Abstract Reinforcement learning often struggles to accomplish a sparse-reward long-horizon task in a complex environment. Goal-conditioned reinforcement learning (GCRL) has been employed to tackle this difficult problem via a curriculum of easy-to-reach sub-goals. In GCRL, exploring novel sub-goals is essential for the agent to ultimately find the pathway to the desired goal. How to explore novel sub-goals efficiently is one of the most challenging issues in GCRL. Several goal exploration methods have been proposed to address this issue but still struggle to find the desired goals efficiently. In this paper, we propose a novel learning objective by optimizing the entropy of both achieved and new goals to be explored for more efficient goal exploration in sub-goal selection based GCRL. To optimize this objective, we first explore and exploit the frequently occurring goal-transition patterns mined in the environments similar to the current task to compose skills via skill learning. Then, the pre-trained skills are applied in goal exploration with theoretical justification. Evaluation on a variety of spare-reward long-horizon benchmark tasks suggests that incorporating our method into several state-of-the-art GCRL baselines significantly boosts their exploration efficiency while improving or maintaining their performance.

Feature importance ranking has become a powerful tool for explainable AI. However, its nature of combinatorial optimization poses a great challenge for deep learning. In this paper, we propose a novel dual-net architecture consisting of operator and selector for discovery of an optimal feature subset of a fixed size and ranking the importance of those features in the optimal subset simultaneously. During learning, the operator is trained for a supervised learning task via optimal feature subset candidates generated by the selector that learns predicting the learning performance of the operator working on different optimal subset candidates. We develop an alternate learning algorithm that trains two nets jointly and incorporates a stochastic local search procedure into learning to address the combinatorial optimization challenge. In deployment, the selector generates an optimal feature subset and ranks feature importance, while the operator makes predictions based on the optimal subset for test data. A thorough evaluation on synthetic, benchmark and real data sets suggests that our approach outperforms several state-of-the-art feature importance ranking and supervised feature selection methods. (Our source code is available: https://github.com/maksym33/FeatureImportanceDL)

Speech emotion recognition plays an important role in building more intelligent and human-like agents. Due to the difficulty of collecting speech emotional data, an increasingly popular solution is leveraging a related and rich source corpus to help address the target corpus. However, domain shift between the corpora poses a serious challenge, making domain shift adaptation difficult to function even on the recognition of positive/negative emotions. In this work, we propose class-wise adversarial domain adaptation to address this challenge by reducing the shift for all classes between different corpora. Experiments on the well-known corpora EMODB and Aibo demonstrate that our method is effective even when only a very limited number of target labeled examples are provided.

Among the main challenges in procedural content generation (PCG), content quality assurance and dynamic difficulty adjustment (DDA) of game content in real time are two major issues concerned in adaptive content generation. Motivated by the recent learning-based PCG framework, we propose a novel approach to seamlessly address two issues in Super Mario Bros (SMB). To address the quality assurance issue, we exploit the synergy between rule-based and learning-based methods to produce quality game segments in SMB, named constructive primitives (CPs). By means of CPs, we propose a DDA algorithm that controls a CP-based level generator to adjust the content difficulty rapidly based on players' real-time game playing performance. We have conducted extensive simulations with sophisticated SMB agents of different types for thorough evaluation. Experimental results suggest that our approach can effectively assure content quality in terms of generic quality measurements and dynamically adjust game difficulty in real time as informed by the game completion rate.

Various aspects of computer game design, including adaptive elements of game levels, characteristics of “bot” behavior, and player matching in multiplayer games, would ideally be sensitive to a player's skill level. Yet, while game difficulty and player learning have been explored in the context of games, there has been little work analyzing skill per se, and how this is related to the interaction of a player with the controls of the game - the player's input. To this end, we present a data set of 476 game logs from over 40 players of a first-person shooter game (Red Eclipse) as a basis of a case study. We then extract features from the keyboard and mouse input and provide an analysis in relation to skill. Finally, we show that a player's skill can be predicted using less than a minute of their keyboard presses. We suggest that the techniques used here are useful for adapting games to match players' skill levels rapidly, arguably more rapidly than solutions based on performance averaging such as TrueSkill.

