Neural networks are computational models inspired by the structure and function of biological nervous systems. Their basis is the mathematical model of an artificial neuron. These neurons are arranged into layers and interconnected by links to which weights are assigned. The network processes data by passing input information through these interconnected neurons, which transform the signal using activation functions. Through a process called learning, the weights are gradually adjusted on the basis of the data presented, so that the network adapts to solving a particular task such as classifying data or making predictions.
Put simply, a neural network can be pictured as a system that learns to solve problems somewhat as the human brain does. Instead of being programmed precisely for every task, it learns from examples. Show the network thousands of pictures of dogs and it will learn by itself to recognise the key features, after which it can identify a dog even in a picture it has never seen before. This is how it is used for face recognition, language translation or weather forecasting, because it is good at finding complex patterns in data.
Technical definition: Artificial neural networks are non-linear computational models and data structures inspired by biological neural networks. They consist of a large number of interconnected simple computational units, so-called artificial neurons, typically organised into an input layer, one or more hidden layers and an output layer. Every link between neurons has an assigned numerical value – a weight. A neuron aggregates the weighted inputs from its links, applies an activation function that determines its output, and passes that on to further neurons. The capacity to learn is realised by iteratively adjusting these weights using algorithms such as backpropagation, with the aim of minimising the difference between predicted and actual output on the training data.
Etymology and context: The term derives from biology, where a neuron denotes a nerve cell and a network their interconnection. The first mathematical model of a neuron was proposed by Warren McCulloch and Walter Pitts in 1943. The concept developed further, for example with Frank Rosenblatt’s Perceptron model (1958). Neural networks are today a fundamental tool in machine learning and artificial intelligence, especially in the form of deep neural networks (deep learning), which thanks to a large number of hidden layers and to computing power achieve high success rates in solving complex tasks such as image processing, natural language processing and autonomous driving.