Relevance Estimation Using Sparse Multidimensional Scaling: Application to Classification and Regression – We present a method to use unsupervised feature learning (similar to Sparse Multi-Class Classification) over large class images (e.g., MNIST and CIFAR-10). Under certain assumptions about the image representations, we establish a new classifier for the task of classification over the MNIST dataset. In this paper, we describe our method, show how it can be used to detect and model the unseen classes and its class labels, and obtain efficient classification results. We demonstrate our method on the MNIST classification task, which achieved state-of-the-art performance when compared to the model-based classification approach, and we show the effectiveness of the proposed method by using different methods to separate the unseen classes in each dataset, and to model the unseen class labels.

In this paper, we propose a novel algorithm for recognizing human actions in videos using the deep convolutional neural network (CNN) in a low rank dimension setting. The proposed algorithm aims at recognizing multiple human actions within a single image. The new CNN model is built using convolutional neural network (CNN) feature and then the data is collected to extract individual actions. The network produces global object detection and segmentation scores of different human actions in a sequence and then the network takes advantage of the visual features extracted from the entire sequence to learn their local actions. The network is trained to recognize human actions with no supervision, which is an improvement over prior works. In contrast to previous works, both CNN and the CNN feature learning is applied in supervised manner and the features extracted from the whole sequence are used to identify the local actions. Moreover, the CNN features were pre-trained to be able to discriminate the actions on the test set. Results shows that the proposed CNN model can be used for real-time action recognition.

Unsupervised learning of motion

Towards Better Analysis of Hierarchical Data Clustering with Applications to Topic Modeling

Relevance Estimation Using Sparse Multidimensional Scaling: Application to Classification and Regression

Stochastic Lifted Bayesian Networks

An Online Corpus of Electronic Medical RecordsIn this paper, we propose a novel algorithm for recognizing human actions in videos using the deep convolutional neural network (CNN) in a low rank dimension setting. The proposed algorithm aims at recognizing multiple human actions within a single image. The new CNN model is built using convolutional neural network (CNN) feature and then the data is collected to extract individual actions. The network produces global object detection and segmentation scores of different human actions in a sequence and then the network takes advantage of the visual features extracted from the entire sequence to learn their local actions. The network is trained to recognize human actions with no supervision, which is an improvement over prior works. In contrast to previous works, both CNN and the CNN feature learning is applied in supervised manner and the features extracted from the whole sequence are used to identify the local actions. Moreover, the CNN features were pre-trained to be able to discriminate the actions on the test set. Results shows that the proposed CNN model can be used for real-time action recognition.