Center for Human Modeling and Simulation

An End-to-End Discriminative Approach to Machine Translation

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:45 PDT

We present a perceptron-style discriminative approach to machine translation in which large feature sets can be exploited. Unlike discriminative reranking approaches, our system can take advantage of learned features in all stages of decoding. We first discuss several challenges to error-driven discriminative approaches. In particular, we explore different ways of updating parameters given a training example. We find that making frequent but smaller updates is preferable to making fewer but larger updates. Then, we discuss an array of features and show both how they quantitatively increase BLEU score and how they qualitatively interact on specific examples. One particular feature we investigate is a novel way to introduce learning into the initial phrase extraction process, which has previously been entirely heuristic.

Sparsity in Dependency Grammar Induction

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:42 PDT

A strong inductive bias is essential in unsupervised grammar induction. We explore a particular sparsity bias in dependency grammars that encourages a small number of unique dependency types. Specifically, we investigate sparsity-inducing penalties on the posterior distributions of parent-child POS tag pairs in the posterior regularization (PR) framework of Graça et al. (2007). In experiments with 12 languages, we achieve substantial gains over the standard expectation maximization (EM) baseline, with average improvement in attachment accuracy of 6.3%. Further, our method outperforms models based on a standard Bayesian sparsity-inducing prior by an average of 4.9%. On English in particular, we show that our approach improves on several other state-of-the-art techniques.

Learning Sparse Markov Network Structure via Ensemble-of-Trees Model

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:40 PDT

Learning the sparse structure of a general Markov network is a hard computational problem. One of the main difficulties is the computation of the generally intractable partition function. To circumvent this difficulty, we propose to learn the network structure using an ensemble-of- trees (ET) model. The ET model was first introduced by Meil˘a and Jaakkola (2006), and it represents a multivariate distribution using a mixture of all possible spanning trees. The advantage of the ET model is that, although it needs to sum over super exponentially many trees, its partition function as well as data likelihood can be computed in a closed form. Furthermore, because the ET model tends to represent a Markov network using as small number of trees as possible, it provides a natural regularization for finding a sparse network structure. Our simulation results show that the proposed ET approach is able to accurately recover the true Markov network connectivity and outperform the state-of-art approaches for both discrete and continuous random variable networks when a small number of data samples is available. Furthermore, we also demonstrate the usage of the ET model for discovering the network of words from blog posts.

Generative-Discriminitive Basis Learning for Medical Imaging

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:33 PDT

This paper presents a novel dimensionality reduction method for classification in medical imaging. The goal is to transform very high-dimensional input (typically, millions of voxels) to a low-dimensional representation (small number of constructed features) that preserves discriminative signal and is clinically interpretable. We formulate the task as a constrained optimization problem that combines generative and discriminative objectives and show how to extend it to the semisupervised learning (SSL) setting. We propose a novel largescale algorithm to solve the resulting optimization problem. In the fully supervised case, we demonstrate accuracy rates that are better than or comparable to state-of-the-art algorithms on several datasets while producing a representation of the group difference that is consistent with prior clinical reports. Effectiveness of the proposed algorithm for SSL is evaluated with both benchmark and medical imaging datasets. In the benchmark datasets, the results are better than or comparable to the state-of-the-art methods for SSL. For evaluation of the SSL setting in medical datasets, we use images of subjects with Mild Cognitive Impairment (MCI), which is believed to be a precursor to Alzheimer’s disease (AD), as unlabeled data. AD subjects and Normal Control (NC) subjects are used as labeled data, and we try to predict conversion from MCI to AD on follow-up. The semi-supervised extension of this method not only improves the generalization accuracy for the labeled data (AD/NC) slightly but is also able to predict subjects which are likely to converge to AD.

