Predicting protein-ligand binding sites by deep learning with single-label sparse predictor learning – Deep learning models are known to be a promising and efficient approach to statistical inference. This work investigates the performance of an efficient non-parametric predictor learning method for a broad class of sparse estimation problems. In this work, we show that the problem of sparse prediction is significantly larger than that of Bayesian estimation of the same function in the data set, and indeed is exponentially smaller than that of non-parametric inference schemes. This is because the number of parameters grows exponentially with the number of examples. We discuss a new non-parametric predictor learning method, which is robust to the size of the predictor, and we show how it can be used to learn to predict the number of examples for a given class by learning from the data. We provide empirical results that demonstrate that the predictor learning method achieves state-of-the-art performance when all the parameters of the predictor are sparse.

In this paper, we propose the solution to the problem of learning a Bayesian posterior using a low-dimensional Euclidean space in particular. We have proposed a novel general framework based on the notion of a low-dimensional Euclidean space. The idea is to map the space into a low-dimensional space using a finite-dimensional Euclidean embedding on the space. Our new formulation in this framework yields a convex relaxation of the posterior probability distribution as a low-dimensional unit and a vector embedding that encodes the posterior probability distribution. The result of the method is that, in a non-convex setting, an unknown variable of interest is given to the posterior probability distribution and the posterior likelihood of the embedding is obtained with the minimax relaxation. We also propose a novel way to learn the embedding using an orthogonal dictionary learning algorithm. Experiments on both synthetic and real data show that the embedding can achieve state-of-the-art performance and outperforms Euclidean-based posterior estimation.

Convex Dictionary Learning using Marginalized Tensors and Tensor Completion

Distributed Sparse Signal Recovery

# Predicting protein-ligand binding sites by deep learning with single-label sparse predictor learning

Recurrent Topic Models for Sequential Segmentation

Fast, Accurate Metric LearningIn this paper, we propose the solution to the problem of learning a Bayesian posterior using a low-dimensional Euclidean space in particular. We have proposed a novel general framework based on the notion of a low-dimensional Euclidean space. The idea is to map the space into a low-dimensional space using a finite-dimensional Euclidean embedding on the space. Our new formulation in this framework yields a convex relaxation of the posterior probability distribution as a low-dimensional unit and a vector embedding that encodes the posterior probability distribution. The result of the method is that, in a non-convex setting, an unknown variable of interest is given to the posterior probability distribution and the posterior likelihood of the embedding is obtained with the minimax relaxation. We also propose a novel way to learn the embedding using an orthogonal dictionary learning algorithm. Experiments on both synthetic and real data show that the embedding can achieve state-of-the-art performance and outperforms Euclidean-based posterior estimation.