A carregar...
2 resultados
Resultados da pesquisa
A mostrar 1 - 2 de 2
- Modular dynamic neural network: a continual learning architecturePublication . Turner, Daniel; Cardoso, Pedro; Rodrigues, JoãoLearning 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.
- Continual learning for object classification: integrating AutoML for binary classification tasks within a modular dynamic architecturePublication . Turner, Daniel; Cardoso, Pedro; Rodrigues, JoaoFor humans it is quite easy to identify a new object after learning to identify existing ones, but not for a machine. Deep neural networks (DNN) are the foundation of the current state-of-the-art methods for training machines to recognize sets of objects. The issue is that any modification to the DNN weights that were trained to classify an initial set of objects has the potential to seriously impair the network’s ability to make those initial classifications; this behaviour is referred to as catastrophic forgetting (CF). This paper presents a continual learning (CL) architecture that can deal with CF. The architecture is composed of two primary parts: (i) The feature extraction component, which is based on the ResNet50 backbone and (ii) the modular dynamic classification (MDC) component. The latter is made up of multiple sub-networks that gradually assemble themselves into a tree-like structure that reorganizes itself as it learns over time, so that each sub-network can operate independently. The MDC relies heavily on binary classification, and here the application of automated machine learning (AutoML) was introduced, where each binary classifier is tailored on-the-fly, and is/can be different from object to object. The strategy involves a calculated selection from a predefined list of model types and parameters, optimizing them for their respective tasks. Results demonstrate that we advanced the adaptability and performance of the network, emphasizing the transformative potential of AutoML in modular CL approaches. Tests on the CORe50 dataset showed accuracy results of 81.1%, which are above the state of the art for CL architectures.