This paper presents a novel framework to extract highly compact and discriminative features for face video retrieval tasks using the deep convolutional neural network (CNN). The face video retrieval task is to find the videos containing the face of a specific person from a database with a face image or a face video of the same person as a query. A key challenge is to extract discriminative features with small storage space from face videos with large intra-class variations caused by different angle, illumination, and facial expression. In recent years, the CNN-based binary hashing and metric learning methods showed notable progress in image/video retrieval tasks. However, the existing CNN-based binary hashing and metric learning have limitations in terms of inevitable information loss and storage inefficiency, respectively. To cope with these problems, the proposed framework consists of two parts: first, a novel loss function using a radial basis function kernel (RBF Loss) is introduced to train a neural network to generate compact and discriminative high-level features, and secondly, an optimized quantization using a logistic function (Logistic Quantization) is suggested to convert a real-valued feature to a 1-byte integer with the minimum information loss. Through the face video retrieval experiments on a challenging TV series data set (ICT-TV), it is demonstrated that the proposed framework outperforms the existing state-of-the-art feature extraction methods. Furthermore, the effectiveness of RBF loss was also demonstrated through the image classification and retrieval experiments on the CIFAR-10 and Fashion-MNIST data sets with LeNet-5.

IAM ABOUT to start my second year of service as the Editor-in-Chief (EiC) of the IEEE TRANSACTIONS ON NEURAL NETWORKS (TNN). Needless to say, my first year as the EiC has been full of excitement and challenges. Transitioning this position from my predecessor to me went very smoothly during the months of September 2009 to January 2010. During the past year, we have accumulated 50+ Associate Editors (AEs) handling roughly 600 new submissions (not counting resubmissions and revised submissions). With the help of these AEs and my predecessor, I was quickly able to learn to do my job, and as such, the transition had very few glitches. The easy part of my job is checking whether a submission is in compliance with our guidelines and where it is within the scope of the TRANSACTIONS, before it is assigned to an AE for handling. The difficult part of my job has been dealing with some papers with three or more reviewers, all of whom agreed to review them but for some reason failed to respond to repeated automatic-review reminders. AEs handling these papers have to take several extra steps to remind reviewers through phone calls or e-mails, look for replacement reviewers, or review the papers themselves. Most authors have been appreciative of the work of the AEs and reviewers, and they accept our decisions without a problem. The backlog of papers has been kept short over the last year. We have maintained an organized printing and paperacceptance schedule, with papers typically printed in the journal within 2‐3 months of acceptance. Our page budget has been kept constant in the past few years (roughly 2060 pages per year), and we expect to hold the same page count for next year.

The output of a classifier is usually determined by the value of a discriminant function and a decision is made based on this output which does not necessarily represent the posterior probability for the soft decision of classification. In this context, it is desirable that the output of a classifier be calibrated in such a way to give the meaning of the posterior probability of class membership. This paper presents a new method of postprocessing for the probabilistic scaling of classifier's output. For this purpose, the output of a classifier is analyzed and the distribution of the output is described by the beta distribution parameters. For more accurate approximation of class output distribution, the beta distribution parameters as well as the kernel parameters describing the discriminant function are adjusted in such a way to improve the uniformity of beta cumulative distribution function (CDF) values for the given class output samples. As a result, the classifier with the proposed scaling method referred to as the class probability output network (CPON) can provide accurate posterior probabilities for the soft decision of classification. To show the effectiveness of the proposed method, the simulation for pattern classification using the support vector machine (SVM) classifiers is performed for the University of California at Irvine (UCI) data sets. The simulation results using the SVM classifiers with the proposed CPON demonstrated a statistically meaningful performance improvement over the SVM and SVM-related classifiers, and also other probabilistic scaling methods.

This paper presents a new method of zero-crossing based binau-ral mask estimation for sound segregation under the condition that multiple sound sources are present simultaneously. The masking is determined by the estimated sound source direc-tions using the spatial cues such as inter-aural time differences (ITDs) and inter-aural intensity differences (IIDs). In the sug-gested method, the estimation of ITDs is utilizing the statistical properties of zero-crossings detected from binaural filter-bank outputs. We also consider the estimation of ITDs with the aid of IID samples to cope with the phase ambiguities of ITD samples in high frequencies. For the masking method, we consider to use the target-to-total power ratio in each segment of the time-frequency domain. We show that this power ratio is optimal from the view point of reconstructing the target speech signal. As a result, the proposed method is able to provide an accurate estimate of sound source directions and also a good masking scheme for speech segregation while offering significantly less computational complexity compared to cross-correlation-based methods. 1.