Music information retrieval faces a challenge in modeling contextualized musical concepts formulated by a set of co-occurring tags. In this article, we investigate the suitability of our recently proposed approach based on a Siamese neural network in fighting off this challenge. By means of tag features and probabilistic topic models, the network captures contextualized semantics from tags via unsupervised learning. This leads to a distributed semantics space and a potential solution to the out of vocabulary problem, which has yet to be sufficiently addressed. We explore the nature of the resultant music-based semantics and address computational needs. We conduct experiments on three public music tag collections—namely, CAL500, MagTag5K and Million Song Dataset—and compare our approach to a number of state-of-the-art semantics learning approaches. Comparative results suggest that this approach outperforms previous approaches in terms of semantic priming and music tag completion.

Mobile devices have become an integral part of our routine activities. Some of the activities involve the storage or access of sensitive data (e.g. on-line banking, paperless prescription services, etc.). These mobile electronic services (e-Services) typically require a method to securely identify and authenticate a claimed identity. Currently, e-Services typically use a knowledge-based authentication method by demonstrating the knowledge of a secret (e.g. password), but it is vulnerable to a number of security attacks, e.g. shoulder spoofing and brute force attacks. To thwart the attacks and to make the authentication method more secure, this paper describes our efforts in investigating the benefits of integrating touch dynamics biometrics, into a PIN-based authentication method. It reports the collection of a comprehensive reference dataset from 150 subjects, the extraction of feature data from the dataset, and the classifications and the use of the feature data to identify a user. Experimental results show that, even when the PIN is exposed, 9 out of 10 impersonation attempts can be successfully detected.

Spatial memory refers to the part of the memory system that encodes, stores, recognizes and recalls spatial information about the environment and the agent's orientation within it. Such information is required to be able to navigate to goal locations, and is vitally important for any embodied agent, or model thereof, for reaching goals in a spatially extended environment. In this paper, a number of computationally implemented cognitive models of spatial memory are reviewed and compared. Three categories of models are considered: symbolic models, neural network models, and models that are part of a systems-level cognitive architecture. Representative models from each category are described and compared in a number of dimensions along which simulation models can differ (level of modeling, types of representation, structural accuracy, generality and abstraction, environment complexity), including their possible mapping to the underlying neural substrate. Neural mappings are rarely explicated in the context of behaviorally validated models, but they could be useful to cognitive modeling research by providing a new approach for investigating a model's plausibility. Finally, suggested experimental neuroscience methods are described for verifying the biological plausibility of computational cognitive models of spatial memory, and open questions for the field of spatial memory modeling are outlined.

Procedural content generation (PCG) has recently become one of the hottest topics in computational intelligence and AI game research. While some substantial progress has been made in this area, there are still several challenges ranging from content evaluation to personalized content generation. In this paper, we present a novel PCG framework based on machine learning, named learning-based procedure content generation (LBPCG), to tackle a number of challenging problems. By exploring and exploiting information gained in game development and public player test, our framework can generate robust content adaptable to end-user or target players on-line with minimal interruption to their gameplay experience. As the data-driven methodology is emphasized in our framework, we develop learning-based enabling techniques to implement the various models required in our framework. For a proof of concept, we have developed a prototype based on the classic open source first-person shooter game, Quake. Simulation results suggest that our framework is promising in generating quality content.

