An artificial neuron is the basic computational unit of a neural network, inspired by the function of biological neurons in the brain. Its job is to accept one or more input signals, process them and pass an output signal on. Each input has an assigned weight determining its importance. The neuron sums all these weighted inputs and applies an activation function to the result. That function decides whether, and with what intensity, the signal will be passed to the next neurons in the network.
Put simply, an artificial neuron can be pictured as a little decision-making clerk. The clerk receives several pieces of information (inputs), each carrying a different importance (weight). On the basis of the sum of this weighted information he then decides whether the message is significant enough to pass on to his superior. The whole neural network is then like a large company full of such clerks passing information to one another so as to reach a final decision together.
Technical definition: An artificial neuron is a mathematical function representing the basic building block of artificial neural networks. It is a model inspired by the biological neuron. The neuron receives a set of input values x1, x2, …, xn, where each input xi is assigned a synaptic weight wi. Within the body of the neuron a weighted sum of the inputs is computed, often with the addition of a threshold value (bias, b). The resulting value, the internal potential, is then transformed by an activation function φ, which determines the neuron’s output y = φ(∑ wi xi + b). The activation function may be linear or non-linear (step, sigmoid, ReLU) and fundamentally affects the computational capabilities of the network.
Etymology and context: The name is a direct analogy to the biological neuron, whose function it attempts to imitate in simplified form. The model’s inputs correspond to dendrites, the weights represent synaptic strength, the body of the neuron performs the summation, and the axon with its activation function decides whether the signal is transmitted. The first formal model of a neuron, known as the McCulloch-Pitts neuron, was introduced in 1943. Today artificial neurons, arranged into complex architectures (layers), are the key element of machine learning and deep learning algorithms, where they make it possible to solve tasks such as classification, regression and pattern recognition.