Estimating the generalization error is one of the key ingredients of supervised learning since a good generalization error estimator can be used for model selection. An unbiased generalization error estimator called the subspace information criterion (SIC) is shown to be useful for model selection, but its range of application is limited to linear learning methods. In this paper, we extend SIC to be applicable to non-linear learning.

This paper describes two different types of GaAs-HBT compatible, base-collector diode 0/20-dB step attenuators--diode-linearizer type and harmonics-trap type--for 3.5-GHz-band wireless applications. The two attenuators use an AC-coupled, stacked type diode switch topology featuring high power handling capability with low bias current operation. Compared to a conventional diode switch topology, this topology can improve the capability of more than 6 dB with the same bias current. In addition, successful incorporation of a shunt diode linearizer and second- and third-harmonic traps into the attenuators gives the IM3 distortion improvement of more than 7 dB in the high power ranging from 16 dBm to 18 dBm even in the 20-dB attenuation mode when IM3 distortion levels are basically easy to degrade. Measurement results show that both the attenuators are capable of delivering power handling capability (P0.2 dB) of more than 18 dBm with IM3 levels of less than -35 dBc at an 18-dBm input power while drawing low bias currents of 3.8 mA and 6.8 mA in the thru and attenuation modes from 0/5-V complementary supplies. Measured insertion losses of the linearizer-type and harmonics-trap type attenuators in the thru mode are as low as 1.4 dB and 2.5 dB, respectively.

To realize highly conductive and wear-durable thin films, we deposited metal doped carbon films onto silicon substrates by RF sputtering method. The dopant metals were various precious metals and transition metals. The electrical conductivity and wear durability of the deposited films were evaluated. We have found that Ir doping especially increased the electrical conductivity for the given amount of dopant metal. The wear durability of Ir-doped carbon films did not deteriorate even below a 7 at.% Ir concentration, and the conductivity of 7 at.% Ir-doped carbon was twenty times that of a non-doped carbon thin film.

Contact resistance is caused by constriction resistance and film resistance through contact layers. It is well known that a surface film causes non-linear voltage and current characteristics. The origin of non-linearity is caused by tunneling electron through thin insulation barrier or jumping over the thick barrier (Shottky barrier) on the contact surface. In this paper, a new idea causing nonlinear property by only current constriction which flows through very small contact spot area, if there is no film layer, is proposed by the two dimensional contact model. The contact model, used in this paper, is a two dimensional type narrow path of contact area (short bridge) made by thin copper foil of 0.035 mm on a glass epoxy resin board. The contact part is made by scraping with an electric drill as a single bridge shape of 0.1 mm wide and 0.3 mm long on the centre of a board (100 mm100 mm). The 3rd harmonic distortion voltage was measured by using a Component Linearity Test Equipment (Type CLT1 made by Radiometer Electronics Company) which the system supplies a pure sine wave current of 10 kHz and detects a distortion voltage of 30 kHz by a narrow band pass filter circuit. The sensitivity of the Component Linearity Test Equipment (CLT1) is under a 10-9 volt. Four bridge samples were examined for the comparison of nonlinear distortion voltage. The distortion voltage of a sample (A) (0.1 mm wide, 0.3 mm long) is too larger than the one of the sample (B) (0.2 mm wide, 0.3 mm long) at the same applied voltage which resistance is not so different each other. It seems that current constriction to the spot (A) may heat up higher and cool down lower than (B). It would be also guessed that the power dissipation of 20 kHz cause temperature oscillation of 20 kHz, then it causes a component of contact resistance of 20 kHz, and therefore the product of 10 kHz current and 20 kHz resistance component cause 30 kHz component distortion voltage.

