The word “algorithm” derives from the name of the Muslim mathematician and astronomer Muhammad ibn Musa al-Khwarizmi (c.780–c.850), author of one of the most important works of medieval mathematics, translated into Latin in the 12th century under the title “Algoritmi de Numero Indorum” (“Al-Khwarizmi concerning Hindu numbers”). Initially, it referred to the study of the decimal numbering system created by the Hindus, but its meaning has evolved a lot over the centuries, as I have already described.
In the 20th century, the term was appropriated by electronic computing: algorithm came to mean a finite sequence of explicit operations to solve a problem or make a calculation automatically. The (human) programmer wrote the sequence in a language suitable for a computer, and the machine performed the operations exactly as prescribed.
But that paradigm was broken by the advent of machine learning methods. The vast majority of algorithms today are much shorter than the codes we wrote twenty or thirty years ago, and initially they are incapable of performing the tasks they set out to do. What makes them so (extraordinarily) useful is their ability to learn how to do it!
The crucial element in many of these modern algorithms is the concept of a neural network. It goes back to the idea of the “disorganized type B machine” described by the English mathematician Alan Turing (1912–1954), a pioneer of computing, in his work “Intelligent Machines”, which he wrote in 1948 but was not published until after his death.
It is a logical model inspired by the structure of our neurons. Like a biological neuron, a neural network takes several inputs (“stimuli”) and produces a response (“trigger”), which depends both on the stimuli and certain parameters that are part of the definition of the neural network.
Initially, the algorithm is “educated” using known data: the parameters of neural networks are calibrated so that they reproduce known correct responses. From there, it is ready to be used to make decisions in new situations.
This procedure can be performed quite automatically, with little human intervention. And it’s very effective, frighteningly effective…
.
