An Empirical Evaluation of Unsupervised Deep Learning for Visual Tracking and Recognition – We present a large-scale evaluation of Deep Convolutional Neural Networks (DNNs) on large data collections from CNNs. We compare the results obtained on a set of the most recent set of CNNs that were used. Most of the previously published DNN evaluations were made on large datasets. This paper will provide a step-by-step overview of both CNNs and DNNs in the context of the data collection. Our purpose is both to highlight the strengths and weaknesses of DCNNs and to discuss the best DNN in two main tasks: 3D image segmentation and 2D scene segmentation.

We address the problem of learning an optimal model of a target image to generate a given set of features. We build on the success of recent progress in neural networks to model the problem. While in the past we have proposed methods for learning to learn features, our approach is based on the first order optimization of the weights of a convolutional neural network model, which allows our solution to take the form of the learning process. We demonstrate that our approach outperforms prior state-of-the-art learning algorithms with a very strong performance on classification tasks of small sample sizes. In particular, we show that the learned features improve significantly when compared to traditional state-of-the-art representations.

Theorem Proving: The Devil is in the Tails! Part II: Theoretical Analysis of Evidence, Beliefs and Realizations

Multi-Winner Semi-Supervised Clustering using a Structured Boltzmann Machine

An Empirical Evaluation of Unsupervised Deep Learning for Visual Tracking and Recognition

Fast and Accurate Low Rank Estimation Using Multi-resolution Pooling

Deep Learning, A Measure of Deep Inference, and a Quantitative AlgorithmWe address the problem of learning an optimal model of a target image to generate a given set of features. We build on the success of recent progress in neural networks to model the problem. While in the past we have proposed methods for learning to learn features, our approach is based on the first order optimization of the weights of a convolutional neural network model, which allows our solution to take the form of the learning process. We demonstrate that our approach outperforms prior state-of-the-art learning algorithms with a very strong performance on classification tasks of small sample sizes. In particular, we show that the learned features improve significantly when compared to traditional state-of-the-art representations.