Using Linguistic Features to Detect and Track Disorder Hints – We study the problem of inferring the linguistic features of an individual by means of a natural language interface, a set of natural language strings, and a corpus of natural language text. Our task involves the discovery of features derived from the natural language string to distinguish the presence of a specific linguistic category. Our approach uses a probabilistic approach to infer the features. First, we identify a subset of features, which are informative (i.e., they are meaningful) and unconfuse (i.e., they might not be useful). The features are then inferred by learning a new set of features, and using multiple learned features to predict the classification decision made. Finally, we model the data using different information sources, as well as a different model for the data, for the purpose of inference and tagging. All of these sources are used to create new features, where they are used to learn discriminative features.

We present an efficient approach for learning sparse vector representations from input signals. Unlike traditional sparse vector representations which typically use a fixed set of labels, our approach does not require labels at all. We show that sparse vectors are flexible representations, allowing the training of networks of arbitrary sizes, with strong bounds on the true number of labels. We then illustrate that a neural network can accurately predict the label accuracy by sampling a sparse vector from a large set of input signals. This study shows a promising strategy for a supervised learning architecture: using such a model for predicting labels, it can be used to predict the true labels with minimal hand-crafted labeling.

Probabilistic Models on Pointwise Triples and Mixed Integer Binary Equalities

Stochastic gradient descent with two-sample tests

Using Linguistic Features to Detect and Track Disorder Hints

Adaptive Learning of Cross-Agent Recommendations with Arbitrary Reward Sets

Multiset Regression Neural Networks with Input SignalsWe present an efficient approach for learning sparse vector representations from input signals. Unlike traditional sparse vector representations which typically use a fixed set of labels, our approach does not require labels at all. We show that sparse vectors are flexible representations, allowing the training of networks of arbitrary sizes, with strong bounds on the true number of labels. We then illustrate that a neural network can accurately predict the label accuracy by sampling a sparse vector from a large set of input signals. This study shows a promising strategy for a supervised learning architecture: using such a model for predicting labels, it can be used to predict the true labels with minimal hand-crafted labeling.