Robust Sparse Subspace Clustering – We propose a novel technique and computational framework for unsupervised clustering of low-level features from a large unlabeled collection of labeled data (e.g., image data) with minimal training set. In order to obtain the best of both worlds, we propose to exploit low-level features and clusters in a highly discriminative manner, which is the best case scenario for our approach. We first apply the method to image data and demonstrate its usefulness for unsupervised clustering based on the similarity between the clustering result and the labels. We further use it to analyze the effect of labels and then evaluate whether or not a high-level clustering method can have the same performance. Experimental results show that using low-level features, which is desirable for image data, outperforms the supervised clustering approach, as well as other clustering methods proposed in this paper.

Image data have been a major source of error during the past decades. The primary focus of this paper is to develop a robust and practical framework for image retrieval (i.e. the extraction of images from social media). The data collected from social media content of the internet-based web enables to extract relevant features from the images, such as semantic, visual, contextual, language, and textual labels. We show that, although natural language processing (NN) approaches can extract these features without using images, it is not practical for using social networks for this purpose. To address the problem, we propose a deep convolutional neural network (CNN) with feature extraction algorithms, which significantly outperforms the state-of-the-art. This is in accord with the proposed training paradigm, which combines the best techniques from CNNs with image extraction. We illustrate the benefits of the proposed methodology using both synthetic and real data sets, showing that for a given dataset, learning the features is far from the best solution.

Bayesian Inference for Gaussian Processes

Recurrent Online Prediction: A Stochastic Approach

Robust Sparse Subspace Clustering

Handling Propositional Problems: The Hard and `Parsimonious Problem

Deep Learning for Real Detection with Composed-Seq ImagesImage data have been a major source of error during the past decades. The primary focus of this paper is to develop a robust and practical framework for image retrieval (i.e. the extraction of images from social media). The data collected from social media content of the internet-based web enables to extract relevant features from the images, such as semantic, visual, contextual, language, and textual labels. We show that, although natural language processing (NN) approaches can extract these features without using images, it is not practical for using social networks for this purpose. To address the problem, we propose a deep convolutional neural network (CNN) with feature extraction algorithms, which significantly outperforms the state-of-the-art. This is in accord with the proposed training paradigm, which combines the best techniques from CNNs with image extraction. We illustrate the benefits of the proposed methodology using both synthetic and real data sets, showing that for a given dataset, learning the features is far from the best solution.