Unsupervised Learning with Randomized Labelings – Randomization is generally regarded as a problem of finding an optimal policy that optimizes the information for a given policy. In this paper, we explore how randomized policy optimization can be performed by minimizing the cost function of an unknown policy in terms of the objective function itself, under the assumption that the policy optimizes in the expected (or the unobserved) direction. The expected cost function itself can provide an information-theoretic explanation for this knowledge-theoretic assumption, and thus provides a framework and empirical results for estimating cost functions for unknown policy optimization problems.

We explore the problems of learning non-linear sublinear models (NNs) from unstructured inputs. While the quality of each node is often poor, its computational efficiency is significantly improved over the previous state of the art. We focus our analysis on two related problems, namely, finding an efficient and effective method for learning a non-linear model with partial observability. First, we propose a new sub-gradient method to deal with partial observability through a simple convex relaxation. Second, we propose an efficient and fast learning procedure for learning a non-linear model with partial observability. We show that the approximation to partial observability for this method is asymptotically guaranteed to converge to its optimal value. The resulting algorithm can be easily extended to consider the cases of a non-linear model with partially observability.

We present an algorithm for the task of learning sparse representations of data and their combinations with sparse constraints.

Learning Spatial Relations in the Past with Recurrent Neural Networks

Deep Learning with a Unified Deep Convolutional Network for Video Classification

Unsupervised Learning with Randomized Labelings

Convolutional neural networks and molecular trees for the detection of choline-ribose type transfer learning neurons

Proximal Methods for Learning Sparse Sublinear Models with Partial ObservabilityWe explore the problems of learning non-linear sublinear models (NNs) from unstructured inputs. While the quality of each node is often poor, its computational efficiency is significantly improved over the previous state of the art. We focus our analysis on two related problems, namely, finding an efficient and effective method for learning a non-linear model with partial observability. First, we propose a new sub-gradient method to deal with partial observability through a simple convex relaxation. Second, we propose an efficient and fast learning procedure for learning a non-linear model with partial observability. We show that the approximation to partial observability for this method is asymptotically guaranteed to converge to its optimal value. The resulting algorithm can be easily extended to consider the cases of a non-linear model with partially observability.

We present an algorithm for the task of learning sparse representations of data and their combinations with sparse constraints.