Learning Determinantal Point Processes

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:30 PDT

Determinantal point processes (DPPs), which arise in random matrix theory and quantum physics, are natural models for subset selection problems where diversity is preferred. Among many remarkable properties, DPPs other tractable algorithms for exact inference, including computing marginal probabilities and sampling; how- ever, an important open question has been how to learn a DPP from labeled training data. In this paper we propose a natural feature-based parameterization of conditional DPPs, and show how it leads to a convex and efficient learning formulation. We analyze the relationship between our model and binary Markov random fields with repulsive potentials, which are qualitatively similar but computationally intractable. Finally, we apply our approach to the task of extractive summarization, where the goal is to choose a small subset of sentences conveying the most important information from a set of documents. In this task there is a fundamental tradeoff between sentences that are highly relevant to the collection as a whole, and sentences that are diverse and not repetitive. Our parameterization allows us to naturally balance these two characteristics. We evaluate our system on data from the DUC 2003/04 multi- document summarization task, achieving state-of-the-art results.

Learning Sparse Markov Network Structure via Ensemble-of-Trees Models

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:20 PDT

Learning the sparse structure of a general Markov network is a hard computational problem. One of the main difficulties is the computation of the generally intractable partition function. To circumvent this difficulty, we propose to learn the network structure using an ensemble-of- trees (ET) model. The ET model was first introduced by Meil˘a and Jaakkola (2006), and it represents a multivariate distribution using a mixture of all possible spanning trees. The advantage of the ET model is that, although it needs to sum over super-exponentially many trees, its partition function as well as data likelihood can be computed in a closed form. Furthermore, because the ET model tends to represent a Markov network using as small number of trees as possible, it provides a natural regularization for finding a sparse network structure. Our simulation results show that the proposed ET approach is able to accurately recover the true Markov network connectivity and outperform the state-of-art approaches for both discrete and continuous random variable network swhen a small number of data samples is available. Furthermore, we also demonstrate the usage of the ET model for discovering the network of words from blog posts.

A permutation-augmented sampler for DP mixture models

Ben Taskar et al. — Wed, 11 Jul 2012 08:42:13 PDT

We introduce a new inference algorithm for Dirichlet process mixture models. While Gibbs sampling and variational methods focus on local moves, the new algorithm makes more global moves. This is done by introducing a permutation of the data points as an auxiliary variable. The algorithm is a blocked sampler which alternates between sampling the clustering and sampling the permutation. The key to the efficiency of this approach is that it is possible to use dynamic programming to consider all exponentially many clusterings consistent with a given permutation. We also show that random projections can be used to effectively sample the permutation. The result is a stochastic hill-climbing algorithm that yields burn-in times significantly smaller than those of collapsed Gibbs sampling.

Discrimintive Image Warping with Attribute Flow

Jianbo Shi et al. — Wed, 11 Jul 2012 08:42:04 PDT

We address the problem of finding deformation between two images for the purpose of recognizing objects. The challenge is that discriminative features are often transformation-variant (e.g. histogram of oriented gradients, texture), while transformation-invariant features (e.g. intensity, color) are often not discriminative. We introduce the concept of attribute flow which explicitly models how image attributes vary with its deformation. We develop a non-parametric method to approximate this using histogram matching, which can be solved efficiently using linear programming. Our method produces dense correspondence between images, and utilizes discriminative, transformation-variant features for simultaneous detection and alignment. Experiments on ETHZ shape categories dataset show that we can accurately recognize highly deformable objects with few training examples.

Perceptually Realistic Behavior through Alibi Generation

Ben Sunshine-Hill et al. — Tue, 09 Nov 2010 10:20:00 PST

Real-time pedestrian simulation for open-world games involves aggressive behavior simpliﬁcation and culling to keep computational cost under control, but it is diffﬁcult to predict whether these techniques will become unrealistic in certain situations. We propose a method of perceptually simulating highly realistic pedestrian behavior in virtual cities in real- time. Designers build a highly realistic simulation, from which a perceptually identical “perceptual simulation” is generated. Although the perceptual simulation simulates only a small portion of the world at a time, and does so with inexpensive approximations, it can be statistically guaranteed that the results are perceptually indistinguishable from those of the original simulation.

Empirical developments in retraction

B K. Redman et al. — Thu, 03 Sep 2009 11:24:59 PDT

This study provides current data on key questions about retraction of scientific articles. Findings confirm that the rate of retractions remains low but is increasing. The most commonly cited reason for retraction was research error or inability to reproduce results; the rate from research misconduct is an underestimate, since some retractions necessitated by research misconduct were reported as being due to inability to reproduce. Retraction by parties other than authors is increasing, especially for research misconduct. Although retractions are on average occurring sooner after publication than in the past, citation analysis shows that they are not being recognised by subsequent users of the work. Findings suggest that editors and institutional officials are taking more responsibility for correcting the scientific record but that reasons published in the retraction notice are not always reliable. More aggressive means of notification to the scientific community appear to be necessary.