Speech signals convey different types of information which vary from linguistic to speaker-specific and should be used in different tasks. However, it is hard to extract a special type of information such that nearly all acoustic representations of speech present all kinds of information as a whole. The use of the same representation in different tasks creates a difficulty in achieving good performance in either speech or speaker recognition. In this paper, we present a deep neural architecture to explore speaker-specific characteristics from popular Mel-frequency cepstral coefficients. For learning, we propose an objective function consisting of contrastive cost in terms of speaker similarity and dissimilarity as well as data reconstruction cost used as regularization to normalize non-speaker related information. Learning deep architecture is done by a greedy layerwise local unsupervised training for initialization and a global supervised discriminative training for extracting a speaker-specific representation. By means of two narrow-band benchmark corpora, we demonstrate that our deep architecture generates a robust overcomplete speech representation in characterizing various speakers and the use of this new representation yields a favorite performance in speaker verification.

This paper presents our investigations on automatic emotional state recognition from speech signals using ensemble based methods based on different acoustic representations/feature measures. In our work, we employ various types of acoustic feature measures where none of the feature measures is optimal for emotional state classification. It is observed that different feature measures may be complementary and used simultaneously to yield a robust classification performance. Therefore, we employ a probabilistic method of combining classifiers based on different feature measures. The combination method that uses different feature measures simultaneously yields high recognition rates on various emotional speech corpora for both full feature set and language-independent feature subset. The ensemble method also outperforms a composite-feature representation and two other methods reported in literature. In addition, the classification accuracies achieved by our combination method are competitive with those mentioned in literature for different emotional speech corpora.

Speech conveys different yet mixed information ranging from linguistic to speaker-specific components, and each of them should be exclusively used in a specific task. However, it is extremely difficult to extract a specific information component given the fact that nearly all existing acoustic representations carry all types of speech information. Thus, the use of the same representation in both speech and speaker recognition hinders a system from producing better performance due to interference of irrelevant information. In this paper, we present a deep neural architecture to extract speaker-specific information from MFCCs. As a result, a multi-objective loss function is proposed for learning speaker-specific characteristics and regularization via normalizing interference of non-speaker related information and avoiding information loss. With LDC benchmark corpora and a Chinese speech corpus, we demonstrate that a resultant speaker-specific representation is insensitive to text/languages spoken and environmental mismatches and hence outperforms MFCCs and other state-of-the-art techniques in speaker recognition. We discuss relevant issues and relate our approach to previous work.

Semi-supervised learning concerns the problem of learning in the presence of labeled and unlabeled data. Several boosting algorithms have been extended to semi-supervised learning with various strategies. To our knowledge, however, none of them takes all three semi-supervised assumptions, i.e., smoothness, cluster, and manifold assumptions, together into account during boosting learning. In this paper, we propose a novel cost functional consisting of the margin cost on labeled data and the regularization penalty on unlabeled data based on three fundamental semi-supervised assumptions. Thus, minimizing our proposed cost functional with a greedy yet stagewise functional optimization procedure leads to a generic boosting framework for semi-supervised learning. Extensive experiments demonstrate that our algorithm yields favorite results for benchmark and real-world classification tasks in comparison to state-of-the-art semi-supervised learning algorithms, including newly developed boosting algorithms. Finally, we discuss relevant issues and relate our algorithm to the previous work.

Temporal data clustering provides underpinning techniques for discovering the intrinsic structure and condensing information over temporal data. In this paper, we present a temporal data clustering framework via a weighted clustering ensemble of multiple partitions produced by initial clustering analysis on different temporal data representations. In our approach, we propose a novel weighted consensus function guided by clustering validation criteria to reconcile initial partitions to candidate consensus partitions from different perspectives, and then, introduce an agreement function to further reconcile those candidate consensus partitions to a final partition. As a result, the proposed weighted clustering ensemble algorithm provides an effective enabling technique for the joint use of different representations, which cuts the information loss in a single representation and exploits various information sources underlying temporal data. In addition, our approach tends to capture the intrinsic structure of a data set, e.g., the number of clusters. Our approach has been evaluated with benchmark time series, motion trajectory, and time-series data stream clustering tasks. Simulation results demonstrate that our approach yields favorite results for a variety of temporal data clustering tasks. As our weighted cluster ensemble algorithm can combine any input partitions to generate a clustering ensemble, we also investigate its limitation by formal analysis and empirical studies.

