Textbook of AI

The McCulloch–Pitts neuron, introduced in Warren McCulloch and Walter Pitts's 1943 paper A Logical Calculus of the Ideas Immanent in Nervous Activity, is the first mathematical model of a neuron and the foundational unit of artificial neural networks. It abstracts the neuron to its bare logical essence: an all-or-nothing thresholding device that computes a Boolean function of its inputs.

The model

Each unit receives a finite number of binary inputs $x_1, \ldots, x_n \in \{0,1\}$, computes their weighted sum, and fires (outputs $1$) if and only if the sum exceeds a fixed threshold $\theta$:

$$y = \begin{cases} 1 & \text{if } \sum_i w_i x_i \geq \theta \\ 0 & \text{otherwise}. \end{cases}$$

Excitatory synapses contribute $w_i = +1$; inhibitory synapses are absolute, a single active inhibitory input forces $y = 0$ regardless of excitatory activity (modelled in some presentations as $w_i = -\infty$). Time advances in synchronous discrete steps.

What McCulloch–Pitts networks can compute

McCulloch and Pitts proved that:

• Any propositional logic expression is computable by a network of such units. AND requires two inputs and threshold $2$; OR requires threshold $1$; NOT requires a single inhibitory input. Because $\{\text{AND}, \text{OR}, \text{NOT}\}$ is functionally complete, all Boolean functions follow.
• Recurrent networks of McCulloch–Pitts neurons can implement arbitrary finite-state machines. Loops in the connectivity graph give the network memory; the state is encoded in which subset of neurons fired in the previous time step.

Together, these results constitute the first formal demonstration that neural tissue, abstracted to its logical structure, is computationally universal, every Turing-computable function on bounded-length inputs is realisable by a McCulloch–Pitts network. The paper appeared in the Bulletin of Mathematical Biophysics and was an explicit attempt to bring mathematical logic (the authors leaned heavily on Carnap and Whitehead/Russell) into neuroscience.

Biological adequacy

The model is biologically minimal. Real neurons have:

• Continuous firing rates and graded postsynaptic potentials, not binary outputs.
• Complex temporal dynamics: refractory periods, spike-timing-dependent plasticity, short-term facilitation and depression.
• Dendritic computation, the dendritic tree itself performs nontrivial nonlinear integration.
• Stochastic synaptic transmission and neuromodulation.

McCulloch and Pitts knew this; their goal was a minimum sufficient model for logical computation, not a faithful biophysical one.

Legacy

The McCulloch–Pitts neuron has no learning rule, weights are fixed by the modeller. Adding a learning rule to determine real-valued weights from data was the contribution of Donald Hebb (1949) and, more decisively, of Frank Rosenblatt's perceptron (1958). Every artificial neuron with a hard threshold (the step function) is a McCulloch–Pitts neuron; every neuron with a sigmoid, tanh, ReLU or GELU activation is a continuous generalisation of one. The architectures of LeNet, AlexNet, ResNets, transformers and modern frontier models all stack such units in their billions.

The 1943 paper also seeded the Macy Conferences, cybernetics, and the cognitive-science programme that views the mind in terms of information processing, making McCulloch and Pitts founders not just of neural networks but of much of computational cognitive science.