A method for flexibly allocating and reallocating optical access network (OAN) resources, including fibers and equipment, using the constraint satisfaction problem (CSP) is described. OAN resource allocation during service delivery provisioning involves various input conditions and allocation sequences, so an OAN resource allocation method has to support various workflow patterns. Furthermore, exception processing, such as reallocating OAN resources once they are allocated, is inevitable, especially during the spread of service using optical fiber and during the deployment of an optical access network. However, it is almost impossible to describe all workflow patterns including exception processes. Improving the efficiency of these exception processes, as well as that of the typical processes, is important for reducing the service delivery time. Describing all these patterns and process flows increases development cost. The CSP can be used to search for solutions without having to fix the process sequence and input conditions beforehand. We have formulated the conditions for OAN resource allocation and reallocation as a CSP. Use of this method makes it possible to handle various allocation workflow patterns including exception processes. Evaluation of the solution search time demonstrated its feasibility.

In this letter, we discuss the problem of receive antenna selection in the downlink of multiuser multiple-input multiple-output (MIMO) systems with Tomlinson-Harashima precoding (THP), where the number of receivers is assumed equal to that of transmit antennas. Based on the criterion of maximum system sum-capacity, a per-layer receive antenna selection scheme is proposed. This scheme, which selects one receive antenna for each receiver, can well exploit the nonlinear and successive characteristics of THP. Two models are established for the proposed per-layer scheme and the conventional per-user scheme. Both the theoretical analysis and simulation results indicate that the proposed scheme can greatly improve the equivalent channel power gains and the system sum-capacity.

This paper presents a novel feature extraction algorithm based on particle swarms for processing hyperspectral imagery data. Particle swarm optimization, originally developed for global optimization over continuous spaces, is extended to deal with the problem of feature extraction. A formulation utilizing two swarms of particles was developed to optimize simultaneously a desired performance criterion and the number of selected features. Candidate feature sets were evaluated on a regression problem. Artificial neural networks were trained to construct linear and nonlinear models of chemical concentration of glucose in soybean crops. Experimental results utilizing real-world hyperspectral datasets demonstrate the viability of the method. The particle swarms-based approach presented superior performance in comparison with conventional feature extraction methods, on both linear and nonlinear models.

The possibility of using three kinds of new type composite materials as material for high speed sliding contacts was investigated. The results of this investigation were compared with the results of the low speed tests that were reported earlier. As a result of the above, it was discovered that for high speed rotation in the range from 0.014 m/s to 2 m/s, the order of merit did not significantly change. Based on this, it was concluded that if solid lubricant is effectively supplied to the sliding surface, the influence by frictional heat generated by high speed is slight. Of the three kinds of composite material, it was clarified that composite material (CMML-1) had the lowest contact resistance and Composite Material (CMML-3) had the lowest maximum frictional coefficient of friction. 'CM' and 'ML' are initialisms for 'Composite Material' and 'Material of Lubrication' respectively. The number that is attached to the material name is a numeric value that was set by this laboratory.

A press-fit connection is a solderless electrical connection technology, which utilizes the mechanical contact force generated between through-holes on a printed circuit board (PCB) and terminals with a width slightly larger than the through-hole diameter. This technology has been widely noted recently as a measure against the "Lead Free Requirement" of materials comprising electric/electronic devices, especially in the area of automobile connector. For the application of this technology to automobile connectors, we have to take into account the severe requirement, such as (1) the adaptation to wider through-hole diameter tolerance range and (2) the establishment of connection reliability for the various PCB surface treatments. As a result, we have determined the minimum and maximum contact forces satisfying the long term connection reliability and designed the terminal shape, which has been refined the N-shape cross section developed before, by using three dimensional finite element methods (FEM). Furthermore, we have developed a new type of hard tin plating on terminals, thus preventing the scraping-off of tin during the insertion process, that could result in a short-circuit on the PCB, for the Organic Solderability Preservative (OSP) treated PCB. The press-fit connector for the automobile airbag Electronic Control Units (ECUs) we developed has been able to transfer to the mass-production phase successfully from August 2005.

