The artificial neuron or perceptron was born in the 50s (yes, nothing new!) and it is just the mathematical representation of the real or biological neuron. A biological neuron works by receiving a signal that is later transferred to the nucleus, where it is processed and sent to the axon to exit the neuron. In the figure (left), you can see how the two neurons, real and artificial, look alike. They are both composed by a part that receives information (dendrites versus input variables, x ), a processing structure (nucleus versus activation function, f ), and the part that transfers the information to other neurons (axon versus output variables, y ). The ANN is just a bunch of neurons organized into different layers (right figure). But how does an ANN learn?

Like the human brain, the ANN learns by looking at examples or samples from a problem to be solved. The key is to find the optimal contribution (the so-called weights, w ) of each input variable of each neuron in the ANN. This is done by running an optimization process (known as backpropagation) that is able to transfer the information through the different layers, comparing the values of the input and output variables for each given example.

Let’s go through a simple case: we want to train an ANN to distinguish between beer and coffee. In order to do this, we first need to define several input variables that cover different characteristics of our drinks: color, temperature, sparkling or not, and level of alcohol. To train our ANN we need several samples (the more, the better) so that we cover all possible options. Once all samples have been processed and the optimization process is finished, one can observe that certain input variables contribute more to predict a correct answer (high weights) than others. For instance, the color (there are brown or black beers) or the temperature (coffee can be served with ice) are not so important as the fact to be sparkling or the level of alcohol (clearly typical features of beer). After training, we can just give a new (and unseen) sample to our ANN applying the optimized weights, and we will obtain the answer.

This is a very trivial example, but deep ANNs (with many layers) are able to solve very complicated problems, with so many input variables that would be impossible to process for a human brain.

If you want to know more: https://www.youtube.com/watch?v=bfmFfD2RIcg&t=249s