Axiomatic Properties of Two-Stream Convolutional Neural Networks – The success and popularity of artificial neural networks has been largely attributed to the ability to generalize from training data. However, the importance of the training data is not fully understood. On the contrary, it is becoming more and more clear that the training data is not generalizable. In this work, we show that the generalization ability of neural networks for the task of recognition is largely dependent on its local representation over the global context, where the input data is a global context. The proposed framework uses one recurrent representation of the global context to perform local attention based discriminative models on feature maps of the local context, and learns local attention patterns for extracting the global context for the training data. Our experimental results show that the proposed framework can improve the generalization ability of neural networks, while learning relevant local attention patterns.
Many of the recent proposals for visual concept recognition have focused on the task of learning visual concepts. In this work, we propose a visual concept recognition model trained on convolutional neural network (CNN) models to learn visual concepts from a sequence of images. After training on the CNN model, a discriminator classifier is trained on this dataset to determine whether visual concepts are present in the images. Experiments show that the proposed model learns the visual concept representations of CNNs for visual concepts without using any visual concept labels and on a set of visual concept datasets, showing that the learned visual concepts represent higher recognition rates, and that visual concepts are more likely to be learned than image labels.
A Computational Study of Learning Functions in Statistical Language Models
A deep learning model for the identification of drivers with susceptibility to fraud
Axiomatic Properties of Two-Stream Convolutional Neural Networks
Multi-view Graph Convolutional Neural Network
Learning Visual Concepts from Text in Natural ScenesMany of the recent proposals for visual concept recognition have focused on the task of learning visual concepts. In this work, we propose a visual concept recognition model trained on convolutional neural network (CNN) models to learn visual concepts from a sequence of images. After training on the CNN model, a discriminator classifier is trained on this dataset to determine whether visual concepts are present in the images. Experiments show that the proposed model learns the visual concept representations of CNNs for visual concepts without using any visual concept labels and on a set of visual concept datasets, showing that the learned visual concepts represent higher recognition rates, and that visual concepts are more likely to be learned than image labels.