因果就会恢复应有的复杂度。

The popular account of Perceptrons reads like a cold case with unusually tidy evidence. In 1969 Marvin Minsky and Seymour Papert proved that neural networks had no future; funding disappeared; connectionism went quiet. The story has a title, two culprits, and a consequence, so it can be told in a few sentences. Nearly every qualifier in the original result has to be removed to make it work.

The book’s subtitle was An Introduction to Computational Geometry. Its subject was a class of threshold machines, with particular attention to single-layer perceptrons. Weighted inputs are combined and passed across a threshold; geometrically, the machine draws a line or hyperplane between classes. XOR arranges its positive and negative cases so that no single line separates them. No adjustment of the weights can make that linear boundary appear. This is a precise result about a defined model class and a defined function—not a theorem that every neural network must fail forever.

The technical argument extended far beyond the diagram of XOR that later became famous. Minsky and Papert examined predicates, the reach of connections, and the computational cost of capturing global properties with locally wired devices. Their analysis separated two very different failures: a system that happens not to train successfully and a system whose form cannot represent the answer at all. After the extravagant claims surrounding perceptrons in the late 1950s, such boundary work did not make research less scientific. It forced promises and evidence back into contact.

The distortion began as the result traveled. Single-layer perceptron contracted to neural network; cannot represent certain functions expanded to the whole program is hopeless. Multilayer networks can construct boundaries unavailable to one layer. But the existence of that expressive capacity in 1969 did not mean researchers already possessed a stable, reusable way to train it, or the compute and data to make it practical. When backpropagation was clearly demonstrated for multilayer connectionist models in the mid-1980s, XOR recovered its narrower meaning: a test that exposes the limits of one layer, not an order forbidding hidden ones.

Connectionist research did lose momentum, and the book’s criticism affected the climate. Yet funding institutions, hardware limits, reproducibility, academic fashion, and exhausted promises belong to the same causal account. Saying that two authors “killed neural networks” grants one volume implausible power over institutions while turning unsolved technical problems into an overly neat victim story. The later expanded edition, with new discussion of parallel computation and emerging directions, also looks less like a permanent ban than an argument continuing as the field changed.

Perceptrons leaves two boundaries worth preserving. One belongs to models: what a line cannot separate, and what a particular connection scheme must pay to capture a global property. The other belongs to historical storytelling: how far a rigorous result can expand once words such as single-layer, particular function, and particular connectivity disappear. The book need not be absolved of influence. Its qualifiers simply need to be restored; once they are, the history regains the complexity that the indictment removed.