In this paper we explore the problem of biasing unsupervised models to favor sparsity. We extend the posterior regularization framework [8] to encourage the model to achieve posterior sparsity on the unlabeled training data. We apply this new method to learn first-order HMMs for unsupervised part-of-speech (POS) tagging, and show that HMMs learned this way consistently and significantly out-performs both EM-trained HMMs, and HMMs with a sparsity-inducing Dirichlet prior trained by variational EM. We evaluate these HMMs on three languages — English, Bulgarian and Portuguese — under four conditions. We find that our method always improves performance with respect to both baselines, while variational Bayes actually degrades performance in most cases. We increase accuracy with respect to EM by 2.5%-8.7% absolute and we see improvements even in a semisupervised condition where a limited dictionary is provided.

For many computer vision applications, the ideal image feature would be invariant to multiple confounding image properties, such as illumination and viewing angle. Recently, deep architectures trained in an unsupervised manner have been proposed as an automatic method for extracting useful features. However, outside of using these learning algorithms in a classifier, they can be sometimes difficult to evaluate. In this paper, we propose a number of empirical tests that directly measure the degree to which these learned features are invariant to different image transforms. We find that deep autoencoders become invariant to increasingly complex image transformations with depth. This further justifies the use of “deep” vs. “shallower” representations. Our performance metrics agree with existing measures of invariance. Our evaluation metrics can also be used to evaluate future work in unsupervised deep learning, and thus help the development of future algorithms.

Contrast statistics of the majority of natural images conform to a Weibull distribution. This property of natural images may facilitate efficient and very rapid extraction of a scenes visual gist. Here we investigate whether a neural response model based on the Weibull contrast distribution captures visual information that humans use to rapidly identify natural scenes. In a learning phase, we measure EEG activity of 32 subjects viewing brief flashes of 800 natural scenes. From these neural measurements and the contrast statistics of the natural image stimuli, we derive an across subject Weibull response model. We use this model to predict the responses to a large set of new scenes and estimate which scene the subject viewed by finding the best match between the model predictions and the observed EEG responses. In almost 90 percent of the cases our model accurately predicts the observed scene. Moreover, in most failed cases, the scene mistaken for the observed scene is visually similar to the observed scene itself. These results suggest that Weibull contrast statistics of natural images contain a considerable amount of scene gist information to warrant rapid identification of natural images.

We show that convex KL-regularized objective functions are obtained from a PAC-Bayes risk bound when using convex loss functions for the stochastic Gibbs classifier that upper-bound the standard zero-one loss used for the weighted majority vote. By restricting ourselves to a class of posteriors, that we call quasi uniform, we propose a simple coordinate descent learning algorithm to minimize the proposed KL-regularized cost function. We show that standard ell_p-regularized objective functions currently used, such as ridge regression and ell_p-regularized boosting, are obtained from a relaxation of the KL divergence between the quasi uniform posterior and the uniform prior. We present numerical experiments where the proposed learning algorithm generally outperforms ridge regression and AdaBoost.

We present a new empirical risk minimization framework for approximating functions from training samples for low-dimensional regression applications where a lattice (look-up table) is stored and interpolated at run-time for an efficient hardware implementation. Rather than evaluating a fitted function at the lattice nodes without regard to the fact that samples will be interpolated, the proposed lattice regression approach estimates the lattice to minimize the interpolation error on the given training samples. Experiments show that lattice regression can reduce mean test error compared to Gaussian process regression for digital color management of printers, an application for which linearly interpolating a look-up table (LUT) is standard. Simulations confirm that lattice regression performs consistently better than the naive approach to learning the lattice, particularly when the density of training samples is low.

Little work has been done to directly combine the outputs of multiple supervised and unsupervised models. However, it can increase the accuracy and applicability of ensemble methods. First, we can boost the diversity of classification ensemble by incorporating multiple clustering outputs, each of which provides grouping constraints for the joint label predictions of a set of related objects. Secondly, ensemble of supervised models is limited in applications which have no access to raw data but to the meta-level model outputs. In this paper, we aim at calculating a consolidated classification solution for a set of objects by maximizing the consensus among both supervised predictions and unsupervised grouping constraints. We seek a global optimal label assignment for the target objects, which is different from the result of traditional majority voting and model combination approaches. We cast the problem into an optimization problem on a bipartite graph, where the objective function favors smoothness in the conditional probability estimates over the graph, as well as penalizes deviation from initial labeling of supervised models. We solve the problem through iterative propagation of conditional probability estimates among neighboring nodes, and interpret the method as conducting a constrained embedding in a transformed space, as well as a ranking on the graph. Experimental results on three real applications demonstrate the benefits of the proposed method over existing alternatives.

Image representation based on image bases provides a framework for understanding neural representation of visual perception. A recent fMRI study has shown that arbitrary contrast-defined visual images can be reconstructed from fMRI activity patterns using a combination of multi-scale local image bases. In the reconstruction model, the mapping from an fMRI activity pattern to the contrasts of the image bases was learned from measured fMRI responses to visual images. But the shapes of the images bases were fixed, and thus may not be optimal for reconstruction. Here, we propose a method to build a reconstruction model in which image bases are automatically extracted from the measured data. We constructed a probabilistic model that relates the fMRI activity space to the visual image space via a set of latent variables. The mapping from the latent variables to the visual image space can be regarded as a set of image bases. We found that spatially localized, multi-scale image bases were estimated near the fovea, and that the model using the estimated image bases was able to accurately reconstruct novel visual images. The proposed method provides a means to discover a novel functional mapping between stimuli and brain activity patterns.

This is often referred to as the MAP (maximum a--posteriori) problem. Of particular interest is the case of MAP in graphical Inodels7 i.e., models where the prob--

Orthogonal matching pursuit (OMP) is a widely used greedy algorithm for recovering sparse vectors from linear measurements. A well-known analysis of Tropp and Gilbert shows that OMP can recover a k-sparse n-dimensional real vector from m = 4k log(n) noise-free random linear measurements with a probability that goes to one as n goes to infinity. This work shows strengthens this result by showing that a lower number of measurements, m = 2k log(n-k), is in fact sufficient for asymptotic recovery. Moreover, this number of measurements is also sufficient for detection of the sparsity pattern (support) of the vector with measurement errors provided the signal-to-noise ratio (SNR) scales to infinity. The scaling m = 2k log(n-k) exactly matches the number of measurements required by the more complex lasso for signal recovery.

Multiple object class learning and detection is a challenging problem due to the large number of object classes and their high visual variability. Specialized detectors usually excel in performance, while joint representations optimize sharing and reduce inference time --- but are complex to train. Conveniently, sequential learning of categories cuts down training time by transferring existing knowledge to novel classes, but cannot fully exploit the richness of shareability and might depend on ordering in learning. In hierarchical frameworks these issues have been little explored. In this paper, we show how different types of multi-class learning can be done within one generative hierarchical framework and provide a rigorous experimental analysis of various object class learning strategies as the number of classes grows. Specifically, we propose, evaluate and compare three important types of multi-class learning: 1.) independent training of individual categories, 2.) joint training of classes, 3.) sequential learning of classes. We explore and compare their computational behavior (space and time) and detection performance as a function of the number of learned classes on several recognition data sets.