Modular dynamic neural network: a continual learning architecture

• Autor(a) principal: Turner, Daniel
• Data de Publicação: 2021
• Outros Autores: Cardoso, Pedro, Rodrigues, João
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.1/17460
• topic: Continual learning; Neural networks; Catastrophic forgetting; Object recognition

Learning to recognize a new object after having learned to recognize other objects may be a simple task for a human, but not for machines. The present go-to approaches for teaching a machine to recognize a set of objects are based on the use of deep neural networks (DNN). So, intuitively, the solution for teaching new objects on the fly to a machine should be DNN. The problem is that the trained DNN weights used to classify the initial set of objects are extremely fragile, meaning that any change to those weights can severely damage the capacity to perform the initial recognitions; this phenomenon is known as catastrophic forgetting (CF). This paper presents a new (DNN) continual learning (CL) architecture that can deal with CF, the modular dynamic neural network (MDNN). The presented architecture consists of two main components: (a) the ResNet50-based feature extraction component as the backbone; and (b) the modular dynamic classification component, which consists of multiple sub-networks and progressively builds itself up in a tree-like structure that rearranges itself as it learns over time in such a way that each sub-network can function independently. The main contribution of the paper is a new architecture that is strongly based on its modular dynamic training feature. This modular structure allows for new classes to be added while only altering specific sub-networks in such a way that previously known classes are not forgotten. Tests on the CORe50 dataset showed results above the state of the art for CL architectures.