Artificial Neural Networks
In the middle of the 20th century, research in artificial intelligence experienced a remarkable breakthrough with the work of Warren McCulloch and Walter Pitts. These two researchers, who seemed to have nothing in common, combined their talents to design a mathematical model of the neuron. Their proposal advanced our understanding of the brain and laid the foundation for future machine learning technologies.
In 1943, McCulloch, a seasoned neuropsychiatrist, and Pitts, a young mathematical prodigy, published a seminal article. This groundbreaking text A Logical Calculus of the Ideas Immanent in Nervous Activity established a correspondence between the functioning of biological neurons and logical operations. The authors relied on the all-or-nothing character of neuronal activity to model nerve cells as binary devices, either active or inactive. This simplification enabled them to construct an equivalence with Boolean logic and its true/false values.
The backgrounds of these two researchers deserve attention. McCulloch, trained in psychiatry, had long harbored philosophical questions about the logical nature of thought and its physiological substrate. During the 1930s, he devoted himself to research in neurophysiology and collaborated notably with J. G. Dusser de Barenne at Yale, where they studied the localization of brain functions. Pitts, for his part, stood out for his precocious genius in mathematics. Without formal academic training, he developed a passion for mathematical logic from a very young age, devouring Bertrand Russell’s Principia Mathematica. In the early 1940s, he joined Nicolas Rashevsky’s team at the University of Chicago, a pioneering group in applying mathematics to biology.
The meeting between McCulloch and Pitts occurred in 1942, orchestrated by Jerome Lettvin. Despite their significant age difference, they quickly discovered their common interests. Pitts was enthusiastic about the concept of a logical machine capable of executing reasoning, an idea he connected to the work of Leibniz and Turing. McCulloch, for his part, had been attempting since the 1920s to translate neuronal activity into logical calculus. Together, they wondered whether the nervous system might function as such a logical machine.
Their 1943 article boldly combines elements of neurophysiology, philosophy, and mathematics. They present a simplified model of the neuron as a binary entity that activates or not according to the sum of its excitatory and inhibitory inputs, emitting a signal when this sum exceeds a certain threshold. By connecting these idealized neurons, they demonstrate the construction of networks using fundamental logical functions (AND, OR, NOT). Their approach follows an axiomatic method: from simplifying assumptions about neuronal activity, they build a logical calculus of relationships between neurons.
The authors themselves emphasize the theoretical limitations of their model. They do not seek to faithfully represent biological neurons, but rather to establish a formal framework for reasoning about hypothetical networks with known properties. They acknowledge that the activity of real neurons proves more continuous than discrete and that phenomena such as learning permanently modify the structure of networks. Their ambition lies elsewhere: to provide a mathematical tool for rigorously manipulating known networks and easily creating networks with desired properties.
This theoretical approach to biology, of which Nicolas Rashevsky was an ardent promoter, marks McCulloch and Pitts’s article. A physicist by training, Rashevsky founded a program of mathematical biophysics at Chicago in the 1930s. According to him, mathematical biology should proceed through abstractions and idealizations, similar to physics. Just as the physicist studies ideal concepts such as the material point or perfect fluid, the theoretical biologist must start from simplified models to progressively grasp the complexity of living systems. Mathematics, according to Rashevsky, offers a valuable framework for formalizing biological phenomena and deducing properties through rigorous reasoning.
It is in this spirit that McCulloch and Pitts shaped their model of logical neural networks. Their project aims not so much to describe the real nervous system as to show how a network of simple elements, properly connected, develops logical functions and, potentially, complex reasoning. Their work also connects to contemporary reflections on the logical foundations of computation, illustrated by the Turing machine. In 1936, Turing had shown how to define the process of computation by an abstract machine manipulating symbols according to precise rules. McCulloch and Pitts drew inspiration from this idea to conceive the brain as a logical machine, whose elementary components would be all-or-nothing neurons.
The impact of this article exceeded all expectations. On the theoretical level, it laid the foundations of computational neuroscience and cognitive science: the idea that complex cognitive functions arise from networks of simple units inspired numerous works in artificial intelligence and neuroscience. The article also presents a philosophical dimension by proposing a mechanistic vision of the mind, where thought would reduce to logical operations executed by neural networks. This conception had a lasting impact on debates about the nature of intelligence and its relationship to biological substrate.
The proposed model certainly has important limitations, which the authors acknowledge. It idealizes the neuron as a binary entity with a fixed threshold, whereas biological neurons possess continuous and adaptive dynamics. It assumes a fixed network structure, when neural connections constantly reconfigure through plasticity. Nevertheless, through its formal clarity, the article opened a vast field of research. It showed how a mathematical approach illuminates the functioning of the nervous system and, more broadly, the biological bases of intelligence.
This work testifies to the richness of interdisciplinary exchanges. It emerges at the crossroads of philosophical questions about the mind, experimental advances in neurophysiology, and formal developments in mathematical logic. It is through this intersection that McCulloch and Pitts arrive at a new vision, where the brain appears as a system for logical information processing. While their model remains rudimentary, it sketches the conceptual outlines of future research on formal neural networks and their applications.