A Novel Continual Learning Approach for Competitive Neural Networks

Esteban J. Palomo ^[1] ; Juan Miguel Ortiz-de-Lazcano-Lobato ^[1] ; José David Fernández-Rodríguez ^[1] ; Ezequiel López-Rubio ^[1] ; Rosa María Maza-Quiroga ^[1]
1. [1] Universidad de Málaga
  
  Universidad de Málaga
  
  Málaga, España
Localización: Bio-inspired Systems and Applications: from Robotics to Ambient Intelligence: 9th International Work-Conference on the Interplay Between Natural and Artificial Computation, IWINAC 2022, Puerto de la Cruz, Tenerife, Spain, May 31 – June 3, 2022, Proceedings, Part II / José Manuel Ferrández Vicente (dir. congr.), José Ramón Álvarez Sánchez (dir. congr.), Félix de la Paz López (dir. congr.), Hojjat Adeli (aut.), 2022, ISBN 978-3-031-06527-9, págs. 223-232
Idioma: inglés
Texto completo no disponible (Saber más ...)
Resumen
- Continual learning tries to address the stability-plasticity dilemma to avoid catastrophic forgetting when dealing with non-stationary distributions. Prior works focused on supervised or reinforcement learning, but few works have considered continual learning for unsupervised learning methods. In this paper, a novel approach to provide continual learning for competitive neural networks is proposed. To this end, we have proposed a different learning rate function that can cope with non-stationary distributions by adapting the model to learn continuously. Experimental results performed with different synthetic images that change over time confirm the performance of our proposal.