Deep learning model combination and regularization using convolutional neural networks

• Autor(a) principal: Frazão, Xavier Marques
• Data de Publicação: 2014
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.6/5605
• topic: Convolutional Neural Networks; Network Ensemble; Object Recognition; Regularization; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Convolutional neural networks (CNNs) were inspired by biology. They are hierarchical neural networks whose convolutional layers alternate with subsampling layers, reminiscent of simple and complex cells in the primary visual cortex [Fuk86a]. In the last years, CNNs have emerged as a powerful machine learning model and achieved the best results in many object recognition benchmarks [ZF13, HSK+12, LCY14, CMMS12]. In this dissertation, we introduce two new proposals for convolutional neural networks. The first, is a method to combine the output probabilities of CNNs which we call Weighted Convolutional Neural Network Ensemble. Each network has an associated weight that makes networks with better performance have a greater influence at the time to classify a pattern when compared to networks that performed worse. This new approach produces better results than the common method that combines the networks doing just the average of the output probabilities to make the predictions. The second, which we call DropAll, is a generalization of two well-known methods for regularization of fully-connected layers within convolutional neural networks, DropOut [HSK+12] and DropConnect [WZZ+13]. Applying these methods amounts to sub-sampling a neural network by dropping units. When training with DropOut, a randomly selected subset of the output layer’s activations are dropped, when training with DropConnect we drop a randomly subsets of weights. With DropAll we can perform both methods simultaneously. We show the validity of our proposals by improving the classification error on a common image classification benchmark.