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

2.4. CONCLUSION 23 where Y is the true label one-hot encoded and ˆ Y is the model prediction as above. Example. To follow previous example, if models predicts class "3" with ˆ Y = [0 . 2 , 0 . 1 , 0 . 7] and GT class is "3" then Y = [0 , 0 , 1] and then L = − 1 × log(0 . 7) Figure 2.1: On left, scheme of DNNs used with one (or more) hidden layers. On right, focus on one neuron h j in hidden layer. 2.4 Conclusion Here, we introduce some theoretical point around point process and time series that can be inter- esting for our application. Some models was presented, in particular we precise how we train a neural network from labelled data (supervised learning). Rockfall event can be naturally modeled by point process in temporal/spatial domain. As RN1 data are sparse, we aggregate spatially data and look at temporal dynamics (previous chapter): a point process with an inhomogeneous seasonal intensity highly correlated with rainfall point pro- cess. Therefore, quantity of interest would be event time ( T i ) with some features (sum of size ?...). Such point processes can be modeled by neural networks, see Annex B for a recent bibliography. In next chapter, we focus on simple models which target is the daily number of rockfall NB t with sometimes transformation on targets.