This paper proposes an image inpainting algorithm based on multiple linear models and matrix rank minimization. Several inpainting algorithms have been previously proposed based on the assumption that an image can be modeled using autoregressive (AR) models. However, these algorithms perform poorly when applied to natural photographs because they assume that an image is modeled by a position-invariant linear model with a fixed model order. In order to improve inpainting quality, this work introduces a multiple AR model and proposes an image inpainting algorithm based on multiple matrix rank minimization with sparse regularization. In doing so, a practical algorithm is provided based on the iterative partial matrix shrinkage algorithm, with numerical examples showing the effectiveness of the proposed algorithm.

Link prediction, the computational problem of determining whether there is a link between two objects, is important in machine learning and data mining. Feature-based link prediction, in which the feature vectors of the two objects are given, is of particular interest because it can also be used for various identification-related problems. Although the factorization machine and the higher-order factorization machine (HOFM) are widely used for feature-based link prediction, they use feature combinations not only across the two objects but also from the same object. Feature combinations from the same object are irrelevant to major link prediction problems such as predicting identity because using them increases computational cost and degrades accuracy. In this paper, we present novel models that use higher-order feature combinations only across the two objects. Since there were no algorithms for efficiently computing higher-order feature combinations only across two objects, we derive one by leveraging reported and newly obtained results of calculating the ANOVA kernel. We present an efficient coordinate descent algorithm for proposed models. We also improve the effectiveness of the existing one for the HOFM. Furthermore, we extend proposed models to a deep neural network. Experimental results demonstrated the effectiveness of our proposed models.

Pulse Pairs (PPs) generated by Distance Measure Equipment (DME) cause severe interference on L-band Digital Aeronautical Communication System type 1 (L-DACS1) which is based on Orthogonal Frequency Division Multiplexing (OFDM). In this paper, a novel and practical PP mitigation approach is proposed. Different from previous work, it adopts only time domain methods to mitigate interference, so it will not affect the subsequent signal processing in frequency domain. At the receiver side, the proposed approach can precisely reconstruct the deformed PPs (DPPs) which are often overlapped and have various parameters. Firstly, a filter bank and a correlation scheme are jointly used to detect non-overlapped DPPs, also a weighted average scheme is used to automatically measure the waveform of DPP. Secondly, based on the measured waveform, sparse estimation is used to estimate the precise positions of DPPs. Finally, the parameters of each DPP are estimated by a non-linear estimator. The key point of this step is, a piecewise linear model is used to approximate the non-linear carrier frequency of each DPP. Numerical simulations show that comparing with existing work, the proposed approach is more robust, closer to interference free environment and its Bit Error Rate is reduced by about 10dB.

We present the adaptation of the acoustic models of hidden Markov models (HMMs) to the target speaker and noise environment using bilinear models. Acoustic models trained from various speakers and noise conditions are decomposed to build the bases that capture the interaction between the two factors. The model for the target speaker and noise is represented as a product of bases and two weight vectors. In experiments using the AURORA4 corpus, the bilinear model outperforms the linear model.

This paper proposes a so called quasi-linear support vector machine (SVM), which is an SVM with a composite quasi-linear kernel. In the quasi-linear SVM model, the nonlinear separation hyperplane is approximated by multiple local linear models with interpolation. Instead of building multiple local SVM models separately, the quasi-linear SVM realizes the multi local linear model approach in the kernel level. That is, it is built exactly in the same way as a single SVM model, by composing a quasi-linear kernel. A guided partitioning method is proposed to obtain the local partitions for the composition of quasi-linear kernel function. Experiment results on artificial data and benchmark datasets show that the proposed method is effective and improves classification performances.

Large-signal-based nonlinear models are developed to analyze a variety of dynamic performances in a resonant tunneling diode (RTD) with peripheral circuits such as an integrated broad band bow-tie antenna, a bias circuit and a bias stabilizer circuit. Dynamic modes of the RTD are classified by the time-domain analysis with the model. On the basis of our model, we suggest a possibility to discuss a terahertz order oscillation mode control, and the ASK modulation in several tens Gbit/sec in the RTD with the broad band antenna. Validity of the model and analysis is shown by explaining measured results of modulated oscillation signals in fabricated triple-barrier RTDs.

