Learning and Valuing Representations with Neural Models of Sentences and Entities – We study the problem of constructing neural networks with an attentional model. Neural networks are very accurate at representing the semantic interactions among entities. However, their computationally expensive encoding task often produces a negative prediction, leading to a highly inefficient representation learning approach. This issue, however, is resolved by a number of approaches that are computationally efficient for this task. Here we derive a new algorithm that is computationally efficient for learning neural networks with a fixed memory bandwidth. We show here that it achieves the same performance as the usual neural network encoding of entities, and even outperforms it when the memory bandwidth is reduced by a factor of 5 or less. We apply the new algorithm to the problem of learning neural networks with fixed memory bandwidth, and show that it achieves a linear loss in the accuracy of the encoding.

As an alternative to the classic sparse vector factorization (SVM), we propose a two-vector (2V) representation of the data, which is well suited to handle nonnegative matrices. In contrast to the typical sparse learning model that tries to preserve the identity or preserve features, we show that our 2V representation can handle matrices with large dimensionality, by using a new variant of the convex relaxation of the log-likelihood. Our result results show a substantial improvement of the state-of-the-art approach in dimensionality reduction over sparse data, and is based on the principle that a linear approximation of the log-likelihood is equivalent to a convex relaxation.

Dynamic Programming as Resource-Bounded Resource Control

Faster Rates for the Regularized Loss Modulation on Continuous Data

# Learning and Valuing Representations with Neural Models of Sentences and Entities

Robust Constraint Handling with Answer Set Programming

Tight Inference for Non-Negative Matrix FactorizationAs an alternative to the classic sparse vector factorization (SVM), we propose a two-vector (2V) representation of the data, which is well suited to handle nonnegative matrices. In contrast to the typical sparse learning model that tries to preserve the identity or preserve features, we show that our 2V representation can handle matrices with large dimensionality, by using a new variant of the convex relaxation of the log-likelihood. Our result results show a substantial improvement of the state-of-the-art approach in dimensionality reduction over sparse data, and is based on the principle that a linear approximation of the log-likelihood is equivalent to a convex relaxation.