Real-time Inverse Kinematics Techniques for Anthropomorphic Limbs

Deepak Tolani et al. — Wed, 10 Dec 2008 07:09:17 PST

In this paper we develop a set of inverse kinematics algorithms suitable for an anthropomorphic arm or leg. We use a combination of analytical and numerical methods to solve generalized inverse kinematics problems including position, orientation, and aiming constraints. Our combination of analytical and numerical methods results in faster and more reliable algorithms than conventional inverse Jacobian and optimization-based techniques. Additionally, unlike conventional numerical algorithms, our methods allow the user to interactively explore all possible solutions using an intuitive set of parameters that define the redundancy of the system.

Animation Control for Real-Time Virtual Humans

Norman I. Badler et al. — Fri, 01 Aug 2008 09:14:29 PDT

The computation speed and control methods needed to portray 3D virtual humans suitable for interactive applications have improved dramatically in recent years. Real-time virtual humans show increasingly complex features along the dimensions of appearance, function, time, autonomy, and individuality. The virtual human architecture we’ve been developing at the University of Pennsylvania is representative of an emerging generation of such architectures and includes low-level motor skills, a mid-level parallel automata controller, and a high-level conceptual representation for driving virtual humans through complex tasks. The architecture—called Jack— provides a level of abstraction generic enough to encompass natural-language instruction representation as well as direct links from those instructions to animation control.

Multi-Level Shape Representation Using Global Deformations and Locally Adaptive Finite Elements

Dimitris Metaxas et al. — Mon, 19 May 2008 08:14:04 PDT

We present a model-based method for the multi-level shape, pose estimation and abstraction of an object’s surface from range data. The surface shape is estimated based on the parameters of a superquadric that is subjected to global deformations (tapering and bending) and a varying number of levels of local deformations. Local deformations are implemented using locally adaptive finite elements whose shape functions are piecewise cubic functions with C1 continuity. The surface pose is estimated based on the model's translational and rotational degrees of freedom. The algorithm first does a coarse fit, solving for a first approximation to the translation, rotation and global deformation parameters and then does several passes of mesh refinement, by locally subdividing triangles based on the distance between the given datapoints and the model. The adaptive finite element algorithm ensures that during subdivision the desirable finite element mesh generation properties of conformity, non-degeneracy and smoothness are maintained. Each pass of the algorithm uses physics-based modeling techniques to iteratively adjust the global and local parameters of the model in response to forces that are computed from approximation errors between the model and the data. We present results demonstrating the multi-level shape representation for both sparse and dense range data.

A Machine Translation System from English to American Sign Language

Liwei Zhao et al. — Mon, 19 May 2008 08:14:02 PDT

Research in computational linguistics, computer graphics and autonomous agents has led to the development of increasingly sophisticated communicative agents over the past few years, bringing new perspective to machine translation research. The engineering of language-based smooth, expressive, natural-looking human gestures can give us useful insights into the design principles that have evolved in natural communication between people. In this paper we prototype a machine translation system from English to American Sign Language (ASL), taking into account not only linguistic but also visual and spatial information associated with ASL signs.

Human Task Animation from Performance Models and Natural Language Input

Jeffrey Esakov et al. — Fri, 14 Sep 2007 08:31:10 PDT

Graphical manipulation of human figures is essential for certain types of human factors analyses such as reach, clearance, fit, and view. In many situations, however, the animation of simulated people performing various tasks may be based on more complicated functions involving multiple simultaneous reaches, critical timing, resource availability, and human performance capabilities. One rather effective means for creating such a simulation is through a natural language description of the tasks to be carried out. Given an anthropometrically-sized figure and a geometric workplace environment, various simple actions such as reach, turn, and view can be effectively controlled from language commands or standard NASA checklist procedures. The commands may also be generated by external simulation tools. Task timing is determined from actual performance models, if available, such as strength models or Fitts' Law. The resulting action specifications are animated on a Silicon Graphics Iris workstation in real-time.

