A Logical Calculus of the Ideas Immanent in Nervous Activity

In December 1943 Warren S. McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity" in the Bulletin of Mathematical Biophysics, volume 5, number 4, pages 115 to 133. The authors were at the Department of Psychiatry at the Illinois Neuropsychiatric Institute, University of Illinois College of Medicine, and at the University of Chicago.^[1] The paper was reprinted in 1990 in the Bulletin of Mathematical Biology.^[2]

The problem

The paper asks how the activity of nerve cells relates to logic and thought. Its starting point is a basic property of neurons: nervous activity is "all-or-none," meaning a nerve cell either fires or does not. The authors argue that, because of this, neural events and the relations among them can be treated with propositional logic, the logic of statements that are true or false and are combined with words such as "and," "or" and "not."^[1]

The model

The authors describe an idealised neuron with five assumptions. First, its activity is all-or-none. Second, a fixed number of its synapses must be excited within a short period to make it fire. Third, the only significant delay is the synaptic delay. Fourth, any active inhibitory synapse absolutely prevents firing. Fifth, the structure of the net does not change over time.^[1]

A small example shows how this works. Imagine a unit that fires only when two excitatory inputs arrive together, and that stays silent whenever an inhibitory input is active. It behaves like a logical "and" with a veto attached. Wiring many such units together lets a net express more complicated logical expressions. This example is an illustration added here, not a case taken from the paper.

Results

The main result is a correspondence between nets and logic. The behaviour of every such net can be described in logical terms. Nets containing loops, which the paper calls "circles," need more complex logic. The reverse also holds: for any logical expression that meets certain conditions, a net can be built that behaves as the expression describes.^[1]

The paper also links its model to Turing's. A net equipped with a tape, scanners and suitable outputs can compute exactly the numbers that a Turing machine can. The authors present this as a psychological justification of Turing's definition of computability.^[1]

The model does not directly explain learning. The authors argue that nets altered by facilitation or learning can be replaced by equivalent "fictitious" nets whose connections and thresholds never change. They stress that this is a formal equivalence, not a factual explanation. Theorem 7 states that alterable synapses can be replaced by circles.^[1]

Why it mattered

The paper showed that a simple, idealised neuron is enough to build logic, and it tied models of the brain to computation in a precise form. Its strength lies in making a biological picture exact enough to prove theorems about it. Its limits are equally clear. Because the structure of the net is fixed, the model does not account for learning, and the authors do not claim to explain how learning happens.^[1]

What followed

Later research on artificial neural networks grew out of questions like the ones this paper raised: how far can networks of simple units go, and what can they represent? The paper remains a common starting point for anyone who wants to see how neural networks and logic are related.