Hsin-Chia Fu

Divide-and-conquer learning and modular perceptron networks

A novel modular perceptron network (MPN) and divide-and-conquer learning (DCL) schemes for the design of modular neural networks are proposed. When a training process in a multilayer perceptron falls into a local minimum or stalls in a flat region, the proposed DCL scheme is applied to divide the current training data region into two easier to be learned regions. The learning process continues when a self-growing perceptron network and its initial weight estimation are constructed for one of the newly partitioned regions. Another partitioned region will resume the training process on the original perceptron network. Data region partitioning, weight estimating and learning are iteratively repeated until all the training data are completely learned by the MPN. We evaluated and compared the proposed MPN with several representative neural networks on the two-spirals problem and real-world dataset. The MPN achieved better weight learning performance by requiring much less data presentations during the network training phases, and better generalization performance, and less processing time during the retrieving phase.

BIC-Based Speaker Segmentation Using Divide-and-Conquer Strategies With Application to Speaker Diarization

In this paper, we propose three divide-and-conquer approaches for Bayesian information criterion (BlC)-based speaker segmentation. The approaches detect speaker changes by recursively partitioning a large analysis window into two sub-windows and recursively verifying the merging of two adjacent audio segments using DeltaBIC, a widely-adopted distance measure of two audio segments. We compare our approaches to three popular distance-based approaches, namely, Chen and Gopalakrishnans window-growing-based approach, Siegler et al.s fixed-size sliding window approach, and Delacourt and Wellekenss DISTBIC approach, by performing computational cost analysis and conducting speaker change detection experiments on two broadcast news data sets. The results show that the proposed approaches are more efficient and achieve higher segmentation accuracy than the compared distance-based approaches. In addition, we apply the segmentation approaches discussed in this paper to the speaker diarization task. The experiment results show that a more effective segmentation approach leads to better diarization accuracy.

Model-Based Clustering by Probabilistic Self-Organizing Maps

In this paper, we consider the learning process of a probabilistic self-organizing map (PbSOM) as a model-based data clustering procedure that preserves the topological relationships between data clusters in a neural network. Based on this concept, we develop a coupling-likelihood mixture model for the PbSOM that extends the reference vectors in Kohonens self-organizing map (SOM) to multivariate Gaussian distributions. We also derive three expectation-maximization (EM)-type algorithms, called the SOCEM, SOEM, and SODAEM algorithms, for learning the model (PbSOM) based on the maximum-likelihood criterion. SOCEM is derived by using the classification EM (CEM) algorithm to maximize the classification likelihood; SOEM is derived by using the EM algorithm to maximize the mixture likelihood; and SODAEM is a deterministic annealing (DA) variant of SOCEM and SOEM. Moreover, by shrinking the neighborhood size, SOCEM and SOEM can be interpreted, respectively, as DA variants of the CEM and EM algorithms for Gaussian model-based clustering. The experimental results show that the proposed PbSOM learning algorithms achieve comparable data clustering performance to that of the deterministic annealing EM (DAEM) approach, while maintaining the topology-preserving property.

An EM based multiple instance learning method for image classification

In this paper, we propose an EM based learning algorithm to provide a comprehensive procedure for maximizing the measurement of diverse density on given multiple Instances. Furthermore, the new EM based learning framework converts an MI problem into a single-instance treatment by using EM to maximize the instance responsibility for the corresponding label of each bag. To learn a desired image class, a user may select a set of exemplar images and label them to be conceptual related (positive) or conceptual unrelated (negative) images. A positive image consists of at least one object that the user may be interested, and a negative image should not contain any object that the user may be interested. By using the proposed EM based learning algorithm, an image retrieval prototype system is implemented. Experimental results show that for only a few times of relearning cycles, the prototype system can retrieve users favor images from WWW over Internet.

User adaptive handwriting recognition by self-growing probabilistic decision-based neural networks

It is generally agreed that, for a given handwriting recognition task, a user dependent system usually outperforms a user independent system, as long as a sufficient amount of training data is available. When the amount of user training data is limited, however, such a performance gain is not guaranteed. One way to improve the performance is to make use of existing knowledge, contained in a rich multiuser data base, so that a minimum amount of training data is sufficient to initialize a model for the new user.We mainly address the user adaption issues for a handwriting recognition system. Based on self-growing probabilistic decision-based neural networks (SPDNNs), user adaptation of the parameters of SPDNN is formulated as incremental reinforced and antireinforced learning procedures, which are easily integrated into the batched training procedures of the SPDNN. In this study, we developed 1) an SPDNN based handwriting recognition system; 2) a two-stage recognition structure; and 3) a three-phase training methodology for a) a global coarse classifier (stage 1); b) a user independent hand written character recognizer (stage 2); and c) a user adaptation module on a personal computer. With training and testing on a 600-word commonly used Chinese character set, the recognition results indicate that the user adaptation module significantly improved the recognition accuracy. The average recognition rate increased from 44.2% to 82.4% in five adapting cycles, and the performance could finally increase up to 90.2% in ten adapting cycles.

