McCulloch and Pitts (1943): The First Artificial Neuron, Explained

The paper at a glance

In 1943, Warren McCulloch (a neurophysiologist) and Walter Pitts (a logician, barely 18) published "A Logical Calculus of the Ideas Immanent in Nervous Activity". Their claim was simple and enormous: a neuron can be described as a logical gate.

That single idea became the ancestor of every neural network — from the Perceptron in 1958 to the Transformers behind today's chatbots.

The problem: a neuron either fires or it doesn't

In the 1940s, neuroscience had settled on the all-or-none law: a neuron either fires an action potential or stays silent. There is no half-firing.

Logic, meanwhile, had Boolean algebra — AND, OR, NOT — the mathematics of true and false.

McCulloch and Pitts noticed the bridge: if a neuron fires only when enough inputs arrive, it is computing a threshold function. And threshold functions can express Boolean logic.

Feynman explanation: the night watchman

Imagine a night watchman at a gate. He opens the gate only if at least two of three guards report in.

One guard shows up → gate stays shut.
Two guards show up → gate opens.
A stop-order from the supervisor → gate stays shut, no matter how many guards arrive.

That is the McCulloch-Pitts neuron:

Inputs are the guards (0 or 1).
Weights say how much each guard counts.
Threshold is the minimum number of reports needed.
Inhibition is the supervisor's stop-order.

The math

The neuron sums its weighted inputs and compares the result to a threshold:

y = ⎩ ⎨ ⎧ 10 if i \sum w_{i} x_{i} \geq θ otherwise

The original 1943 paper added a biological twist: any inhibitory input vetoes the output, regardless of excitation:

y = 1 ⟺ exc \sum x_{i} \geq θ \land inh \sum x_{i} = 0

AND, OR, NOT from one neuron

Change the weights and the threshold, and the same neuron becomes a different logic gate.

AND gate

Two inputs, both must be 1. Set $w_{1} = w_{2} = 1$ and $θ = 2$ .

$x_{1}$	$x_{2}$	$\sum w_{i} x_{i}$	Output
0	0	0	0
0	1	1	0
1	0	1	0
1	1	2	1

OR gate

At least one input must be 1. Set $w_{1} = w_{2} = 1$ and $θ = 1$ .

$x_{1}$	$x_{2}$	$\sum w_{i} x_{i}$	Output
0	0	0	0
0	1	1	1
1	0	1	1
1	1	2	1

NOT gate

Use an inhibitory connection: an inhibitory input forces the output to 0.

$x_{1}$ (excitatory)	$x_{2}$ (inhibitory)	Output
0	0	0
1	0	1
0	1	0
1	1	0

Because these gates can be wired together, networks of McCulloch-Pitts neurons can compute any Boolean function. That is their universality.

What the paper actually claimed

The paper was careful:

Nervous activity can be described by a logical calculus.
Networks of threshold neurons can realize any finite logical expression.
Such networks could, in principle, perform perception, memory, and learning.

But it did not claim that the neurons learn. The weights were fixed by design.

What it got wrong

No learning. Weights and thresholds had to be set by hand. This lasted until 1958, when Rosenblatt's Perceptron introduced a learning rule.
Biological oversimplification. Real neurons have timing, refractory periods, and plastic synapses; the model has none of that.
Linearly separable only. A single threshold gate cannot compute XOR. Minsky and Papert highlighted this in 1969, helping trigger the first AI winter.

Why it still matters

Every modern network descends from this model:

1943 concept	Modern descendant
Threshold activation	ReLU, sigmoid, tanh, GELU
Weighted summation	Linear layers (matrix multiply)
Binary inputs	Continuous embeddings
Fixed weights	Learned by backpropagation
Single gate	Billions of parameters in layers

The structure — inputs, weights, sum, nonlinearity — is still the skeleton of every neural network.

One line

McCulloch and Pitts did not build a thinking machine; they built the logic of a single neuron — and every network since has been a conversation with that idea.

What is next

The next paper asks the question that launched computer science itself: can machines think?

Next up: Turing (1950) — Computing Machinery and Intelligence. (coming soon)