Terrain Navigation Skills and Reasoning

Hyeongseok Ko et al. — Wed, 12 Sep 2007 08:46:07 PDT

We describe a real-time model of terrain traversal by simulated human agents. Agent navigation includes a variety of simulated sensors, terrain reasoning with behavioral constraints, and detailed simulation of a variety of locomotion techniques. Our Kinematic Locomotion Generation Module (KLOG) generates various terrain navigation skills as well as both rhythmic and non-rhythmic variations of these skills. The terrain navigation skills include curved path walking, lateral or backward stepping, running, and the transitions between walking and running for motion continuity. Locomotion attributes such as pelvis rotation and translation and torso flexion and twist are used to modify the KLOG skills so that realistic looking rhythmic locomotion or non-rhythmic variations, such as ducking under a low hanging branch of a tree, can be achieved. The path through the terrain is incrementally computed by a behavioral reasoning system configuring a behavioral feedback network. A number of sensors acquire information on object range, passageways, obstacles, terrain type, exposure to hostile agents and so on. The behavioral reasoner weighs this information along with collision avoidance, cost, danger minimization, locomotion types and other behaviors available to the agent and incrementally attempts to reach a goal location. Since the system is reactive, it can respond to moving obstacles, changing terrain, or unexpected events due to hostile agents or the effects of limited perception.

Efficient Re-rendering of Naturally Illuminated Environments

Jeffry S. Nimeroff et al. — Wed, 12 Sep 2007 07:53:59 PDT

We present a method for the efficient re-rendering of a scene under a directional illuminant at an arbitrary orientation. We take advantage of the linearity of the rendering operator with respect to illumination for a fixed scene and camera geometry. Re-rendering is accomplished via linear combination of a set of pre-rendered "basis" images. The theory of steerable functions provides the machinery to derive an appropriate set of basis images. We demonstrate the technique on both simple and complex scenes illuminated by an approximation to natural skylight. We show re-rendering simulations under conditions of varying sun position and cloudiness.

Model-based Analysis of Cardiac Motion from Tagged MRI Data

Jinah Park et al. — Wed, 12 Sep 2007 07:34:11 PDT

We develop a new method for analyzing the motion of the left ventricle (LV) of a heart from tagged MRI data. Our technique is based on the development of a new class of physics-based deformable models whose parameters are functions allowing the definition of new parameterized primitives and parameterized deformations. These parameter functions improve the accuracy of shape description through the use of a few intuitive parameters such as functional twisting. Furthermore, these parameters require no complex post-processing in order to be used by a physician. Using a physics-based approach, we convert these geometric models into deformable models that deform due to forces exerted from the datapoints and conform to the given dataset. We present experiments involving the extraction of shape and motion of the LV from MRI-SPAMM data based on a few parameter functions. Furthermore, by plotting the variations over time of the extracted model parameters from normal and abnormal heart data we are able to characterize quantitatively their differences.

A Framework for Global Illumination in Animated Environments

Jeffry S. Nimeroff et al. — Wed, 12 Sep 2007 06:12:07 PDT

We describe a new framework for efficiently computing and storing global illumination effects for complex, animated environments. The new framework allows the rapid generation of sequences representing any arbitrary path in a "view space" within an environment in which both the viewer and objects move. The global illumination is stored as time sequences of range images at base locations that span the view space. We present algorithms for determining locations for these base images, and the time steps required to adequately capture the effects of object motion. We also present algorithms for computing the global illumination in the base images that exploit spatial and temporal coherence by considering direct and indirect illumination separately. We discuss an initial implementation using the new framework. Results from our implementation demonstrate the efficient generation of multiple tours through a complex space, and a tour of an environment in which objects move.

Simulation and analysis of complex human tasks

Norman I. Badler et al. — Tue, 11 Sep 2007 12:56:23 PDT

We discuss how the combination of a realistic human figure with a high-level behavioral control interface allow the construction of detailed simulations of humans performing manual tasks from which inferences about human performance requirements can be made. The Jack human modeling environment facilitates the real-time simulation of humans performing sequences of tasks such as walking, lifting, reaching, and grasping in a complex simulated environment. Analysis capabilities include strength, reachability, and visibility; moreover results from these tests can affect an unfolding simulation.