Similarity search is a task fundamental to many machine learning and data analytics applications, where distance metric learning plays an important role. However, since modern online applications continuously produce objects with new characteristics which tend to change over time, state-of-the-art similarity search using distance metric learning methods tends to fail when deployed in such applications without taking the change into consideration. In this work, we propose a Distance Metric Learning-based Continuous Similarity Search approach (CSS for short) to account for the dynamic nature of such data. CSS system adopts an online metric learning model to achieve distance metric evolving to adapt the dynamic nature of continuous data without large latency. To improve the accuracy of online metric learning model, a compact labeled dataset which is representative of the updated data is dynamically updated. Also, to accelerate similarity search, CSS includes an online maintained Locality Sensitive Hashing index to accelerate the similarity search. One, our labeled data update strategy progressively enriches the labeled data to assure continued representativeness, yet without excessively growing its size to ensure that the computation costs of metric learning remain bounded. Two, our continuous distance metric learning strategy ensures that each update only requires one linear time k-NN search in contrast to the cubic time complexity of relearning the distance metric from scratch. Three, our LSH update mechanism leverages our theoretical insight that the LSH built based on the original distance metric is equally effective in supporting similarity search using the new distance metric as long as the transform matrix learned for the new distance metric is reversible. This important observation empowers CSS to avoid the modification of LSH in most cases. Our experimental study using real-world public datasets and large synthetic datasets confirms the effectiveness of CSS in improving the accuracy of classification and information retrieval tasks. Also, CSS achieves 3 orders of magnitude speedup of our incremental distance metric learning strategy (and its three underlying components) over the state-of-art methods.

Person re-identification (Re-ID) has become an intense research area in recent years. The main goal of this topic is to recognize and match individuals over time at the same or different locations. This task is challenging due to the variation of illumination, viewpoints, pedestrians’ appearance and partial occlusion. Previous works mainly focus on finding robust features and metric learning. Many metric learning methods convert the Re-ID problem to a matrix decomposition problem by Fisher discriminant analysis (FDA). Mahalanobis distance metric learning is a popular method to measure similarity; however, since directly extracted descriptors usually have high dimensionality, it’s intractable to learn a high-dimensional semi-positive definite (SPD) matrix. Dimensionality reduction is used to project high-dimensional descriptors to a lower-dimensional space while preserving those discriminative information. In this paper, the kernel Fisher discriminant analysis (KLFDA) [38] is used to reduce dimensionality given that kernelization method can greatly improve Re-ID performance for nonlinearity. Inspired by [47], an SPD matrix is then learned on lower-dimensional descriptors based on the limitation that the maximum intraclass distance is at least one unit smaller than the minimum interclass distance. This method is proved to have excellent performance compared with other advanced metric learning.

A good distance metric can improve the accuracy of a nearest neighbour classifier. Xing et al. (2002) proposed distance metric learning to find a linear transformation of the data so that observations of different classes are better separated. For high-dimensional problems where many un-informative variables are present, it is attractive to select a sparse distance metric, both to increase predictive accuracy but also to aid interpretation of the result. In this thesis, we investigate three different types of sparsity assumption for distance metric learning and show that sparse recovery is possible under each type of sparsity assumption with an appropriate choice of L1-type penalty. We show that a lasso penalty promotes learning a transformation matrix having lots of zero entries, a group lasso penalty recovers a transformation matrix having zero rows/columns and a trace norm penalty allows us to learn a low rank transformation matrix. The regularization allows us to consider a large number of covariates and we apply the technique to an expanded set of basis called rule ensemble to allow for a more flexible fit. Finally, we illustrate an application of the metric learning problem via a document retrieval example and discuss how similarity-based information can be applied to learn a classifier.

There are a number of clinical options available for speech-language pathologists to choose from to analyze a child's phonological system, select treatment targets, and design intervention. Frequently, each of these areas of clinical options is viewed independently of one another or approached within an eclectic framework. In this article, an integrated and systemic approach is presented which assumes that a dynamic interaction exists among assessment, target selection, and intervention. Systemic Phonological Assessment of Child Speech, the distance metric approach to target selection, and the multiple oppositions treatment approach are described, with examples provided for each component. Finally, a case study is presented that examines the systemic approach of multiple oppositions relative to the approach of minimal pairs.

