A critical practical problem in the field of object recognition is an insufficient number of labeled training images, as manually labeling images is a time consuming task. For this reason, unsupervised learning techniques are used to take advantage of unlabeled training images to extract image representations that are useful for classification. However, unsupervised learning is in general difficult. We propose simplifying the unsupervised training problem considerably by taking the advance of motion information. The output of our method is a model that can generate a vector representation from any static image. However, the model is trained using images with additional motion information. To demonstrate the flobject analysis framework, we extend the latent Dirichlet allocation model to account for word-specific flow vectors. We show that the static image representations extracted using our model achieve higher classification rates and better generalization than standard topic models, spatial pyramid matching, and Gist descriptors.

This thesis explores the use of physics based human motion models in the context of video-based human motion estimation and scene analysis. Two abstract models of human locomotion are described and used as the basis for video-based estimation. These models demonstrate the power of physics based models to provide meaningful cues for estimation without the use of motion capture data. However promising, the abstract nature of these models limit the range of motion they can faithfully capture. A more detailed model of human motion and ground interaction is also described. This model is used to estimate the ground surface which a subject interacts with, the forces driving the motion and, finally, to smooth corrupted motions from existing trackers in a physically realistic fashion. This thesis suggests that one of the key difficulties in using physical models is the discontinuous nature of contact and collisions. Two different approaches to handling ground contacts are demonstrated,one using explicit detection and collision resolution and the other using a continuous approximation. This difficulty also distinguishes the models used here from others used in other areas which often sidestep the issue of collisions.

This thesis introduces principled, broadly applicable, and efficient patch-based models for data processing applications. Recently, "epitomes" were introduced as patch-based probability models that are learned by compiling together a large number of examples of patches from input images. This thesis describes how epitomes can be used to model video data and a significant computational speedup is introduced that can be incorporated into the epitome inference and learning algorithm. In the case of videos, epitomes are estimated so as to model most of the small space-time cubes from the input data. Then, the epitome can be used for various modelling and reconstruction tasks, of which we show results for video super-resolution, video interpolation, and object removal. Besides computational efficiency, an interesting advantage of the epitome as a representation is that it can be reliably estimated even from videos with large amounts of missing data. This ability is illustrated on the task of reconstructing the dropped frames in a video broadcast using only the degraded video. Further, a new patch-based model is introduced, that when applied to epitomes, accounts for the varying geometric configurations of object features. The power of this model is illustrated on tasks such as multiple object registration and detection and missing data interpolation, including a difficult task of photograph relighting.

A critical practical problem in the field of object recognition is an insufficient number of labeled training images, as manually labeling images is a time consuming task. For this reason, unsupervised learning techniques are used to take advantage of unlabeled training images to extract image representations that are useful for classification. However, unsupervised learning is in general difficult. We propose simplifying the unsupervised training problem considerably by taking the advance of motion information. The output of our method is a model that can generate a vector representation from any static image. However, the model is trained using images with additional motion information. To demonstrate the flobject analysis framework, we extend the latent Dirichlet allocation model to account for word-specific flow vectors. We show that the static image representations extracted using our model achieve higher classification rates and better generalization than standard topic models, spatial pyramid matching, and Gist descriptors.

Human face detection and recognition techniques play an important role in applica- tions like face recognition, video surveillance, human computer interface and face image databases. Using color information in images is one of the various possible techniques used for face detection. The novel technique used in this project was the combination of various techniques such as skin color detection, template matching, gradient face de- tection to achieve high accuracy of face detection in frontal faces. The objective in this work was to determine the best rotation angle to achieve optimal detection. Also eye and mouse template matching have been put to test for feature detection.

Feature extraction is an important step in the classification of high-dimensional data such as face images. Furthermore, linear feature extractors are more prevalent due to computational efficiency and preservation of the Gaussianity. This research proposes a simple and fast linear feature extractor approximating the sufficient statistic for Gaussian distributions. This method preserves the discriminatory information in both first and second moments of the data and yields the linear discriminant analysis as a special case. Additionally, an accurate upper bound on the error probability of a plug-in classifier can be used to approximate the number of features minimizing the error probability. Therefore, tighter error bounds are derived in this work based on the Bayes error or the classification error on the trained distributions. These bounds can also be used for performance guarantee and to determine the required number of training samples to guarantee approaching the Bayes classifier performance.

