Improving the Robustness of Deep Neural Networks by Exploiting Connectionist Sampling – Recent research on deep learning has focused on minimizing the computational cost as a condition to perform inference. We propose an adaptive inference algorithm that encourages sub-parameters to be learned from input data to improve inference in a robust way. The objective is to find the optimal parameters of the network using an estimator that learns the best estimates of the underlying latent factors. To this end, for each sub-modular variable, we propose an adaptive estimator that predicts the likelihood that most of the parameters of the network are learned and the worst estimates of the parameters of the network are ignored. This estimator is shown to outperform previous estimators that are able to learn the best estimates. We apply our algorithm to two datasets of synthetic and real data collections.
This paper addresses the problem of learning the shape space of a high-dimensional (H) dimensional data set. To that end our contribution is a Bayesian framework that learns the shape space of the data set by solving a general Bayesian optimization problem. The framework shows that the H-dimensional data sets are not very compact, and hence a novel optimization problem is approached. The proposed framework is evaluated using the PASCAL VOC dataset, where it outperforms state-of-the-art methods in terms of accuracy and complexity.
We demonstrate how to recover an image with low-level features by solving low-level semantic image restoration problems. We propose a new approach that uses visual attention, by exploiting the fact that the human gaze can only appear in the visual domain. The technique is simple: we first extract low-level features from the image. Then the human gaze is recovered by means of our method. The proposed approach is evaluated on different synthetic datasets that provide very promising results.
Improving the Accuracy of the LLE Using Multilayer Perceptron
Improving the Robustness of Deep Neural Networks by Exploiting Connectionist Sampling
A Data Mining Framework for Answering Question Answering over Text
Unsupervised feature learning using adaptive thresholding for object clusteringThis paper addresses the problem of learning the shape space of a high-dimensional (H) dimensional data set. To that end our contribution is a Bayesian framework that learns the shape space of the data set by solving a general Bayesian optimization problem. The framework shows that the H-dimensional data sets are not very compact, and hence a novel optimization problem is approached. The proposed framework is evaluated using the PASCAL VOC dataset, where it outperforms state-of-the-art methods in terms of accuracy and complexity.
We demonstrate how to recover an image with low-level features by solving low-level semantic image restoration problems. We propose a new approach that uses visual attention, by exploiting the fact that the human gaze can only appear in the visual domain. The technique is simple: we first extract low-level features from the image. Then the human gaze is recovered by means of our method. The proposed approach is evaluated on different synthetic datasets that provide very promising results.
Leave a Reply