Recent theoretical and experimental studies indicate that spatial multiplexing (SM) systems have enormous potential for increasing the capacity of corresponding multiple input multiple output (MIMO) channels in rich scattering environments. In this paper, we propose a new recursion based algorithm for Bell Labs layered space time (BLAST) signal detection in SM systems. The new algorithm uses an inflated recursion in the initialization and a deflated recursion in the iteration stage: as a result, the complexity is greatly reduced and the irregularity issues are completely avoided. Compared with the conventional fastest recursive approach, the complexity of our proposal is lower by a factor of 2 and it is also very implementation friendly.

We propose a robust iterative multiuser receiver for decoding convolutional coded code-division multiple access (CDMA) signals in both Gaussian and non-Gaussian channel noise. The receiver is derived from a modified maximum a-posteriori (MAP) algorithm called the max-log-MAP algorithm for robustness against erroneous channel variance estimation. Furthermore, the effect of destructive outliers arising from impulsive noise is mitigated in the proposed receiver by incorporating the robust Huber penalty function into the multiuser detector. The proposed receiver is shown to perform satisfactorily over Gaussian and non-Gaussian impulsive channels. In every iteration, cumulative improvement in the quality of the a-posteriori probabilities is also demonstrated.

In this paper, we describe a two-phase method for biomedical named entity recognition consisting of term boundary detection and biomedical category labeling. The term boundary detection can be defined as a task to assign label sequences to a given sentence, and biomedical category labeling can be viewed as a local classification problem which does not need knowledge of the labels of other named entities in a sentence. The advantage of dividing the recognition process into two phases is that we can measure the effectiveness of models at each phase and select separately the appropriate model for each subtask. In order to obtain a better performance in biomedical named entity recognition, we conducted comparative experiments using several learning methods at each phase. Moreover, results by these machine learning based models are refined by rule-based postprocessing. We tested our methods on the JNLPBA 2004 shared task and the GENIA corpus.

In wireless ad hoc networks, network services are provided through the cooperation of all nodes. Albeit that good teamwork could smoothly run a mobile network, selfish node behaviors would probably cause it to break down. Some examples of these selfish node behaviors would include, "listening only" for saving energy or "receiving the valuable" without forwarding the packets to others. To cope with this problem, we propose PDM, a price-demand function based pricing model, to restrains the selfish behaviors of mobile nodes. PDM is based on the packet sending requirements of the source nodes and the forwarding cost of relay nods. Using this pricing methodology, the packet forwarding activities will be profitable for the relay node and further stimulate cooperation in the network. In particular, the new model enjoys the merit of giving relay nodes no reason to dishonestly report their forwarding costs, because an honest cost claim has proven to be an optimal strategy for relay nodes. Furthermore, our new model uses a price-demand function to reflect the relationship between the service demand of the source nodes and the service supply of the relay nodes. As a consequence, our approach reduces the source nodes' payments to send packets, and at the same time guarantees that the packets sent by the source nodes are delivered to the destination.

This paper proposes an OFDM mobile radio packet system that employs a new protocol of automatic repeat request (ARQ) for nonlinear multiuser detection (MUD) with log likelihood ratio combining (LC) on the appropriate bits. The conventional metric combining (MC) MUD separates collided packets by using nonlinear MUD, accumulates the Euclidian distance metrics of the received subcarrier symbols in the packets, and then achieves throughput improvement. However, when MC-MUD detects a packet error, it makes user terminals retransmit the same packets so as to reproduce the collision of the same packets. The proposed LC-MUD scheme simplifies the ARQ protocol and requires no reproduction of the same packet collision. The computer simulations demonstrate the superior throughput of LC-MUD to that of MC-MUD, and further improvement due to adaptive modulation and coding (AMC) optimized for the nonlinear MUD in LC-MUD.

