We emphasize the need for input-dependent regularization in the context of conditional density models (also: discriminative models) like Gaussian process predictors. This can be achieved by a simple modification of the standard Bayesian data generation model underlying these techniques. Specifically, we allow the latent target function to be apriori dependent on the distribution of the input po...

This work proposes an alternative to ordered subsets to improve the convergence speed of list-mode expectationmaximization image reconstruction algorithms. Instead of subdividing the input data into non-overlapping subsets, the stream of measured coincidence events is immediately processed online. The reconstruction algorithm maintains a sliding window covering the events that contribute to the...

We present an approach to train a model for classifying ice and open water directly using the polygon-wise ice concentration available from ice charts. This can be considered as a “learning from label proportions” (LFLP) problem which has been studied in the last decade and applied to many real-world applications. Our approach is based on convolutional neural networks (CNNs), which have been sh...

We obtain improved running times for two algorithms for clustering data: the expectationmaximization (EM) algorithm and Lloyd's algorithm. The EM algorithm is a heuristic for finding a mixture of k normal distributions in Rd that maximizes the probability of drawing n given data points. Lloyd's algorithm is a special case of this algorithm in which the covariance matrix of each normally-distrib...

This paper shows how a text classifier’s need for labeled training documents can be reduced by taking advantage of a large pool of unlabeled documents. We modify the Query-by-Committee (QBC) method of active learning to use the unlabeled pool for explicitly estimating document density when selecting examples for labeling. Then active learning is combined with ExpectationMaximization in order to...

A practical method for creating a high dimensional index structure that adapts to the data distribution and scales well with the database size, is presented. Typical media descriptors, such as texture features, are high dimensional and are not uniformly distributed in the feature space. The performance of many existing methods degrade if the data is not uniformly distributed. The proposed metho...

Clustering in high-dimensional spaces is a difficult problem which is recurrent in many domains, for example in image analysis. The difficulty is due to the fact that highdimensional data usually live in different low-dimensional subspaces hidden in the original space. This paper presents a family of Gaussian mixture models designed for highdimensional data which combine the ideas of subspace c...

Consider the problem of tting a nite Gaussian mixture, with an unknown number of components, to observed data. This paper proposes a new minimum description length (MDL) type criterion, termed MMDL (formixtureMDL), to select the number of components of the model. MMDL is based on the identi cation of an \equivalent sample size", for each component, which does not coincide with the full sample s...

A deep generative model is developed for representation and analysis of images, based on a hierarchical convolutional dictionary-learning framework. Stochastic unpooling is employed to link consecutive layers in the model, yielding top-down image generation. A Bayesian support vector machine is linked to the top-layer features, yielding max-margin discrimination. Deep deconvolutional inference ...

This paper presents a fast variational Bayesian method for linear state-space models. The standard variational Bayesian expectationmaximization (VB-EM) algorithm is improved by a parameter expansion which optimizes the rotation of the latent space. With this approach, the inference is orders of magnitude faster than the standard method. The speed of the proposed method is demonstrated on an art...