We introduce a framework for unsupervised learning of structured predictors with overlapping, global features. Each input's latent representation is predicted conditional on the observable data using a feature-rich conditional random field. Then a reconstruction of the input is (re)generated, conditional on the latent structure, using models for which maximum likelihood estimation has a closed-form. Our autoencoder formulation enables efficient learning without making unrealistic independence assumptions or restricting the kinds of features that can be used. We illustrate insightful connections to traditional autoencoders, posterior regularization and multi-view learning. We show competitive results with instantiations of the model for two canonical NLP tasks: part-of-speech induction and bitext word alignment, and show that training our model can be substantially more efficient than comparable feature-rich baselines.
Structural support vector machines (SSVMs) are amongst the best performing models for structured computer vision tasks, such as semantic image segmentation or human pose estimation. Training SSVMs, however, is computationally costly, because it requires repeated calls to a structured prediction subroutine (called \emph{max-oracle}), which has to solve an optimization problem itself, e.g. a graph cut. In this work, we introduce a new algorithm for SSVM training that is more efficient than earlier techniques when the max-oracle is computationally expensive, as it is frequently the case in computer vision tasks. The main idea is to (i) combine the recent stochastic Block-Coordinate Frank-Wolfe algorithm with efficient hyperplane caching, and (ii) use an automatic selection rule for deciding whether to call the exact max-oracle or to rely on an approximate one based on the cached hyperplanes. We show experimentally that this strategy leads to faster convergence to the optimum with respect to the number of requires oracle calls, and that this translates into faster convergence with respect to the total runtime when the max-oracle is slow compared to the other steps of the algorithm. A publicly available C++ implementation is provided at http://pub.ist.ac.at/~vnk/papers/SVM.html .
We introduce a purely feed-forward architecture for semantic segmentation. We map small image elements (superpixels) to rich feature representations extracted from a sequence of nested regions of increasing extent. These regions are obtained by "zooming out" from the superpixel all the way to scene-level resolution. This approach exploits statistical structure in the image and in the label space without setting up explicit structured prediction mechanisms, and thus avoids complex and expensive inference. Instead superpixels are classified by a feedforward multilayer network. Our architecture achieves new state of the art performance in semantic segmentation, obtaining 64.4% average accuracy on the PASCAL VOC 2012 test set.
When evaluating computer vision systems, we are often concerned with performance on a task-specific evaluation measure such as the Intersection-Over-Union score used in the PASCAL VOC image segmentation challenge. Ideally, our systems would be tuned specifically to these evaluation measures. However, despite much work on loss-aware structured prediction, top performing systems do not use these techniques. In this work, we seek to address this problem, incorporating loss-aware prediction in a manner that is amenable to the approaches taken by top performing systems. Our main idea is to simultaneously leverage two systems: a highly tuned pipeline system as is found on top of leaderboards, and a traditional CRF. We show how to combine high quality candidate solutions from the pipeline with the probabilistic approach of the CRF that is amenable to loss-aware prediction. The result is that we can use loss-aware prediction methodology to improve performance of the highly tuned pipeline system.
We describe a model for multi-target tracking based on associating collections of candidate detections across frames of a video. In order to model pairwise interactions between different tracks, such as suppression of overlapping tracks and contextual cues about co-occurence of different objects, we augment a standard min-cost flow objective with quadratic terms between detection variables. We learn the parameters of this model using structured prediction and a loss function which approximates the multi-target tracking accuracy. We evaluate two different approaches to finding an optimal set of tracks under model objective based on an LP relaxation and a novel greedy extension to dynamic programming that handles pairwise interactions. We find the greedy algorithm achieves equivalent performance to the LP relaxation while being 2-7x faster than a commercial solver. The resulting model with learned parameters outperforms existing methods across several categories on the KITTI tracking benchmark.
We consider the class of optimization problems arising from computationally intensive L1-regularized M-estimators, where the function or gradient values are very expensive to compute. A particular instance of interest is the L1-regularized MLE for learning Conditional Random Fields (CRFs), which are a popular class of statistical models for varied structured prediction problems such as sequence labeling, alignment, and classification with label taxonomy. L1-regularized MLEs for CRFs are particularly expensive to optimize since computing the gradient values requires an expensive inference step. In this work, we propose the use of a carefully constructed proximal quasi-Newton algorithm for such computationally intensive M-estimation problems, where we employ an aggressive active set selection technique. In a key contribution of the paper, we show that the proximal quasi-Newton method is provably super-linearly convergent, even in the absence of strong convexity, by leveraging a restricted variant of strong convexity. In our experiments, the proposed algorithm converges considerably faster than current state-of-the-art on the problems of sequence labeling and hierarchical classification.