As digital images become ubiquitous in many applications, the need for efficient and effective retrieval techniques is more demanding than ever. Query by Example (QBE) and Query by Concept (QBC) are among the most popular query models. The former model accepts example images as queries and searches for similar ones based on low-level features such as colors and textures. The latter model allows queries to be expressed in the form of high-level semantics or concept words, such as "boat" or "car," and finds images that match the specified concepts. Recent research has focused on the connections between these two models and attempts to close the semantic-gap between them. This research involves finding the best method that maps a set of low-level features into high-level concepts. Automatic annotation techniques are investigated in this dissertation to facilitate QBC. In this approach, sets of training images are used to discover the relationship between low-level features and predetermined high-level concepts. The best mapping with respect to the training sets is proposed and used to analyze images, annotating them with the matched concept words. One principal difference between QBE and QBC is that, while similarity matching in QBE must be done at the query time, QBC performs concept exploration off-line. This difference allows QBC techniques to shift the time-consuming task of determining similarity away from the query time, thus facilitating the additional processing time required for increasingly accurate matching. Consequently, QBC's primary design objective is to achieve accurate annotation within a reasonable processing time. This objective is the guiding principle in the design of the following proposed methods which facilitate image annotation: 1.A novel dynamic similarity function. This technique allows users to query with multiple examples: relevant, irrelevant or neutral. It uses the range distance in each group to automatically determine weights in the distance function. Among the advantages of this technique are higher precision and recall rates with fast matching time. 2.Object recognition based on skeletal graphs. The topologies of objects' skeletal graphs are captured and compared at the node level. Such graph representation allows preservation of the skeletal graph's coherence without sacrificing the flexibility of matching similar portions of graphs across different levels. The technique is robust to translation, scaling, and rotation invariants at object level. This technique achieves high precision and recall rates with reasonable matching time and storage space. 3.ASIA (Automatic Sampling-based Image Annotation) is a technique based on a new sampling-based matching framework allowing users to identify their area of interest. ASIA eliminates noise, or irrelevant areas of the image. ASIA is robust to translation, scaling, and rotation invariants at the object level. This technique also achieves high precision and recall rates. While the above techniques may not be the fastest when contrasted with some other recent QBE techniques, they very effectively perform image annotation. The results of applying these processes are accurately annotated database images to which QBC may then be applied. The results of extensive experiments are presented to substantiate the performance advantages of the proposed techniques and allow them to be compared with other recent high-performance techniques. Additionally, a discussion on merging the proposed techniques into a highly effective annotation system is also detailed.

A large part of the empirical work on differentiated products markets has focused on demand estimation and the pricing behavior of firms. These two themes are key inputs in important applications such as the merging of two firms or the introduction of new products. The validity of inferences, therefore, depends on accurate demand estimates and sound assumptions about the pricing behavior of firms. This dissertation makes a contribution to this literature in two ways. First, it adds to previous techniques of estimating demand for differentiated products. Second, it extends previous analyses of pricing behavior to models of price leadership that, while important, have received limited attention. The investigation focuses on the U.S. brewing industry, where price leadership appears to be an important type of firm behavior. The analysis is conducted in two stages. In the first stage, the recent Distance Metric (DM) method devised by Pinkse, Slade and Brett is used to estimate the demand for 64 brands of beer in 58 major metropolitan areas of the United States. This study adds to previous applications of the DM method (Pinkse and Slade; Slade 2004) by employing a demand specification that is more flexible and also by estimating advertising substitution coefficients for numerous beer brands. In the second stage, different pricing models are compared and ranked by exploiting the exogenous change in the federal excise tax of 1991. Demand estimates of the first stage are used to compute the implied marginal costs for the different models of pricing behavior prior to the tax increase. Then, the tax increase is added to the these pre-tax increase marginal costs, and equilibrium prices for all brands are simulated for each model of pricing behavior. These "predicted" prices are then compared to actual prices for model assessment. Results indicate that Bertrand-Nash predicts the pricing behavior of firms more closely than other models, although Stackelberg leadership yields results that are not substanitally different from the Bertrand-Nash model. Nevertheless, Bertrand-Nash tends to under-predict prices of more price-elastic brands and to over-predict prices of less price- elastic brands. An implication of this result is that Anheuser-Busch could exert more market power by increasing the price of its highly inelastic brands, especially Budweiser. Overall, actual price movements as a result of the tax increase tend to be more similar across brands than predicted by any of the models considered. While this pattern is not inconsistent with leadership behavior, leadership models considered in this dissertation do not conform with this pattern. / Ph. D.

A systemic perspective was employed in completing a phonological analysis and developing an intervention plan for Jarrod, a 7;0 year old child who exhibited a severe speech sound disorder characterized by inconsistency. Results of the Systemic Phonological Analysis of Child Speech (SPACS) revealed a limited sound system that was characterized by phonotactic inventory constraints, positional constraints, and sequence constraints. Mapping the child-to-adult sound systems through phoneme collapses revealed a logical and symmetrical system that maintained systematicity, yet permitted variability. Based on the organizational principles suggested by the phoneme collapses, targets were identified for intervention using the distance metric approach, which is based on the function of sounds within a given system rather than the characteristics of a given sound, and assumes that targets will interact dynamically with the child's unique sound system. Finally, a multiple oppositions treatment approach intended to facilitate learning across phoneme collapses and lead to system-wide phonological restructuring was described.

There are a number of clinical options available for speech-language pathologists to choose from to analyze a child's phonological system, select treatment targets, and design intervention. Frequently, each of these areas of clinical options is viewed independently of one another or approached within an eclectic framework. In this article, an integrated and systemic approach is presented which assumes that a dynamic interaction exists among assessment, target selection, and intervention. Systemic Phonological Assessment of Child Speech, the distance metric approach to target selection, and the multiple oppositions treatment approach are described, with examples provided for each component. Finally, a case study is presented that examines the systemic approach of multiple oppositions relative to the approach of minimal pairs.

In this work, we propose a novel method for a supervised dimensionality reduc- tion, which learns weights of a neural network using an evolutionary algorithm, CMA-ES, optimising the success rate of the k-NN classifier. If no activation func- tions are used in the neural network, the algorithm essentially performs a linear transformation, which can also be used inside of the Mahalanobis distance. There- fore our method can be considered to be a metric learning algorithm. By adding activations to the neural network, the algorithm can learn non-linear transfor- mations as well. We consider reductions to low-dimensional spaces, which are useful for data visualisation, and demonstrate that the resulting projections pro- vide better performance than other dimensionality reduction techniques and also that the visualisations provide better distinctions between the classes in the data thanks to the locality of the k-NN classifier. 1

No description available.