RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

4.1.3. Neural Network (NN) Receiving its name after the human brain this method mimics its logic having a series of neurons that activate each other. The simplest model is created by a first layer, with one neuron per feature, a series of layers with various neurons and a final layer with the output that gives the confidence of the model for each classification (Figure 21). Each neuron receives a value from the previous layer of neurons, transforms the value and sends this value to the next layer. The value each neuron from a layer receives is a linear combination of a number of neurons from the previous layers (Figure 22). For each neuron there will be one weight for each neuron on the previous layer that is connected to this neuron and a bias. Each neuron is not necessarily linked to neurons from the previous layer. This linear combination is the input value that will than be inserted on an activation function that will transform it into a value between 0 and 1. This value will then be passed on to the next layer until the last layer which will provide the probability the value is a certain class. Figure 21: example of a Neural Network structure Figure 22: representation of a neuron (x_i the input, y output, Sigma for the sum of the inputs, and g is the activation function) 29