Several research fields have to deal with very large classification problems, e.g. handwritten character recognition and speech recognition. Many works have proposed methods to address problems with large number of samples, but few works have been done concerning problems with large numbers of classes. CombNET-II was one of the first methods proposed for such a kind of task. It consists of a sequential clustering VQ based gating network (stem network) and several Multilayer Perceptron (MLP) based expert classifiers (branch networks). With the objectives of increasing the classification accuracy and providing a more flexible model, this paper proposes a new model based on the CombNET-II structure, the CombNET-III. The new model, intended for, but not limited to, problems with large number of classes, replaces the branch networks MLP with multiclass Support Vector Machines (SVM). It also introduces a new probabilistic framework that outputs posterior class probabilities, enabling the model to be applied in different scenarios (e.g. together with Hidden Markov Models). These changes permit the use of a larger number of smaller clusters, which reduce the complexity of the final classifiers. Moreover, the use of binary SVM with probabilistic outputs and a probabilistic decoding scheme permit the use of a pairwise output encoding on the branch networks, which reduces the computational complexity of the training stage. The experimental results show that the proposed model outperforms both the previous model CombNET-II and a single multiclass SVM, while presenting considerably smaller complexity than the latter. It is also confirmed that CombNET-III classification accuracy scales better with the increasing number of clusters, in comparison with CombNET-II.

In this paper, a method of detecting overlapping speech segments in meetings is proposed. It is known that the eigenvalue distribution of the spatial correlation matrix calculated from a multiple microphone input reflects information on the number and relative power of sound sources. However, in a reverberant sound field, the feature of the number of sources in the eigenvalue distribution is degraded by the room reverberation. In the Support Vector Machines approach, the eigenvalue distribution is classified into two classes (overlapping speech segments and single speech segments). In the Support Vector Regression approach, the relative power of sound sources is estimated by using the eigenvalue distribution, and overlapping speech segments are detected based on the estimated relative power. The salient feature of this approach is that the sensitivity of detecting overlapping speech segments can be controlled simply by changing the threshold value of the relative power. The proposed method was evaluated using recorded data of an actual meeting.

To achieve a very high data rate within a limited frequency band in orthogonal frequency division multiplexing (OFDM) systems, multi-input multi-output (MIMO) techniques are very promising. Moreover, if a transmitter has the channel state information (CSI), the achievable spectrum efficiency can be maximized using the eigenbeam-space division multiplexing (E-SDM). However, this scheme demands accurate channel estimation. Therefore, in a closed-loop transmission scheme, an increase in the amount of feedback is absolutely necessary for the E-SDM. This paper describes a downlink beamforming method that significantly reduces the amount of feedback needed by using the common transmission weight vectors in all sub-carriers, compared to the amount required for E-SDM. The proposed method also applies transmission diversity to compensate for the quality. The effectiveness of the proposed method was confirmed using computer simulations in both Ricean and Rayleigh fading environments.

Field assembly of optical connectors is demanded because of the wide use of optical fiber in telecommunications systems. We propose a new assembling techniques that enable us to assemble connectors anywhere quickly and cost effectively. The key points are an adhesive technique and a polishing technique. In this report, we focus mainly on our a new polishing machine, which is suitable for optical connector ends machining on-site. The machine which is small and light weight can finish optical connector ends easily in a short time with enough low cost.

We achieved first dynamic all-optical signal processing with a bandgap-engineered MZI SOA all-optical switch. The wide-gap Selective Area Growth (SAG) technique was used to provide multi-bandgap materials with a single step epitaxy. The maximum photoluminescence (PL) peak shift obtained between the active region and the passive region was 192 nm. The static current switching with the fabricated switch indicated a large carrier induced refractive index change; up to 14 π phase shift was obtained with 60 mA injection in the SOA. The carrier recovery time of the SOA for obtaining a phase shift of π was estimated to be 250-300 ps. A clear eye pattern was obtained in 2.5 Gbps all-optical wavelength conversion. This is the first all-optical wavelength conversion demonstration with a bandgap-engineered PIC with either selective area growth or quantum-well intermixing techniques.

This paper presents high-performance millimeter-wave passive devices using MEMS technology. The purpose of this paper is to show the possibility of MEMS technology as an enabling technology for millimeter-waves. The loss and cost issues, which have been the inherent barrier for commercialization of mm-waves, can be solved by RF MEMS technology. Successful demonstrations of MEMS technology for mm-waves include novel CPW transmission lines, digital impedance tuners, analog tunable band-pass filters, reconfigurable low-pass filters, V-band digital distributed phase shifters and 2-D mechanical beam-steering antennas. All these circuits were implemented for 30-65 GHz frequency range, and show the state-of-the-art performance, which is beyond the limit set by the conventional technology.

