JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
Back to the library
人工智慧 1943

神經活動中內在思想的邏輯演算

沃倫·麥卡洛克 與 沃爾特·皮茨

把神經元看作「全或無」,由它們組成的網路,便成了一台演算邏輯的機器。

Choose your version
In depth · the introduction

1943 年,一位腦科學家與一個自學成才的離家少年邏輯學家,寫下了關於單個神經細胞的第一個方程——又幾乎是無心地,畫出了人工智慧的第一張藍圖。

把這個想法拆開看

麥卡洛克與皮茨從一個事實出發:神經細胞是「全或無」的——它要麼激發,要麼不激發,像一個開關,而非一個可調的旋鈕。於是他們提出一個大膽的問題——如果單個神經元只是一個開關,那麼由開關組成的整張網路,又能做什麼?

他們的答案是:只要接得對,這樣的網路就能演算邏輯。給一個神經元一條規則——「只有至少兩個輸入為開時才激發」——它的行為便是邏輯「且」。換一條規則,你便得到「或」,或「非」。而由於全部邏輯、乃至全部計算,都能由這寥寥幾樣拼出,一張由這些微小的「是 / 否」決定組成的網路,原則上便能計算任何電腦能算的東西。照此看來,大腦,是一台由開關砌成的龐大機器。

它從哪裡來

這對搭檔是一組不太可能的組合。沃倫·麥卡洛克是一位年屆四旬、富有哲思的神經生理學家,被一個問題縈繞:大腦那濕漉漉的物質,究竟何以能承載邏輯?沃爾特·皮茨則剛剛走出少年時代——一個離家出走的孩子,自學了形式邏輯,少時便寫信給伯特蘭·羅素,又師從邏輯學家魯道夫·卡爾納普,後被麥卡洛克收留於家中。

他們把神經生理學,與懷特海和羅素的形式邏輯、以及阿蘭·圖靈那個嶄新的「通用計算機器」之念融在一起,寫出一篇符號密佈、幾乎無法卒讀的論文。然而,對的人讀到了它:約翰·馮·諾伊曼——兩年後,他把他們的「邏輯神經元」,建進了現代儲存程式計算機的設計之中。

它為何重要

這是頭一回,有人為一個古老的問題給出了精確的、數學的回答:機器,原則上,能思考嗎?透過表明大腦自身的硬體可以被描述為計算,這篇論文一舉開啟了兩個領域。它告訴工程師:由簡單開關組成的網路能計算任何東西——這一許諾,一直延伸到今天的神經網路。它也告訴腦科學家:心智,或許可以被理解為正在被處理的資訊。此後的每一個人工神經網路,以及你此刻或許正在使用的 AI,都從這一想法承襲而來。

一個便於想像的畫面

想像一個酒吧門口的保安,在決定今晚要不要照常營業。只有到場的、獲准的客人足夠多,他才放行——這就是一個閾值。可名單上有一個被禁的名字,只要那人一進門,今晚就取消,無論場子有多滿。這條「客人夠多,除非那個被禁的人出現」的規則,正是麥卡洛克—皮茨神經元:把興奮性輸入與一個閾值相比,但讓任何抑制都能一筆勾銷一切。在下方,親手試試。

兩個輸入接入一個閾值神經元。每個突觸或為興奮性(+),或為抑制性(−)。當啟用的興奮性輸入之數達到閾值 θ、且沒有抑制性輸入處於啟用時,神經元便激發。預設按鈕分別給出 OR(θ=1)、AND(θ=2)與 NOT(單個抑制性輸入,θ=0);一張即時真值表,給出每一種輸入組合下的輸出。

它的位置

這篇論文是思想史上的一道樞紐。在它身後,站著喬治·布林——1854 年,他把邏輯變成了代數;還有阿蘭·圖靈——1930 年代,他把計算變成了機器能執行的事。在它身前,則是整個現代 AI:法蘭克·羅森布拉特的感知機(1958)賦予神經元學習的能力;數十年後,AlexNet(2012)與 Transformer(2017)把數以百萬計的這種單元堆疊起來,造出我們如今逕直稱作 AI 的系統。麥卡洛克與皮茨,正是這條漫長譜系的起點。

The original document
Original source text
Warren S. McCulloch & Walter Pitts · Bulletin of Mathematical Biophysics, Vol. 5, pp. 115–133 · 1943
Abstract
Because of the "all-or-none" character of nervous activity, neural events and the relations among them can be treated by means of propositional logic.
It is found that the behavior of every net can be described in these terms, with the addition of more complicated logical means for nets containing circles; and that for any logical expression satisfying certain conditions, one can find a net behaving in the fashion it describes.
The five physical assumptions
From the known physiology of the neuron the authors abstract five idealizing assumptions. They are stated plainly, and everything in the paper is built on them:
1. The activity of the neuron is an "all-or-none" process.
2. A certain fixed number of synapses must be excited within the period of latent addition in order to excite a neuron at any time, and this number is independent of previous activity and position on the neuron.
3. The only significant delay within the nervous system is synaptic delay.
4. The activity of any inhibitory synapse absolutely prevents excitation of the neuron at that time.
5. The structure of the net does not change with time.
From neurons to propositions
On these assumptions each neuron becomes a two-valued device, firing or not firing at each tick of a discrete clock set by the synaptic delay. Its firing at one instant is read as a proposition asserting that the conditions for it held an instant before. The paper then shows that nets without closed loops realize exactly the expressions of propositional logic — conjunction, disjunction, negation and their combinations — while nets that contain loops ("circles") acquire memory, letting a fact reverberate and so depend on the whole past.
[ … ]
The reach of the result
Building on Turing's 1936 notion of computability, the authors argue that, apart from the provision of an unlimited tape, such nets can compute whatever is effectively computable — that the logic of the idealized brain and the logic of the universal machine are, in this precise sense, the same. The closing sections turn to consequences for psychology and the theory of knowledge. The notation, borrowed from Carnap and Russell, is famously forbidding; the full nineteen pages, with every theorem and proof, are at the source below.
Chicago · 1943