Authors:
Bihi Sabiri
1
;
Bouchra El Asri
1
and
Maryem Rhanoui
2
Affiliations:
1
IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University in Rabat, Morocco
;
2
Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco
Keyword(s):
Data Overfitting, Machine Learning, Dropout, Deep Learning, Convolutional Neural Network, Max Pooling, Early Stopping.
Abstract:
The objective of a deep learning neural network is to have a final model that performs well both on the data used to train it and the new data on which the model will be used to make predictions. Overfitting refers to the fact that the predictive model produced by the machine learning algorithm adapts well to the training set. In this case, the predictive model will capture the generalizable correlations and the noise produced by the data and will be able to give very good predictions on the data of the training set, but it will predict badly on the data that it has not yet seen during his learning phase. This paper proposes two techniques among many others to reduce or prevent overfitting. Furthermore, by analyzing dynamics during training, we propose a consensus classification algorithm that avoids overfitting, we investigate the performance of these two types of techniques in convolutional neural network. Early stopping allowing to save the hyper-parameters of a model at the right
time. And the dropout making the learning of the model harder allowing to gain up to more than 50% by decreasing the loss rate of the model.
(More)