In this paper, Morse code is selected as a communication adaptive device for persons whose hand coordination and dexterity are impaired by such ailments as amyotrophic lateral sclerosis, multiple sclerosis, muscular dystrophy, and other severe handicaps. Morse code is composed of a series of dots, dashes, and space intervals, and each element is transmitted by sending a signal for a defined length of time. A suitable adaptive automatic recognition method is needed for persons with disabilities due to their difficulty in maintaining a stable typing rate. To overcome this problem, the proposed method combines the support vector machines method with a variable degree variable step size LMS algorithm. The method is divided into five stages: tone recognition, space recognition, training process, adaptive processing, and character recognition. Statistical analyses demonstrated that the proposed method elicited a better recognition rate in comparison to alternative methods from the literature.

In the field of industrial manufacturing, visual pattern inspection is an important task to prevent the inclusion of incorrect parts. There have been demands for such methods able to handle factors caused by positional and rotational alignment, and illumination changes. In this paper, we propose a discrimination method called Trinarized broad-edge and Gray-scale Hybrid Matching (TGHM). The method is highly reliable due to gray-scale cross correlation which has a high pattern discrimination efficiency, with high-speed position and rotation alignment using the characteristics of trinarized broad-edge representation which has high data compressibility and illumination-resistant variability. In an example in which the method is applied to mis-collation inspection equipment of a bookbinding machine, it is confirmed that the processing speed is 24,000 sheets/hour, the error detection rate is 100.0%, and the mis-alarm rate is less than 0.002%, and it is verified that the method is practical.

Target word selection is one of the most important and difficult tasks in English-Korean Machine Translation. It effects on the overall translation accuracy of machine translation systems. In this paper, we present a new approach to Korean target word selection for an English noun with translation ambiguities using multiple knowledge such as verb frame patterns, sense vectors based on collocations, statistical Korean local context information and co-occurring POS information. Verb frame patterns constructed with dictionary and corpus play an important role in resolving the sparseness problem of collocation data. Sense vectors are a set of collocation data when an English word having target selection ambiguities is to be translated to specific Korean target word. Statistical Korean Local Context Information is an N-gram information generated using Korean corpus. The co-occurring POS information is a statistically significant POS clue which appears with ambiguous word. To evaluate our approach, we applied the method to Tellus-EK system, English-Korean automatic translation system currently developed at ETRI [1],[2]. The experiment showed promising results for diverse sentences from web documents.

In this paper, we present a Korean to Chinese/English/Japanese multilingual Machine Translation (MT) system of closed captions for Digital Television (DTV). Preliminary experiments of our closed caption translation with existing base MT systems had shown unsatisfactory result. In order to achieve more accurate translation with the base MT systems, we adopted live resources of multilingual Named Entities and their translingual equivalences from the Web. We also utilize the program information, which the terrestrial broadcasters offer through DTV transport stream, in order to use program specific dictionaries, including the names of characters, locations and organizations. Two more components are adopted for reducing the ambiguities of parsing and word sense disambiguation; sentence simplification for long sentence segmentation and dynamic domain identification for automatic domain dictionary stacking. With these integrated approaches, we could raise the Mean Opinion Score (MOS) of translation accuracy by 0.40 higher than the base MT systems.

This paper proposes a metric for example matching under the example-based machine translation. Our metric served as similarity measure is employed to retrieve the most similar examples to a given query. Basically it makes use of simple information such as lemma and part-of-speech information of typographically mismatched words. In addition, it uses the contiguity information of matched word units to catch the full context. Finally we show the results for the correctness of the proposed metric.

This paper deals with the performance evaluation of two types of dynamic change, called Flush and SCO (Synthetic Cut-Over), for state machine WF-nets. As an evaluation measure of dynamic change for marked graph WF-nets, change time has been used. We first generalize change time so as to apply it to dynamic change for state machine WF-nets. By using its maximum value, we evaluate the worst-case of dynamic change for state machine WF-nets. We call the maximum value as worst change time. Then under the same assumptions as our previous studies, we give methods of calculating worst change time of Flush and SCO dynamic changes. We also clarify the relation on worst change time between them. Finally we evaluate them by comparing the values of worst change time for an actual example of dynamic change.

For the measurement of the 3D surface of micro-solderballs in IC (Integrated Circuit) manufacturing inspection, a binary grating project lenses of high MTF (Modulation Transfer Function) with tilted project plane is designed in this paper. Using a combination of lenses and a tilted optical layout both on object and image plane, the wave-front aberrations are reduced and the nonlinear image distortion is corrected with nonlinearity compensation, This optical lens allows us to project the structured light pattern to the inspected objects efficiently for clear deformed coded imaging, it could be used to online measure 3D shape of micro-solderballs with high precision and accuracy.

This paper investigates the accepting powers of two-way alternating Turing machines (2ATM's) with only existential (universal) states which have inkdots and sublogarithmic space. It is shown that for sublogarithmic space-bounded computations, (i) multi-inkdot 2ATM's with only existential states and the ones with only universal states are incomparable, (ii) k-inkdot 2ATM's are better than k-inkdot 2ATM's with only existential (universal) states, k ≥ 0, and (iii) the class of sets accepted by multi-inkdot 2ATM's with only existential (universal) states is not closed under complementation.