This paper describes a technique for overcoming the model shrinkage problem in automatic speech recognition (ASR), which allows application developers and users to control the model size with less degradation of accuracy. Recently, models for ASR systems tend to be large and this can constitute a bottleneck for developers and users without special knowledge of ASR with respect to introducing the ASR function. Specifically, discriminative language models (DLMs) are usually designed in a high-dimensional parameter space, although DLMs have gained increasing attention as an approach for improving recognition accuracy. Our proposed method can be applied to linear models including DLMs, in which the score of an input sample is given by the inner product of its features and the model parameters, but our proposed method can shrink models in an easy computation by obtaining simple statistics, which are square sums of feature values appearing in a data set. Our experimental results show that our proposed method can shrink a DLM with little degradation in accuracy and perform properly whether or not the data for obtaining the statistics are the same as the data for training the model.

The single-hidden-layer feedforward neural networks (SLFNs) are frequently used in machine learning due to their ability which can form boundaries with arbitrary shapes if the activation function of hidden units is chosen properly. Most learning algorithms for the neural networks based on gradient descent are still slow because of the many learning steps. Recently, a learning algorithm called extreme learning machine (ELM) has been proposed for training SLFNs to overcome this problem. It randomly chooses the input weights and hidden-layer biases, and analytically determines the output weights by the matrix inverse operation. This algorithm can achieve good generalization performance with high learning speed in many applications. However, this algorithm often requires a large number of hidden units and takes long time for classification of new observations. In this paper, a new approach for training SLFNs called least-squares extreme learning machine (LS-ELM) is proposed. Unlike the gradient descent-based algorithms and the ELM, our approach analytically determines the input weights, hidden-layer biases and output weights based on linear models. For training with a large number of input patterns, an online training scheme with sub-blocks of the training set is also introduced. Experimental results for real applications show that our proposed algorithm offers high classification accuracy with a smaller number of hidden units and extremely high speed in both learning and testing.

We present a method for recognition of two-hand gestures. Two-hand gestures include fine-grain descriptions of hands under a complicated background, and have complex dynamic behaviors. Hence, assuming that two-hand gestures are an interacting process of two hands whose shapes and motions are described by switching linear dynamics, we propose a coupled switching linear dynamic model to capture interactions between both hands. The parameters of the model are learned via EM algorithm using approximate computations. Recognition is performed by selection of the model with maximum likelihood out of a few learned models during tracking. We confirmed the effectiveness of the proposed model in tracking and recognition of two-hand gestures through some experiments.

In this paper, we propose a nonlinear control model to characterize the AQM algorithm-GREEN. Based on this model, we analyze its performance and prove that there exists a stable oscillation when in equilibrium. Furthermore, we also investigate the effects of the factors such as bandwidth, round trip time, and load level on the amplitude and frequency of the oscillation. Theoretical analysis and simulation results indicate that GREEN algorithm is insensitive to the network conditions when the link rate and the round trip time are relatively small and becomes more sensitive to the change of network conditions when the bandwidth delay product is relatively high. For GREEN the adaptability to a wide range of network conditions is based on the compromising of the efficiency.

We present a method to track and recognize shape-changing hand gestures simultaneously. The switching linear model using active contour model well corresponds to temporal shapes and motions of hands. However, inference in the switching linear model is computationally intractable, and therefore the learning process cannot be performed via the exact EM (Expectation Maximization) algorithm. Thus, we present an approximate EM algorithm using a collapsing method in which some Gaussians are merged into a single Gaussian. Tracking is performed through the forward algorithm based on Kalman filtering and the collapsing method. We also present a regularized smoothing, which plays a role of reducing jump changes between the training sequences of shape vectors representing complex-variable hand shapes. The recognition process is performed by the selection of a model with the maximum likelihood from some trained models while tracking is being performed. Experiments for several shape-changing hand gestures are demonstrated.