Face detection and eye localization by neural network based color segmentation

This paper presents a neural network based scheme for human face detection and eye localization in color images under an unconstrained scene. A self-growing probabilistic decision-based neural network (SPDNN) is used to learn the conditional distribution for each color classes. Pixels of a color image are first classified into facial or non-facial regions, then pixels in the facial region are followed by eye region segmentation. The class of each pixel is determined by using the conditional distribution of the chrominance components of pixels belonging to each class. The paper demonstrates a successful application of SPDNN to face detection and eye localization on a database of 755 images from 151 persons. Regarding the performance, experimental results are elaborated. As to the processing speed, the face detection and eye localization processes consume approximately 560 ms on a Pentium-II personal computer.

A model-selection-based self-splitting Gaussian mixture learning with application to speaker identification

We propose a self-splitting Gaussian mixture learning (SGML) algorithm for Gaussian mixture modelling. The SGML algorithm is deterministic and is able to find an appropriate number of components of the Gaussian mixture model (GMM) based on a self-splitting validity measure, Bayesian information criterion (BIC). It starts with a single component in the feature space and splits adaptively during the learning process until the most appropriate number of components is found. The SGML algorithm also performs well in learning the GMM with a given component number. In our experiments on clustering of a synthetic data set and the text-independent speaker identification task, we have observed the ability of the SGML for model-based clustering and automatically determining the model complexity of the speaker GMMs for speaker identification.

A Prototypes-Embedded Genetic K-means Algorithm

This paper presents a genetic algorithm (GA) for K-means clustering. Instead of the widely applied string-of-group-numbers encoding, we encode the prototypes of the clusters into the chromosomes. The crossover operator is designed to exchange prototypes between two chromosomes. The one-step K-means algorithm is used as the mutation operator. Hence, the proposed GA is called the prototypes-embedded genetic K-means algorithm (PGKA). With the inherent evolution process of evolutionary algorithms, PGKA has superior performance than the classical K-means algorithm, while comparing to other GA-based approaches, PGKA is more efficient and suitable for large scale data sets

A self-growing probabilistic decision-based neural network with automatic data clustering

In this paper, we propose a new clustering algorithm for a mixture of Gaussian-based neural network and self-growing probabilistic decision-based neural networks (SPDNN). The proposed self-growing cluster learning (SGCL) algorithm is able to find the natural number of prototypes based on a self-growing validity measure, Bayesian information criterion (BIC). The learning process starts from a single prototype randomly initialized in the feature space and grows adaptively during the learning process until most appropriate number of prototypes are found. We have conducted numerical and real-world experiments to demonstrate the effectiveness of the SGCL algorithm. In the results of using SGCL to train the SPDNN for data clustering and speaker identification problems, we have observed a noticeable improvement among various model-based or vector quantization-based classification schemes.

Constructing and application of multimedia TV-news archives

This paper addresses an integrated information mining techniques for broadcasting TV-news. This utilizes technique from the fields of acoustic, image, and video analysis, for information on news story title, newsman and scene identification. The goal is to construct a compact yet meaningful abstraction of broadcast TV-news, allowing users to browse through large amounts of data in a non-linear fashion with flexibility and efficiency. By adding acoustic analysis, a news program can be partitioned into news and commercial clips, with 90% accuracy on a data set of 400h TV-news recorded off the air from July 2005 to August 2006. By applying speaker identification and/or image detection techniques, each news stories can be segmented with a better accuracy of 95.92%. On-screen captions or subtitles are recognized by OCR techniques to produce the text title of each news stories. The extracted title words can be used to link or to navigate more related news contents on the WWW. In cooperation with facial and scene analysis and recognition techniques, OCR results can provide users with multimodal query on specific news stories. Some experimental results are presented and discussed for the system reliability, performance evaluation and comparison.