Speech signals convey various yet mixed information ranging from linguistic to speaker-specific information. However, most of acoustic representations characterize all different kinds of information as whole, which could hinder either a speech or a speaker recognition (SR) system from producing a better performance. In this paper, we propose a novel deep neural architecture (DNA) especially for learning speaker-specific characteristics from mel-frequency cepstral coefficients, an acoustic representation commonly used in both speech recognition and SR, which results in a speaker-specific overcomplete representation. In order to learn intrinsic speaker-specific characteristics, we come up with an objective function consisting of contrastive losses in terms of speaker similarity/dissimilarity and data reconstruction losses used as regularization to normalize the interference of non-speaker-related information. Moreover, we employ a hybrid learning strategy for learning parameters of the deep neural networks: i.e., local yet greedy layerwise unsupervised pretraining for initialization and global supervised learning for the ultimate discriminative goal. With four Linguistic Data Consortium (LDC) benchmarks and two non-English corpora, we demonstrate that our overcomplete representation is robust in characterizing various speakers, no matter whether their utterances have been used in training our DNA, and highly insensitive to text and languages spoken. Extensive comparative studies suggest that our approach yields favorite results in speaker verification and segmentation. Finally, we discuss several issues concerning our proposed approach.

Time series clustering provides underpinning techniques for discovering the intrinsic structure and condensing/summarizing information conveyed in time series, which is demanded in various fields ranging from bioinformatics to video content understanding. In this paper, we present an unsupervised ensemble learning approach to time series clustering by combining rival-penalized competitive learning (RPCL) networks with different representations of time series. In our approach, the RPCL network ensemble is employed for clustering analyses based on different representations of time series whenever available, and an optimal selection function is applied to find out a final consensus partition from multiple partition candidates yielded by applying various consensus functions for the combination of competitive learning results. As a result, our approach first exploits its capability of the RPCL rule in clustering analysis of automatic model selection on individual representations and subsequently applies ensemble learning for the synergy of reconciling diverse partitions resulted from the use of different representations and augmenting RPCL networks in automatic model selection and overcoming its inherent limitation. Our approach has been evaluated on 16 benchmark time series data mining tasks with comparison to state-of-the-art time series clustering techniques. Simulation results demonstrate that our approach yields favorite results in clustering analysis of automatic model selection.

Unsupervised classification or clustering is an important data analysis technique demanded in various fields including machine learning, data mining, pattern recognition, image analysis and bioinformatics. Recently a large number of studies have attempted to improve clustering by combing multiple clustering solutions into a single consolidated clustering ensemble that has the best performance among given clustering solutions. However, the different clustering ensembles have their own behaviors on data of various characteristics. In this paper, we propose a novel approach to data clustering by constructing a clustering ensemble iteratively based on partitions generated on training subsets sampled from the original dataset. To yield a robust clustering ensemble our approach employs a hybrid sampling scheme inspired by both boosting and bagging techniques originally proposed for supervised learning. Our approach has been evaluated on synthetic data and real-world motion trajectory data sets, and experimental results demonstrate that it yields satisfactory performance for a variety of clustering tasks.

This paper investigates the performance of automatic emotional state categorization from speech signals on the Serbian Emotional Speech Corpus, named GEES, against the corresponding human performance. We employ a multistage strategy along with sophisticated features used for automatic emotional state categorization. Our study is the first attempt to apply a machine learning technique to the GEES where the human performance was only available prior to our study. Our investigation indicates that the use of a multistage categorization strategy yields behaviors similar to what human perceives and the performance close to human being's. Copyright © 2008 ISCA.

In the above titled paper (ibid., vol 27, no. 10, pp. 1552-1567, Oct 05), equation (9) contained a typographical error. The corrected equation is presented here.

