Connectionism

Connectionism is an approach to the study of mental processes and cognition that describes the mind and behaviour as phenomena emerging from networks of simple, interconnected units. Its most characteristic models are artificial neural networks, whose capacity to process information depends on the strength of the connections between their units rather than on explicit rules programmed in advance.

The idea has roots in nineteenth-century psychology and neurology. Santiago Ramón y Cajal described the structure of neurons and the way they interconnect, and in 1949 Donald Hebb proposed that the connection between two neurons strengthens when they fire at the same time. Edward Thorndike, for his part, framed learning as the strengthening or weakening of links between stimuli and responses depending on their outcomes. The formal model took a decisive step in 1943, when Warren McCulloch and Walter Pitts showed how a network of idealised neurons could implement logical operations. In 1958 Frank Rosenblatt presented the perceptron, the first network able to learn from examples.

The history of connectionism is usually told as a succession of waves. The first lost momentum after Marvin Minsky and Seymour Papert's 1969 critique of the limits of the simple perceptron. The second gained strength in the 1980s with the book Parallel Distributed Processing, which introduced hidden layers and non-linear activation functions, and with the backpropagation algorithm for adjusting the weights of a network. The third, marked by deep learning, has made possible today's large language models.

Connectionists usually describe learning as a gradual adjustment of connection weights, guided by the error between the network's output and the expected one. Its main appeal is the link it offers between neural activity and cognition, although some critics regard that link as excessively reductionist. A central debate concerns whether connectionism is a genuine alternative to classical symbolic artificial intelligence, and to what extent it can explain high-level phenomena such as the compositionality of language or logical reasoning. The difficulty of interpreting what its models do internally remains one of its main challenges.