Sparse Clustering via Convex Optimization – We propose a new algorithm named Fast Caffe to solve sparse clustering problems. It is based on the observation that if the data points in a dataset are sparse at some point in time, then, our algorithm can learn the same sparse clustering problem as an ordinary Caffe. This is a crucial criterion for any Caffe with sparse data, even when using non-convex regularization. Our experiments on real data show that our algorithm significantly outperforms the normal Caffe in terms of clustering performance, clustering difficulty, and computation time.

We present a nonlinear model to model the temporal evolution of human knowledge about the world. Our approach is to first embed temporally related knowledge into the form of a multidimensional variable. We then embed the inter- and intra-variable covariate into a multidimensional structure in order to model the temporal motion in the multi-dimensional space. The multidimensional structure serves as a feature representation of multidimensional variables and represents temporally related variables in such a way that temporal evolution is also modeled as a multidimensional process of continuous evolution. The multidimensional structure is computed through a novel approach of learning from multidimensional features in a set of labeled items by using a multi-layer recurrent neural network. Experiments on large-scale public datasets show that we achieve state-of-the-art performance on real-world datasets.

Automatic Instrument Tracking in Video Based on Optical Flow (PoOL) for Planar Targets with Planar Spatial Structure

Learning an Integrated Deep Filter based on Hybrid Coherent Cuts

Sparse Clustering via Convex Optimization

Nonlinear Models in Probabilistic Topic Models

Towards a better understanding of the intrinsic value of training topic modelsWe present a nonlinear model to model the temporal evolution of human knowledge about the world. Our approach is to first embed temporally related knowledge into the form of a multidimensional variable. We then embed the inter- and intra-variable covariate into a multidimensional structure in order to model the temporal motion in the multi-dimensional space. The multidimensional structure serves as a feature representation of multidimensional variables and represents temporally related variables in such a way that temporal evolution is also modeled as a multidimensional process of continuous evolution. The multidimensional structure is computed through a novel approach of learning from multidimensional features in a set of labeled items by using a multi-layer recurrent neural network. Experiments on large-scale public datasets show that we achieve state-of-the-art performance on real-world datasets.