This paper analyzes automation surprises in human-machine systems with time information. Automation surprises are phenomena such that the underlying machine's behavior diverges from user's intention and may lead to critical situations. Thus, designing human-machine systems without automation surprises is one of fundamental issues to achieve reliable user interaction with the machines. In this paper, we focus on timed human-machine interaction and address their formal aspects. The presented framework is essentially an extension of untimed human-machine interaction and will cover the previously proposed methodologies. We employ timed automata as a model of human-machine systems with time information. Modeling the human-machine systems as timed automata enables one to deal with not only discrete behavior but also time constraints. Then, by introducing the concept of timed simulation of the machine model and the user model, conditions which guarantee the nonexistence of automation surprises are derived. Finally, we construct a composite model in which a machine model and a user model evolve concurrently and show that automation surprises can be detected by solving a reachability problem in the composite model.

Energy efficiency of cache memories is crucial in designing embedded processors. Reducing energy consumption in the instruction cache is especially important, since the instruction cache consumes a significant portion of total processor energy. This paper proposes a new instruction cache architecture, named Partitioned Instruction Cache (PI-Cache), for reducing dynamic energy consumption in the instruction cache by partitioning it to smaller (less power-consuming) sub-caches. When the proposed PI-Cache is accessed, only one sub-cache is accessed by utilizing the temporal/spatial locality of applications. In the meantime, other sub-caches are not accessed, leading to dynamic energy reduction. The PI-Cache also reduces dynamic energy consumption by eliminating the energy consumed in tag lookup and comparison. Moreover, the performance gap between the conventional instruction cache and the proposed PI-Cache becomes little when the physical cache access time is considered. We evaluated the energy efficiency by running a cycle accurate simulator, SimpleScalar, with power parameters obtained from CACTI. Simulation results show that the PI-Cache improves the energy-delay product by 20%-54% compared to the conventional direct-mapped instruction cache.

A comparison of performances is made of three text-independent speaker identification methods based on dual Penalized Logistic Regression Machine (dPLRM), Support Vector Machine (SVM) and Gaussian Mixture Model (GMM) with experiments by 10 male speakers. The methods are compared for the speech data which were collected over the period of 13 months in 6 utterance-sessions of which the earlier 3 sessions were for obtaining training data of 12 seconds' utterances. Comparisons are made with the Mel-frequency cepstrum (MFC) data versus the log-power spectrum data and also with training data in a single session versus in plural ones. It is shown that dPLRM with the log-power spectrum data is competitive with SVM and GMM methods with MFC data, when trained for the combined data collected in the earlier three sessions. dPLRM outperforms GMM method especially as the amount of training data becomes smaller. Some of these findings have been already reported in [1]-[3].

There has been significant interest in developing new forms of acoustic model, in particular models which allow additional dependencies to be represented than those contained within a standard hidden Markov model (HMM). This paper discusses one such class of models, augmented statistical models. Here, a local exponential approximation is made about some point on a base model. This allows additional dependencies within the data to be modelled than are represented in the base distribution. Augmented models based on Gaussian mixture models (GMMs) and HMMs are briefly described. These augmented models are then related to generative kernels, one approach used for allowing support vector machines (SVMs) to be applied to variable length data. The training of augmented statistical models within an SVM, generative kernel, framework is then discussed. This may be viewed as using maximum margin training to estimate statistical models. Augmented Gaussian mixture models are then evaluated using rescoring on a large vocabulary speech recognition task.

Given the availability of high-speed Ethernet and HW based protocol offload, clustered systems using a commodity network fabric (e.g., TCP/IP over Ethernet) are expected to become more attractive for a range of e-business and data center applications. In this paper, we describe a comprehensive simulation to study the performance of clustered database systems using such a fabric. The simulation model currently supports both TCP and SCTP as the transport protocol and models an Oracle 9i like clustered DBMS running a TPC-C like workload. The model can be used to study a wide variety of issues regarding the performance of clustered DBMS systems including the impact of enhancements to network layers (transport, IP, MAC), QoS mechanisms or latency improvements, and cluster-wide power control issues.

Content-Addressable Network (CAN) provides a mechanism that could retrieve objects in a P2P network by maintaining indices to those objects in a fully decentralized manner. In the CAN system, index caching is a useful technique for reducing the response time of retrieving objects. The key points of effective caching techniques are to improve cache hit ratio by actively sharing caches distributed over the P2P network with every node and to reduce a maintenance and/or routing overhead for locating the cache of a requested index. In this paper, we propose a new caching technique based on the notion of proxy-type caching techniques which have been widely used in WWW systems. It can achieve active cache sharing by incorporating the concept of proxy caching into the index access mechanism and locate a closer proxy cache of a requested index with a little routing overhead. By the result of simulations, we conclude that it can improve the response time of retrieving indices by 30% compared with conventional caching techniques.