Unlike traditional pattern classification, semi-supervised learning provides a novel technique to make use of both labeled and unlabeled data for improving the performance of classification. In general, there are two critical issues for semi-supervised learning of discriminative classifiers; i.e., how to create an initial classifier of a good generalization capability with the limited labeled data and the how to make an effective use of unlabeled data without degradation of the established classifier. To tackle two aforementioned problems, we propose an ensemble learning approach based on a recent active data selection strategy, where ensemble learning would yield good generalization and active data selection tends to choose the unlabeled data more likely resulting in an improvement during semi-supervised learning. By using an ensemble of K-NN classifiers, we demonstrate the effectiveness of our approach on a synthetic data classification and a facial expression recognition tasks.

Semi-supervised inductive learning concerns how to learn a decision rule from a data set containing both labeled and unlabeled data. Several boosting algorithms have been extended to semi-supervised learning with various strategies. To our knowledge, however, none of them takes local smoothness constraints among data into account during ensemble learning. In this paper, we introduce a local smoothness regularizer to semi-supervised boosting algorithms based on the universal optimization framework of margin cost functionals. Our regularizer is applicable to existing semi-supervised boosting algorithms to improve their generalization and speed up their training. Comparative results on synthetic, benchmark and real world tasks demonstrate the effectiveness of our local smoothness regularizer. We discuss relevant issues and relate our regularizer to previous work.

Temporal data clustering provides useful techniques for condensing and summarizing information conveyed in temporal data, which is demanded in various fields ranging from time series analysis to sequential data understanding. In this paper, we propose a novel approach to temporal data clustering by an ensemble of competitive learning networks incorporated by different representations of temporal data. In our approach, competitive learning networks of the rival-penalized learning mechanism are employed for clustering analyses based on different temporal data representations while an optimal selection function is applied to find out a final consensus partition from multiple partition candidates yielded by applying alternative consensus functions to results of competitive learning on different representations. Thanks to its capability of the rival penalized learning rules in automatic model selection and the synergy of fusing diverse partitions on different representations, our ensemble approach yields favorite results, which has been demonstrated in time series and motion trajectory clustering tasks.

A novel adaptive smoothing approach is proposed for noise removal and feature preservation where two distinct measures are simultaneously adopted to detect discontinuities in an image. Inhomogeneity underlying an image is employed as a multiscale measure to detect contextual discontinuities for feature preservation and control of the smoothing speed, while local spatial gradient is used for detection of variable local discontinuities during smoothing. Unlike previous adaptive smoothing approaches, two discontinuity measures are combined in our algorithm for synergy in preserving nontrivial features, which leads to a constrained anisotropic diffusion process that inhomogeneity offers intrinsic constraints for selective smoothing. Thanks to the use of intrinsic constraints, our smoothing scheme is insensitive to termination times and the resultant images in a wide range of iterations are applicable to achieve nearly identical results for various early vision tasks. Our algorithm is formally analyzed and related to anisotropic diffusion. Comparative results indicate that our algorithm yields favorable smoothing results, and its application in extraction of hydrographic objects demonstrates its usefulness as a tool for early vision.

Numerous speech representations have been reported to be useful in speaker recognition. However, there is much less agreement on which speech representation provides a perfect representation of speaker-specific information conveyed in a speech signal. Unlike previous work, we propose an alternative approach to speaker modeling by the simultaneous use of different speech representations in an optimal way. Inspired by our previous empirical studies, we present a soft competition scheme on different speech representations to exploit different speech representations in encoding speaker-specific information. On the basis of this soft competition scheme, we present a parametric statistical model, generalized Gaussian mixture model (GGMM), to characterize a speaker identity based on different speech representations. Moreover, we develop an expectation-maximization algorithm for parameter estimation in the GGMM. The proposed speaker modeling approach has been applied to text-independent speaker recognition and comparative results on the KING speech corpus demonstrate its effectiveness.