Master of Science / Department of Computing and Information Sciences / William H. Hsu / This thesis discusses the background and methodologies necessary for constructing features in order to discover hidden links in relational data. Specifically, we consider the problems of predicting, classifying and annotating friends relations in friends networks, based upon features constructed from network structure and user profile data. I first document a data model for the blog service LiveJournal, and define a set of machine learning problems such as predicting existing links and estimating inter-pair distance. Next, I explain how the problem of classifying a user pair in a social networks, as directly connected or not, poses the problem of selecting and constructing relevant features. In order to construct these features, a genetic programming approach is used to construct multiple symbol trees with base features as their leaves; in this manner, the genetic program selects and constructs features that many not have been considered, but possess better predictive properties than the base features. In order to extract certain graph features from the relatively large social network, a new shortest path search algorithm is presented which computes and operates on a Euclidean embedding of the network. Finally, I present classification results and discuss the properties of the frequently constructed features in order to gain insight on hidden relations that exists in this domain.

Doctor of Philosophy / Department of Computing and Information Sciences / Scott A. DeLoach / There is a growing demand for more effective integration of humans and computing systems, specifically in multiagent and multirobot systems. There are two aspects to consider in human integration: (1) the ability to control an arbitrary number of robots (particularly heterogeneous robots) and (2) integrating humans as peers in computing systems instead of being just users or supervisors. With traditional supervisory control of multirobot systems, the number of robots that a human can manage effectively is between four and six [17]. A limitation of traditional supervisory control is that the human must interact individually with each robot, which limits the upper-bound on the number of robots that a human can control effectively. In this work, I define the concept of "organizational control" together with an autonomous mechanism that can perform task allocation and other low-level housekeeping duties, which significantly reduces the need for the human to interact with individual robots. Humans are very versatile and robust in the types of tasks they can accomplish. However, failures in computing systems are common and thus redundancies are included to mitigate the chance of failure. When all redundancies have failed, system failure will occur and the computing system will be unable to accomplish its tasks. One way to further reduce the chance of a system failure is to integrate humans as peer "agents" in the computing system. As part of the system, humans can be assigned tasks that would have been impossible to complete due to failures.

With the increasing availability of large datasets machine learning techniques are becoming an increasingly attractive alternative to expert-designed approaches to solving complex problems in domains where data is abundant. In this thesis we introduce several models for large sparse discrete datasets. Our approach, which is based on probabilistic models that use distributed representations to alleviate the effects of data sparsity, is applied to statistical language modelling and collaborative filtering. We introduce three probabilistic language models that represent words using learned real-valued vectors. Two of the models are based on the Restricted Boltzmann Machine (RBM) architecture while the third one is a simple deterministic model. We show that the deterministic model outperforms the widely used n-gram models and learns sensible word representations. To reduce the time complexity of training and making predictions with the deterministic model, we introduce a hierarchical version of the model, that can be exponentially faster. The speedup is achieved by structuring the vocabulary as a tree over words and taking advantage of this structure. We propose a simple feature-based algorithm for automatic construction of trees over words from data and show that the resulting models can outperform non-hierarchical neural models as well as the best n-gram models. We then turn our attention to collaborative filtering and show how RBM models can be used to model the distribution of sparse high-dimensional user rating vectors efficiently, presenting inference and learning algorithms that scale linearly in the number of observed ratings. We also introduce the Probabilistic Matrix Factorization model which is based on the probabilistic formulation of the low-rank matrix approximation problem for partially observed matrices. The two models are then extended to allow conditioning on the identities of the rated items whether or not the actual rating values are known. Our results on the Netflix Prize dataset show that both RBM and PMF models outperform online SVD models.

Gesture Recognition provides an efficient human-computer interaction for interactive and intelligent computing. In this work, we address the problem of gesture recognition using the theory of random projection and by formulating the recognition problem as an $\ell_1$-minimization problem. The gesture recognition uses a single 3-axis accelerometer for data acquisition and comprises two main stages: a training stage and a testing stage. For training, the system employs dynamic time warping as well as affinity propagation to create exemplars for each gesture while for testing, the system projects all candidate traces and also the unknown trace onto the same lower dimensional subspace for recognition. A dictionary of 18 gestures is defined and a database of over 3,700 traces is created from 7 subjects on which the system is tested and evaluated. Simulation results reveal a superior performance, in terms of accuracy and computational complexity, compared to other systems in the literature.