To improve the channel estimation accuracy of multiple-input multiple-output (MIMO) multiplexing, we previously proposed iterative QR-decomposition with M-algorithm (QRD-M) with decision directed channel estimation. In this paper, to keep the computational complexity low while further improving the transmission performance, we will modify previously proposed iterative QRD-M by incorporating cyclic redundancy check (CRC) coding. In the proposed method, transmitted signals are ranked according to their results of CRC decoding and the received signal-to-interference plus noise power ratio (SINR). In the modified M-algorithm, since the results of Turbo decoding and CRC decoding are used to generate the surviving symbol replica, the accuracy of signal detection in the following steps can be improved. Furthermore, based on the results of CRC decoding, iterative process can be terminated before reaching the maximum allowable number of iterations. Computer simulation results show that the loss in the required average received signal energy per bit-to-noise power spectrum density ratio Eb/N0 for average packet error rate (PER) = 10-2 is only about 0.4 dB from maximum likelihood detection (Full MLD) with ideal channel estimation.

This paper proposes effective energy feature normalization methods for robust speech recognition in noisy environments. We first develop an energy subtraction method and a modified method for the Log-energy Dynamic Range Normalization (ERN) using inverse function. We then present the hybrid method combining the energy subtraction and the modified ERN. Using Aurora2.0 database for representative evaluations, a significant performance improvement over the ERN method is demonstrated.

This paper presents a new Fuzzy Logic Controller (FLC) having the ability to support rational-powered membership functions. These functions are extended forms of triangular/trapezoidal membership functions, and also those functions which are generated by applying linguistic hedges. A two-input, single-output, nine-rule Takagi-Sugeno-Kang (TSK) type FLC is designed in 0.35 µm standard CMOS technology. This controller can also be used as a standard (Mamdani) type FLC having singleton output membership functions, as well as a Linguistic Hedge FLC (LHFLC). Mixed analog/digital realization of the circuit makes the design programmable and extendable, while having relatively low power consumption. Current mode realization of the circuits leads to simple and intuitive configurations. For a particular set of programming parameters, simulation results of the controller using HSPICE simulator and level 49 parameters (BSIM3v3), show an average power consumption of 5 mW, and an RMS error of 1.32% compared to ideal results obtained from MATLAB software.

A simple, yet effective geometric method is presented to construct the signature sequences for multicarrier code-division multiple access (MC-CDMA) systems. By minimizing the correlation of the effective signature vectors, the signature sequences are recursively determined via projection onto a properly constructed subspace. Conducted simulations verify the effectiveness of the method.

We present a computationally efficient sequential detection scheme using a modified Fano algorithm (MFA) for V-BLAST systems. The proposed algorithm consists of the following three steps: initialization, tree searching, and optimal selection. In the first step, the proposed detection scheme chooses several candidate symbols at the tree level of one. Based on these symbols, the MFA then finds the remaining transmitted symbols from the second tree level in the original tree structure. Finally, an optimal symbol sequence is decided among the most likely candidate sequences searched in the previous step. Computer simulation shows that the proposed scheme yields significant saving in complexity with very small performance degradation compared with that of sphere detection (SD).

We previously proposed a rotation-spreading neural network (R-SAN net). This neural net can recognize the orientation of an object irrespective of its shape, and its shape irrespective of its orientation. The R-SAN net is suitable for orientation recognition of a concentric circular pattern such as an iris image. Previously, variations of ambient lighting conditions affected iris detection. In this study, we introduce brightness normalization for accuracy improvement of iris detection in various lighting conditions. Brightness normalization provides high accuracy iris extraction in severe lighting conditions. A recognition experiment investigated the characteristics of rotation and shape recognition for both learned and un-learned iris images in various plane rotations. The R-SAN net recognized the rotation angle for the learned iris images in arbitrary orientation, but not for un-learned iris images. Thus, the variation of the rotation angle was corrected only for learned irises, but not un-learned irises. Although the R-SAN net rightly recognized the learned irises, it could not completely reject the un-learned irises as unregistered irises. Using the specific orientation recognition characteristics of the R-SAN net, a minimum distance was introduced as a new shape recognition criterion for the R-SAN net. In consequence, the R-SAN net combined with the minimum distance rightly recognized both learned (registered) and un-learned irises; the unregistered irises were correctly rejected.