This paper addresses the problem of real-time speaker change detection in broadcast news, in which no prior knowledge on speakers is assumed. Our speaker segmentation is a "coarse to refine" process, which consists of two stages: pre-segmentation and refinement. In the pre-segmentation stage, a new approach based on Gaussian mixture model-universal background model (GMM-UBM) is proposed to categorize feature vectors into three sets, i.e. reliable speaker-related set, doubtful speaker-related set and unreliable speaker-related set, in order to enhance the effect of the reliable speaker-related feature vectors. Then potential speaker change boundaries are detected based on a novel distance measure. In the refinement stage, incremental speaker adaptation (ISA), which is suitable for real-time requirement, is proposed to obtain considerably precise speaker models so that the potential speaker change boundaries can be confirmed and refined. Experimental results demonstrate that our approach yields satisfactory performance.

This paper addresses a novel algorithm of incremental speaker adaptation (ISA) based on universal background model (UBM) for saving storage and real-time processing. This algorithm can be seen as an extension of traditional speaker adaptation. It consists of two steps, adaptation and combination. It not only considers the speaker's characteristics in limited training data, but also prohibits over-fitting of the updated model. The incremental adaptation algorithm needs little storage and meets the requirement of real-time processing. In order to evaluate the efficiency and effectivity of the proposed approach, a real-time speaker segmentation system for broadcasting news is built. Experiment results demonstrate that our approach yields real time operation and achieves satisfactory performance.

As a new approach of automatic speaker authentication [1], VIV has manifested excellent performance in English, which turns to be complementary with speaker verification. In this paper, we study the VIV for users who speak Mandarin Chinese, the most widely used language in the world. We present our methodologies to build target acoustic models and anti-models for Mandarin VIV. In our experiments of 50 Chinese speakers, the EER reaches 0.00% with robust interval of 8.2%, which indicates that our Mandarin VIV technologies are promising for Mandarin speaker authentication.

Model-based approach is one of methods widely used for speaker identification, where a statistical model is used to characterize a specific speaker's voice but no interspeaker information is involved in its parameter estimation. It is observed that interspeaker information is very helpful in discriminating between different speakers. In this paper, we propose a novel method for the use of interspeaker information to improve performance of a model-based speaker identification system. A neural network is employed to capture the interspeaker information from the output space of those statistical models. In order to sufficiently utilize interspeaker information, a rival penalized encoding rule is proposed to design supervised learning pairs. For better generalization, moreover, a query-based learning algorithm is presented to actively select the input data of interest during training of the neural network. Comparative results on the KING speech corpus show that our method leads to a considerable improvement for a model-based speaker identification system.

This paper presents a progress update of our ISIS 1 trilingual spoken dialog system. As described in [8], this is a conversational system for the stocks domain, and supports interactions in the languages of our region – English and two dialects of Chinese (Mandarin and Cantonese). ISIS provides a system test-bed for our initial explorations with the CORBA architecture, and delegation to KQML (Knowledge Query and Manipulation Language) agents. CORBA offers the advantages of interoperability, scalability and location transparency in client/server systems development. Users can delegate tasks to software agents to help monitor information (e.g. a drop in the price of a pre-specified stock), and generate user alert messages. Our current work presents new research directions in the context of ISIS: (i) automatic incorporation of newly listed stocks into our system’s knowledge base; (ii) switching between on-line interaction and off-line delegation in a single dialog thread. We will also report on enhancements in the system’s architecture and features (e.g. automatic end-point detection). delegation in a single dialog thread. We will also report on enhancements in the system’s architecture and features (e.g. automatic end-point detection). 2. System Architecture Previous work in the development of software infrastructures for dialog systems include [11] 2 [2]. Over the past year, we have continued to explore the development of a spoken dialog system based on the CORBA architecture. This middleware resides in between the operating system and the application layer 1.

The authors propose a framework for object extraction with accurate boundaries. A multilayer perceptron is used to identify seed points through examples, and regions are extracted and localized using a locally coupled network with weight adaptation. A functional system has been developed and applied to hydrographic region extraction from Digital Orthophoto Quarter-Quadrangle images.

