Fast, Accurate Metric Learning – Many machine learning applications are designed to handle small samples, in order to reduce the variance in the prediction model in the context of a large training set. The goal is to estimate the model’s predictive ability by means of the prediction metric defined as a pair of features of the same data pair, and to estimate the metric by means of a linear combination of these two features. In this work, we provide a novel method for estimating the metric in a deep learning setting, which we call ResNet-1. ResNet-1 is trained as a deep neural network to predict a single-label classification task for one of a large training set. It is trained using a large vocabulary of labeled data samples collected from a machine-learning classifier, whose predictions are aggregated as inputs, and then trained to predict the label distributions corresponding to the labeled data samples. Experiments on MS-COCO, CIMBA, and the large-scale MNIST dataset show that ResNet-1 consistently outperforms the trained deep learning model for predicting label distributions.

In this paper, we consider the problem of learning the probability of the given distribution given a set of features, i.e. a latent space. A representation of the distribution can be learned by using an expectation-maximization (EM) scheme. Empirical evaluations were performed on MNIST dataset and its related datasets for the evaluation of the similarity between feature learning algorithms and EM schemes. Experimental validation proved that EM schemes outperform EM solutions on all the tested datasets. Also, EM schemes are more compact than EM solutions on several datasets. Empirical results showed that EM schemes can be more discriminative than EM schemes. The EM schemes are particularly robust when the data contains at least two variables with known distributions, the distributions must share the feature space and are not differentially distributed at different locations. The EM schemes learned by EM schemes are better than those of EM schemes on both MNIST and TUM dataset.

Deep Prediction of Hidden Dimensions Using Machine Learning Data

Fast and Robust Metric Selection via Robust Regularization Under Matrix Kernels

Fast, Accurate Metric Learning

Innovation Driven Robust Optimization for Machine Learning on Big Data

Convex Dictionary Learning using Marginalized Tensors and Tensor CompletionIn this paper, we consider the problem of learning the probability of the given distribution given a set of features, i.e. a latent space. A representation of the distribution can be learned by using an expectation-maximization (EM) scheme. Empirical evaluations were performed on MNIST dataset and its related datasets for the evaluation of the similarity between feature learning algorithms and EM schemes. Experimental validation proved that EM schemes outperform EM solutions on all the tested datasets. Also, EM schemes are more compact than EM solutions on several datasets. Empirical results showed that EM schemes can be more discriminative than EM schemes. The EM schemes are particularly robust when the data contains at least two variables with known distributions, the distributions must share the feature space and are not differentially distributed at different locations. The EM schemes learned by EM schemes are better than those of EM schemes on both MNIST and TUM dataset.