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

We say that a neural network learns from data because of its operating steps. First, all trainable parameters of the model, like weight of connexion and biases are randomly initialised. After, the model improves itself to minimize a loss function that measures the error between predictions and true labels. This is achieved by iterating over each sample of the train dataset: - A prediction is made by the actual model configuration, then the loss is computed and its gradient with respect to all trainable parameters. - Model configuration (all weights and biases) is upgraded, in the direction of gradient with a scaling factor called the learning rate. These two points, forming a learning step, are repeated over the entire dataset and are called a training epoch. After one epoch, model capabilities are evaluated on the validation dataset. Depending on the results, some hyper-parameters could be adjusted like the learning rate. To learn correctly, we repeat many epochs to reach a good predictive model There are many parameters that can be optimized for each problem like the structures for the intermediate layers, changing the activation function, the error function and also the learning rate. This part of the work has been carried on by LISTIC laboratory. They decided to use two types of Neural Networks s: - Dense Neural Network (DNN). are characterized by having all the neurons from a certain layer connected to all neurons from the previous and following layer. Figure 21is an example of a DNN model with all neurons connected to all neurons from neighbor layers. The hidden layer was formed by one or two dense layers using a logistic regression function to activate the neurons in order to observe non linear behaviour of the data. - Convolutional neural networks (CNN). They are famously used to identify patterns on images. Convolution can be used to identify edges, curves and even more complex patterns. The convolution acts like a filter that enhances certain patterns. On images, the convolution kernels, called filters, are matrix while vectors on 1D times series, as for our dataset. 4.2. Performance metrics A common way to observe the performance of a classification model is the confusion matrix. It compares the prediction and the actual values as shown by Figure 23. In order to compare different models, some values, representing the performance of the matrix, are calculated. Many different indicators may be used. The ones chosen were: accuracy, balanced accuracy, precision, recall, and . β 30