While there are many studies on weight regularization, the study on structure regularization is rare. Many existing systems on structured prediction focus on increasing the level of structural dependencies within the model. However, this trend could have been misdirected, because our study suggests that complex structures are actually harmful to generalization ability in structured prediction. To control structure-based overfitting, we propose a structure regularization framework via \emph{structure decomposition}, which decomposes training samples into mini-samples with simpler structures, deriving a model with better generalization power. We show both theoretically and empirically that structure regularization can effectively control overfitting risk and lead to better accuracy. As a by-product, the proposed method can also substantially accelerate the training speed. The method and the theoretical results can apply to general graphical models with arbitrary structures. Experiments on well-known tasks demonstrate that our method can easily beat the benchmark systems on those highly-competitive tasks, achieving state-of-the-art accuracies yet with substantially faster training speed.
Markov Chain Monte Carlo (MCMC) algorithms are often used for approximate inference inside learning, but their slow mixing can be difficult to diagnose and the approximations can seriously degrade learning. To alleviate these issues, we define a new model family using strong Doeblin Markov chains, whose mixing times can be precisely controlled by a parameter. We also develop an algorithm to learn such models, which involves maximizing the data likelihood under the induced stationary distribution of these chains. We show empirical improvements on two challenging inference tasks.
Linguistic structures exhibit a rich array of global phenomena, however commonly used Markov models are unable to adequately describe these phenomena due to their strong locality assumptions. We propose a novel hierarchical model for structured prediction over sequences and trees which exploits global context by conditioning each generation decision on an unbounded context of prior decisions. This builds on the success of Markov models but without imposing a fixed bound in order to better represent global phenomena. To facilitate learning of this large and unbounded model, we use a hierarchical Pitman-Yor process prior which provides a recursive form of smoothing. We propose prediction algorithms based on A* and Markov Chain Monte Carlo sampling. Empirical results demonstrate the potential of our model compared to baseline finite-context Markov models on part-of-speech tagging and syntactic parsing.
Stochastic binary hidden units in a multi-layer perceptron (MLP) network give at least three potential benefits when compared to deterministic MLP networks. (1) They allow to learn one-to-many type of mappings. (2) They can be used in structured prediction problems, where modeling the internal structure of the output is important. (3) Stochasticity has been shown to be an excellent regularizer, which makes generalization performance potentially better in general. However, training stochastic networks is considerably more difficult. We study training using M samples of hidden activations per input. We show that the case M=1 leads to a fundamentally different behavior where the network tries to avoid stochasticity. We propose two new estimators for the training gradient and propose benchmark tests for comparing training algorithms. Our experiments confirm that training stochastic networks is difficult and show that the proposed two estimators perform favorably among all the five known estimators.
We address the problem of \emph{quantification}, a supervised learning task whose goal is, given a class, to estimate the relative frequency (or \emph{prevalence}) of the class in a dataset of unlabelled items. Quantification has several applications in data and text mining, such as estimating the prevalence of positive reviews in a set of reviews of a given product, or estimating the prevalence of a given support issue in a dataset of transcripts of phone calls to tech support. So far, quantification has been addressed by learning a general-purpose classifier, counting the unlabelled items which have been assigned the class, and tuning the obtained counts according to some heuristics. In this paper we depart from the tradition of using general-purpose classifiers, and use instead a supervised learning model for \emph{structured prediction}, capable of generating classifiers directly optimized for the (multivariate and non-linear) function used for evaluating quantification accuracy. The experiments that we have run on 5500 binary high-dimensional datasets (averaging more than 14,000 documents each) show that this method is more accurate, more stable, and more efficient than existing, state-of-the-art quantification methods.
Methods for learning to search for structured prediction typically imitate a reference policy, with existing theoretical guarantees demonstrating low regret compared to that reference. This is unsatisfactory in many applications where the reference policy is suboptimal and the goal of learning is to improve upon it. Can learning to search work even when the reference is poor? We provide a new learning to search algorithm, LOLS, which does well relative to the reference policy, but additionally guarantees low regret compared to deviations from the learned policy: a local-optimality guarantee. Consequently, LOLS can improve upon the reference policy, unlike previous algorithms. This enables us to develop structured contextual bandits, a partial information structured prediction setting with many potential applications.
In this paper, we address the problem of multi-label classification. We consider linear classifiers and propose to learn a prior over the space of labels to directly leverage the performance of such methods. This prior takes the form of a quadratic function of the labels and permits to encode both attractive and repulsive relations between labels. We cast this problem as a structured prediction one aiming at optimizing either the accuracies of the predictors or the F 1-score. This leads to an optimization problem closely related to the max-cut problem, which naturally leads to semidefinite and spectral relaxations. We show on standard datasets how such a general prior can improve the performances of multi-label techniques.