This research monograph was developed in lectures to intermediate- level engineering and statistics students. It admirably surveys the field of feedforward networks considering both theoretical and practical aspects. The books attempts to respond to four fundamental questions concerning computational capabilities of feedforward networks:

Since first proposed by Minsky and Papert (1969), the spiral problem is well known in neural networks. It receives much attention as a benchmark for various learning algorithms. Unlike previous work that emphasizes learning, we approach the problem from a different perspective. We point out that the spiral problem is intrinsically connected to the inside-outside problem proposed by Ullman (1984, 1996). We propose a solution to both problems based on oscillatory correlation using a time-delay network. Our simulation results are qualitatively consistent with human performance, and we interpret human limitations in terms of synchrony and time delays. As a special case, our network without time delays can always distinguish these figures regardless of shape, position, size, and orientation.

Many speaker identification systems are created by model-based approaches, where a statistical model is used to characterize a speaker's voice and no inter-speaker information is used in parameter estimation. It is well known that inter-speaker information is very helpful in discrimination of different speakers. We propose a method for the use of inter-speaker information to improve performance of a model-based speaker identification system. A neural network is employed to capture inter-speaker information from output space of those statistical models. In order to sufficiently utilize inter-speaker information, a rival penalized encoding rule is proposed to design supervised learning pairs for training the neural network. Comparative results in the KING speech corpus show that our method leads to a considerable improvement for a model-based speaker identification system.

ISIS, which abbreviates Intelligent Speech for Information Systems, is a trilingual spoken dialog system (SDS) for the financial domain. It handles two dialects of Chinese (Cantonese and Putonghua), as well as English the predominant languages in our region. The system supports spoken language queries regarding stock market information and simulated personal portfolios. Real-time information is retrieved directly from a dedicated Reuters satellite feed. ISIS provides a system test-bed for our work in multilingual speech recognition and generation, speaker authentication, language understanding and dialog modeling. Furthermore, ISIS supports our initial explorations in: (i) CORBA's interoperability and scalability for SDS development; in conjunction with (ii) asynchronous human-computer interaction by delegation to KQML software agents...

We propose a method for image segmentation based on a neural oscillator network. Unlike previous methods, weight adaptation is adopted during segmentation to remove noise and preserve significant discontinuities in an image. Moreover, a logarithmic grouping rule is proposed to facilitate grouping of oscillators representing pixels with coherent properties. We show that weight adaptation plays the roles of noise removal and feature preservation. In particular, our weight adaptation scheme is insensitive to termination time and the resulting dynamic weights in a wide range of iterations lead to the same segmentation results. A computer algorithm derived from oscillatory dynamics is applied to synthetic and real images and simulation results show that the algorithm yields favorable segmentation results in comparison with other recent algorithms. In addition, the weight adaptation scheme can be directly transformed to a novel feature-preserving smoothing procedure. We also demonstrate that our nonlinear smoothing algorithm achieves good results for various kinds of images.

We propose an alternative method for the use of different feature sets in pattern classification. Unlike traditional methods, e.g. combination of multiple classifiers and use of a composite feature set, our method copes with the problem based on an idea of soft competition on different feature sets, a modular neural network architecture is proposed to implement the idea accordingly. The proposed architecture is interpreted as a generalized finite mixture model and, therefore, parameter estimation is treated as a maximum likelihood problem. An EM algorithm is derived for parameter estimation. Moreover, we propose a heuristic model selection method to fit the proposed architecture to a specific problem. Comparative results are presented for the real world problem of speaker identification.

