Examinando por Autor "Llopis-Ibor, L."

Mostrando 1 - 3 de 3

DC Neural Networks avoid overfitting in one-dimensional nonlinear regression
(Elsevier, 2024-01-11) Beltran-Royo, C.; Llopis-Ibor, L.; Ramirez, I.; Pantrigo, J.J.
In this paper, we analyze Difference of Convex Neural Networks in the context of one-dimensional nonlinear regression. Specifically, we show the surprising ability of the Difference of Convex Multilayer Perceptron (DC-MLP) to avoid overfitting in nonlinear regression. Otherwise said, DC-MLPs self-regularize (do not require additional regularization techniques). Thus, DC-MLPs could result very useful for practical purposes based on one-dimensional nonlinear regression. It turns out that shallow MLPs with a convex activation (ReLU, softplus, etc.) fall in the class of DC-MLPs. On the other hand, we call SQ-MLP the shallow MLP with a Squashing activation (logistic, hyperbolic tangent, etc.). In the numerical experiments, we show that DC-MLPs used for nonlinear regression avoid overfitting, in contrast with SQ-MLPs. We also compare DC-MLPs and SQ-MLPs from a theoretical point of view
Fast incremental learning by transfer learning and hierarchical sequencing
(Elsevier, 2023-02) Llopis-Ibor, L.; Beltrán-Royo, César; Cuesta-Infante, A.; Pantrigo, J.J.
In this paper we address the Class Incremental Learning (CIL) problem, characterized by sequences of data batches in which examples of different classes occur at different times. From a theoretical point of view, we propose a new approach that we call hierarchical sequencing and prove that any CIL task can be sequenced into simple incremental classification tasks by means of the hierarchical sequencing. From a practical point of view, we propose the HILAND method for image classification, which combines the hierarchical sequencing with transfer learning. In our experiments, the HILAND method has obtained state-of-the-art results for the CIL problem, but with far less training effort through transfer learning
Human Activity Recognition with Capsule Networks
(Springer, 2021-09-13) Llopis-Ibor, L.; Cuesta-Infante, A.; Beltrán-Royo, César; Pantrigo, J.J.
Human activity recognition is a challenging problem, where deep learning methods are showing to be very efficient. In this paper we propose the use of capsule networks. This type of networks have proved to generalize better to novel viewpoints than convolutional neural networks. We show that the use of capsule networks into a straightforward architecture, between a convolutional preprocessing stage to extract visual features and a header for carrying out the task, is able to attain competitive results with spatio-temporal data without the use of any kind of recurrent neural network. Moreover, an analysis of the obtained results shows that our architecture is capable of learning the properties that encode the spatio-temporal dynamics of the movements that characterize each activity