We address the task of annotating images with semantic tuples. Solving this problem requires an algorithm which is able to deal with hundreds of classes for each argument of the tuple. In such contexts, data sparsity becomes a key challenge, as there will be a large number of classes for which only a few examples are available. We propose handling this by incorporating feature representations of both the inputs (images) and outputs (argument classes) into a factorized log-linear model, and exploiting the flexibility of scoring functions based on bilinear forms. Experiments show that integrating feature representations of the outputs in the structured prediction model leads to better overall predictions. We also conclude that the best output representation is specific for each type of argument.
Deep structured output learning shows great promise in tasks like semantic image segmentation. We proffer a new, efficient deep structured model learning scheme, in which we show how deep Convolutional Neural Networks (CNNs) can be used to estimate the messages in message passing inference for structured prediction with Conditional Random Fields (CRFs). With such CNN message estimators, we obviate the need to learn or evaluate potential functions for message calculation. This confers significant efficiency for learning, since otherwise when performing structured learning for a CRF with CNN potentials it is necessary to undertake expensive inference for every stochastic gradient iteration. The network output dimension for message estimation is the same as the number of classes, in contrast to the network output for general CNN potential functions in CRFs, which is exponential in the order of the potentials. Hence CNN message learning has fewer network parameters and is more scalable for cases that a large number of classes are involved. We apply our method to semantic image segmentation on the PASCAL VOC 2012 dataset. We achieve an intersection-over-union score of 73.4 on its test set, which is the best reported result for methods using the VOC training images alone. This impressive performance demonstrates the effectiveness and usefulness of our CNN message learning method.
Color constancy is the problem of inferring the color of the light that illuminated a scene, usually so that the illumination color can be removed. Because this problem is underconstrained, it is often solved by modeling the statistical regularities of the colors of natural objects and illumination. In contrast, in this paper we reformulate the problem of color constancy as a 2D spatial localization task in a log-chrominance space, thereby allowing us to apply techniques from object detection and structured prediction to the color constancy problem. By directly learning how to discriminate between correctly white-balanced images and poorly white-balanced images, our model is able to improve performance on standard benchmarks by nearly 40%.
There is a large variety of objects and appliances in human environments, such as stoves, coffee dispensers, juice extractors, and so on. It is challenging for a roboticist to program a robot for each of these object types and for each of their instantiations. In this work, we present a novel approach to manipulation planning based on the idea that many household objects share similarly-operated object parts. We formulate the manipulation planning as a structured prediction problem and design a deep learning model that can handle large noise in the manipulation demonstrations and learns features from three different modalities: point-clouds, language and trajectory. In order to collect a large number of manipulation demonstrations for different objects, we developed a new crowd-sourcing platform called Robobarista. We test our model on our dataset consisting of 116 objects with 249 parts along with 250 language instructions, for which there are 1225 crowd-sourced manipulation demonstrations. We further show that our robot can even manipulate objects it has never seen before.
IllinoisSL is a Java library for learning structured prediction models. It supports structured Support Vector Machines and structured Perceptron. The library consists of a core learning module and several applications, which can be executed from command-lines. Documentation is provided to guide users. In Comparison to other structured learning libraries, IllinoisSL is efficient, general, and easy to use.
Algorithm designers typically assume that the input data is correct, and then proceed to find "optimal" or "sub-optimal" solutions using this input data. However this assumption of correct data does not always hold in practice, especially in the context of online learning systems where the objective is to learn appropriate feature weights given some training samples. Such scenarios necessitate the study of inverse optimization problems where one is given an input instance as well as a desired output and the task is to adjust the input data so that the given output is indeed optimal. Motivated by learning structured prediction models, in this paper we consider inverse optimization with a margin, i.e., we require the given output to be better than all other feasible outputs by a desired margin. We consider such inverse optimization problems for maximum weight matroid basis, matroid intersection, perfect matchings, minimum cost maximum flows, and shortest paths and derive the first known results for such problems with a non-zero margin. The effectiveness of these algorithmic approaches to online learning for structured prediction is also discussed.
Decision trees and randomized forests are widely used in computer vision and machine learning. Standard algorithms for decision tree induction optimize the split functions one node at a time according to some splitting criteria. This greedy procedure often leads to suboptimal trees. In this paper, we present an algorithm for optimizing the split functions at all levels of the tree jointly with the leaf parameters, based on a global objective. We show that the problem of finding optimal linear-combination (oblique) splits for decision trees is related to structured prediction with latent variables, and we formulate a convex-concave upper bound on the tree's empirical loss. The run-time of computing the gradient of the proposed surrogate objective with respect to each training exemplar is quadratic in the the tree depth, and thus training deep trees is feasible. The use of stochastic gradient descent for optimization enables effective training with large datasets. Experiments on several classification benchmarks demonstrate that the resulting non-greedy decision trees outperform greedy decision tree baselines.