In this paper, a dynamically coupled neural oscillator network is proposed for image segmentation. Instead of pair-wise coupling, an ensemble of oscillators coupled in a local region is used for grouping. We introduce a set of neighborhoods to generate dynamical coupling structures associated with a specific oscillator. Based on the proximity and similarity principles, two grouping rules are proposed to explicitly consider the distinct cases of whether an oscillator is inside a homogeneous image region or near a boundary between different regions. The use of dynamical coupling makes our segmentation network robust to noise on an image. For fast computation, a segmentation algorithm is abstracted from the underlying oscillatory dynamics and has been applied to synthetic and real images. Simulation results demonstrate the effectiveness of our oscillator network in image segmentation.

We propose a novel connectionist method for the use of different feature sets in pattern classification. Unlike traditional methods, e.g., combination of multiple classifiers and use of a composite feature set, our method copes with the problem based on an idea of soft competition on different feature sets developed in our earlier work. An alternative modular neural network architecture is proposed to provide a more effective implementation of soft competition on different feature sets. The proposed architecture is interpreted as a generalized finite mixture model and, therefore, parameter estimation is treated as a maximum likelihood problem. An EM algorithm is derived for parameter estimation and, moreover, a model selection method is proposed to fit the proposed architecture to a specific problem. Comparative results are presented for the real world problem of speaker identification.

In practical applications of pattern recognition, there are often different features extracted from raw data which needs recognizing. Methods of combining multiple classifiers with different features are viewed as a general problem in various application areas of pattern recognition. In this paper, a systematic investigation has been made and possible solutions are classified into three frameworks, i.e. linear opinion pools, winner-take-all and evidential reasoning. For combining multiple classifiers with different features, a novel method is presented in the framework of linear opinion pools and a modified training algorithm for associative switch is also proposed in the framework of winner-take-all. In the framework of evidential reasoning, several typical methods are briefly reviewed for use. All aforementioned methods have already been applied to text-independent speaker identification. The simulations show that results yielded by the methods described in this paper are better than not only the individual classifiers' but also ones obtained by combining multiple classifiers with the same feature. It indicates that the use of combining multiple classifiers with different features is an effective way to attack the problem of text-independent speaker identification.

Manuscript received November 22, 1996. K. Chen was with the National Lab of Machine Perception and Center for Information Science, Peking University, Beijing 100871, China. He is now with the Department of Computer and Information Science, The Ohio State University, Columbus, OH 43210-1277 USA. D. Xie was with the National Lab of Machine Perception and Center for Information Science, Peking University, Beijing 100871, China. He is now with the Center for Advanced Computer Studies, The University of Southwestern Louisiana, Lafayette, LA 70504 USA. H. Chi is with the National Lab of Machine Perception and Center for Information Science, Peking University, Beijing 100871, China. Publisher Item Identifier S 1045-9227(97)02888-9. 1K. Chen, D. Xie and H. Chi, IEEE Trans. Neural Networks, vol. 7, no. 5, pp. 1309–1313, Sept. 1996.

In this paper, we extend the Hierarchical Mixture of Experts (HME) to temporal processing and explore it for a substantial problem, that of text-dependent speaker identification. For a specific multiway classification, we propose a generalized Bernoulli density instead of the multinomial logit density to avoid the instability during training. Time-delay technique is applied for spatio-temporal processing in the HME and a combining scheme is presented for combining multiple time-delay HMEs in order to complete a multi-scale analysis for the temporal data. Using the time-delay HME along with the EM algorithm as well as the combination of multiple time-delay HMEs, the speaker identification system has a good performance and yields significantly fast training. We have also addressed some issues about the time-delay techniques in the HME.

A modified hierarchical mixtures of experts (HME) architecture is presented for text-dependent speaker identification. A new gating network is introduced to the original HME architecture for the use of instantaneous and transitional spectral information in text-dependent speaker identification. The statistical model underlying the proposed architecture is presented and learning is treated as a maximum likelihood problem; in particular, an expectation-maximization (EM) algorithm is also proposed for adjusting the parameters of the proposed architecture. An evaluation has been carried out using a database of isolated digit utterances by 10 male speakers. Experimental results demonstrate that the proposed architecture outperforms the original HME architecture in text-dependent speaker identification.