This paper describes an improved nonlinear GaAs FET model and its parameter extraction procedure for almost all operating conditions such as the small-signal condition, the power saturated condition, and the controlled-resistance condition. The model is capable of modeling the gate voltage dependent drain current and its derivatives in the saturated region as well as the drain voltage dependent drain current and its derivatives in the linear region. The model can take into account the frequency dispersion effects of both transconductance and output conductance. The model describes forward conduction and reverse conduction currents. Deriving the capacitance part of the model from unique charge equations satisfies charge conservation. The model accurately predicts voltage-dependent S-parameters, spurious response in an active condition and inter-modulation response in the controlled-resistance condition of a GaAs FET.

This paper proposes an improved nonlinear FET model along with its parameter extraction procedure suitable for the accurate prediction of inter-modulation product's levels (IM) and spurious responses in active and passive applications. This new model allows accurate capture of the drain current behavior and its derivatives with respect to the gate voltage and the drain voltage in the both the saturated and linear regions of the I-V biasing domain. It was found that this model accurately predicts the bias-dependent S-parameters as well as IM's levels for both amplifier and mixer applications up to mm-wave frequencies.

The design and performance of a S-band solid-state power amplifier (SSPA) module are reported. The SSPA module consisted from four MMIC power amplifiers (PA) achieves 90 W output power, 35% power added efficiency (PAE), and 80% power combine efficiency. The MMIC PA achieves over 28 W output power, 25 dB small signal gain, and over 38% PAE. This paper also describes the large-signal circuit design for the MMIC PA using an improved nonlinear FET model developed for high power amplifier applications.

Intricate Speech Communication Mode (I-SC Mode) is observed in verbal interaction during ISDN-TV conferencing. It is characterized by conflicts and multiple interactions of speech. I-SC Mode might cause mental stress to participants and be obstacles for smooth communication. However, the reasons of I-SC Mode on the environment of information transmission are hitherto unknown. Furthermore, analyses on the talks inside a conference site (LT: local talk or a talk inside a local site) and between remote sites (MT: media talk or a talk between remote sites) are originally conceived on assumed differences in cognitive distance and media intimacy. This study deals with communication effects/barriers and cognitive distance/intimacy of media correlated with audio-video transmission signals and speech modes or talk types and response delay in human speech interactions by using an innovated conference model (decision-making transaction model: DT-Model) in synchronous ISDN-TV conference systems (SYN) and asynchronous ones (ASYN). The effects of intricate communication can be predicted to a certain extent and in some ways. In I-SC Mode, because a timely answer can not be received from recipients (or partner), response time delay and response rate are analyzed. These factors are thus analyzed with an innovated dynamic model, where the recognizable acceptance of delay is evaluated. The nonlinear model shows that the larger the response time delay, the lower the response rate becomes. Comparing the response rate between SYN and ASYN, the latter is notably lower than the former. This indicates that the communication efficiency is lower in ASYN. An I-SC Mode is the main mode that occurs during ASYN conferences, and this in turn causes psychological stress. Statistics show the prevalence of a high incidence of complicated plural talks and a low response rate exists as the main factors preventing smooth human-to-human communication. Furthermore, comparing the response delays in face-to-face LT (Tf) and machine-mediated MT (Tm), human communication delay is significantly extended by the effects of initial mechanical delays. Therefore, cognitive intimacy of media is clearly affected by the existence of physical distance.

In this paper we introduce the Piecewise Linear Radial Basis Function Model (PWL-RBFM), a new nonlinear model that uses the well known RBF framework to build a PWL functional approximation by combining an l1 norm with a linear RBF function. A smooth generalization of the PWL-RBF is proposed: it is obtained by substituting the modulus function with the logistic function. These models are applied to several time series prediction tasks.