Costin Barbu

ShareBoost: boosting for multi-view learning with performance guarantees

Algorithms combining multi-view information are known to exponentially quicken classification, and have been applied to many fields. However, they lack the ability to mine most discriminant information sources (or data types) for making predictions. In this paper, we propose an algorithm based on boosting to address these problems. The proposed algorithm builds base classifiers independently from each data type (view) that provides a partial view about an object of interest. Different from AdaBoost, where each view has its own re-sampling weight, our algorithm uses a single re-sampling distribution for all views at each boosting round. This distribution is determined by the view whose training error is minimal. This shared sampling mechanism restricts noise to individual views, thereby reducing sensitivity to noise. Furthermore, in order to establish performance guarantees, we introduce a randomized version of the algorithm, where a winning view is chosen probabilistically. As a result, it can be cast within a multi-armed bandit framework, which allows us to show that with high probability the algorithm seeks out most discriminant views of data for making predictions. We provide experimental results that show its performance against noise and competing techniques.

Fuzzy component based object detection

Component based object detection approaches have been shown to significantly improve object detection performance in adversities such as occlusion, variations in pose, in and out of plane rotation and poor illumination. Even the best object detectors are prone to errors when used in a global object detection scheme (one that uses the whole object as a single entity for detection purpose), due to these problems. We propose a fuzzy approach to object detection that treats an object as a set of constituent components rather than a single entity. The object detection task is completed in two steps. In the first step, candidates for respective components are selected based on their appearance match and handed over to the geometrical configuration classifier. The geometrical configuration classifier is a fuzzy inference engine that selects one candidate for each component such that each candidate is a reasonable match to the corresponding component in terms of appearance and also a good fit for the overall geometrical model. The detected object consists of candidates that are not necessarily the best in terms of appearance match or the closest to the geometrical model in terms of placement. The output is a set of candidates that is an optimal combination satisfying both criteria. We evaluate the technique on a well known face dataset and show that the technique results in detection of most faces in a scale-invariant manner. The technique has been shown to be robust to in-plane rotations and occlusion.

AQS-20 sonar processing enhancement for bathymetry estimation

Bathymetry is used to determine optimal tactics during Mine Warfare operations. Previous work demonstrated that bathymetric data could be acquired from the Volume Search Sonar (VSS) mounted on the AQS-20 system. The VSS transmitter produces a pulse at approximately one-second intervals along the track. The returning pulse from the sea-bottom is received by a group of sensors and beamformed in hardware into two fans (one pitched slightly forward and a second pitched slightly aft). A possible way to increase the accuracy of the bathymetry data is to improve the angle of arrival estimates by processing the adjacent across-track and/or along-track beam pairs. This paper employs narrow-beams monopulse techniques in order to investigate improvements to the bathymetric data over conventional processing. A comparative analysis of the experimental results for both the new and the classical technique is presented.

Classifier fusion using shared sampling distribution for boosting

We present a new framework for classifier fusion that uses a shared sampling distribution for obtaining a weighted classifier ensemble. The weight update process is self regularizing as subsequent classifiers trained on the disjoint views rectify the bias introduced by any classifier in preceding iterations. We provide theoretical guarantees that our approach indeed provides results which are better than the case when boosting is performed separately on different views. The results are shown to outperform other classifier fusion strategies on a well known texture image database.

Distributed multi-modal sensor system for searching a foliage-covered region

We designed and constructed a system that includes aircraft, ground vehicles, and throwable sensors to search a semi-forested region that was partially covered by foliage. The system contained 4 radio-controlled (RC) trucks, 2 aircraft, and 30 SensorMotes (throwable sensors). We also investigated communications links, search strategies, and system architecture. Our system is designed to be low-cost, contain a variety of sensors, and distributed so that the system is robust even if individual components are lost.

An Ensemble Approach to Robust Biometrics Fusion

A clever information fusion algorithm is a key component in designing a robust multimodal biometrics algorithm. We present a novel information fusion approach that can be a very useful tool for multimodal biometrics learning. The proposed technique is a multiple view generalization of AdaBoost in the sense that weak learners from various information sources are selected in each iteration based on lowest weighted error rate. Weak learners trained on individual views in each iteration rectify the bias introduced by learners in preceding iterations resulting in a self regularizing behavior. We compare the classification performance of proposed technique with recent classifier fusion strategies in various domains such as face detection, gender classification and texture classification.

Boosting in classifier fusion vs. fusing boosted classifiers

In this paper we investigate the performance of boosting used for fusing various classifiers. We propose a new boosting - based algorithm for fusion and we show through empirical studies on texture image data sets that it outperforms existing SVM-based classifier fusion technique in terms of accuracy, computational efficiency and robustness.

Frequency domain system identification: a full estimation approach

The purpose of this paper is to address the full estimation issue, where both the order and the parameters are identified for single-input/single-output linear time-invariant systems. The technique introduced here investigates the system identification problem when the phase information is lacking and only the magnitude of the input/output and the frequency data are available. The system identification approach described in this work has been shown to function with good results on both simulated and measured data

Meta latent semantic analysis

Meta latent semantic analysis (MLSA) is a novel approach to automated document analysis and indexing which relies on symbolic ontologies to further enhance the traditional probabilistic latent semantic analysis (LSA) of documents. While LSA is able to discover clusters of related terms and documents in a given collection of documents, the proposed MLSA is able to meta-cluster such clusters by taking into account existing symbolic ontologies relevant for the analyzed collections of documents. Such an approach can be successfully used to improve the performance of fast LSA by random projection.

Boosting information fusion

Ensemble methods provide a principled framework for building high performance classifiers and representing many types of data. As a result, these methods can be useful for making inferences in many domains such as classification and multi-modal biometrics. We introduce a novel ensemble method for combining multiple representations (or views). The method is a multiple view generalization of AdaBoost. Similar to AdaBoost, base classifiers are independently built from each representation. Unlike AdaBoost, however, all data types share the same sampling distribution as the view whose weighted training error is the smallest among all the views. As a result, the most consistent data type dominates over time, thereby significantly reducing sensitivity to noise. In addition, our proposal is provably better than AdaBoost trained on any single type of data. The proposed method is applied to the problems of facial and gender prediction based on biometric traits as well as of protein classification. Experimental results show that our method outperforms several competing techniques including kernel-based data fusion.