This paper presents a new method for estimating interaural time differences (ITDs) in noisy multisource environments. To estimate ITDs, conventional methods use cross-correlation of left and right instantaneous firing rates of neurons. However, these methods require high computational complexity involved in the computation of cross-correlation. Furthermore, they suffer from inaccuracies in estimating ITDs, especially in noisy multisource environments. In this context, we propose a new method using zero-crossing time differences (ZCTDs) detected from binaural filter-bank outputs in which a distinctive feature is to estimate the signal-to-noise ratios (SNRs) of filtered signals using the variances of ITD estimates. As a result, the proposed ZCTD method with SNR estimation is able to provide an accurate estimate of multiple sound source directions and is more robust to noise while offering significantly less computational complexity compared to the conventional cross-correlation based methods

This paper presents a new method of zero-crossing based binaural mask estimation for missing data speech recognition under the condition that multiple sound sources are present simultaneously. The masking is determined by the estimated directions of sound sources using the spatial cues such as inter-aural time differences (ITDs) and inter-aural intensity differences (IIDs). In the suggested method, the estimation of ITDs is utilizing the statistical properties of zero-crossings generated from binaural filter-bank outputs. We also consider the estimation of ITDs with the aid of IID samples to cope with the phase ambiguities of ITD samples in high frequencies. As a result, the proposed method is able to provide an accurate estimate of sound source directions and a good masking scheme for speech recognition while offering significantly less computational complexity compared to cross-correlation based methods.

The prediction risk estimation in nonlinear regres- sion models including artificial neural networks is especially important for problems with limited data since it can be used as a tool for finding the optimal model (or network architecture) minimizing the expected risk. In this paper, we suggest the prediction risk bounds of nonlinear regression models. The suggested bounds are derived from the modulus of continuity for a multivariate function. We also present the model selection criteria referred to as the modulus of continuity information criteria (MCIC) derived from the suggested prediction risk bounds. Through the simulation for function approximation, we have shown that the suggested MCIC is effective in nonlinear model selection problems with limited data.

This paper presents a new method of sound source localization based on zero-crossings generated from binaural filter-bank outputs. To detect the sound source direction, in the conventional methods the inter-aural time differences (ITDs) and the inter-aural intensity differences (IIDs) are estimated using the cross-correlations of neuronal firing rates. However, these methods require high computational complexity involved in the computation of cross-correlations and they suffer from inaccuracies in estimating the ITDs and IIDs, especially in a noisy multisource environment. In this context, we propose a method using zero-crossing and peak-amplitudes (ZCPAs) of binaural filter-bank outputs for accurate and efficient estimation of ITDs and IIDs. The proposed method is able to provide accurate estimation of the sound source direction and robustness to noise while offering significantly less computational complexity compared to cross-correlation based methods.

This paper presents a new form of true risk bounds for the regression of real-valued functions. The goal of machine learning is minimizing the true risk (or general error) for the whole distribution of sample space, not just a set of training samples. However, the true risk cannot be estimated accurately with the finite number of samples. In this sense, we derive the form of true risk bounds which may provide the useful guideline for the optimization of learning models. Through the simulation for the function approximation, we have shown that the prediction of true risk bounds based on the suggested functional norm is well fitted to the empirical data.

This paper presents a new model of auditory preprocessors based on tree structured filter-banks to provide the robustness to noise and the ease of hardware implementation. The robustness to noise is further improved by two components: one is the adaptive Q control of filter-bank in the sense of enhancing the contrast between frequency channels and another is the adaptive gain control of filter-bank outputs in the sense of enhancing the contrast between time frames. As a result, the proposed approach has shown the better performance of speech recognition in noisy environment compared to other approaches of auditory preprocessors. To show the effectiveness of our approach, the simulation for the English digit recognition of TI46-Word database has been performed.

As a preprocessing stage of time series prediction, an analysis of time series is an important issue since the structure of a prediction model such as delay time and embedding dimension which determine the window size, can greatly influence the performance of a prediction model. A new method of determining the optimum window size in the sense of the smoothness (or easiness) of mapping, which is defined by the given data, is suggested for the purpose of determining the structure of a nonlinear prediction model in order to be more faithfully identified to the given data. To show the effectiveness of our approach, the suggested method is applied to determining the optimum window size for the prediction of Mackey-Glass chaotic time series and analyzing ECG heart rate data.

This paper presents a new approach to an auditory model for robust speech recognition in noisy environments. The proposed model consists of cochlear bandpass filters and nonlinear operations in which frequency information of the signal is obtained by zero-crossing intervals. Intensity information is also incorporated by a peak detector and a compressive nonlinearity. The robustness of the zero-crossings in spectral estimation is verified by analyzing the variance of the level-crossing intervals as a function of the crossing level values. Compared with other auditory models, the proposed auditory model is computationally efficient, free from many unknown parameters, and able to serve as a robust front-end for speech recognition in noisy environments. Experimental results of speech recognition demonstrate the robustness of the proposed method in various types of noisy environments.

SPEECH RECOGNITION Doh-Suk Kimy, Jae-Hoon Jeongy, Soo-Young Leey, Rhee M. Kilz yDepartment of Electrical Engineering/ zDivision of Basic Science Korea Advanced Institute of Science and Technology 373-1 Kusong-dong, Yusong-gu, Taejon 305-701, Korea E-mail: [email protected] ABSTRACT Zero-crossings with peak amplitudes (ZCPA) model motivated by human auditory periphery is simple compared with other auditory models, but powerful speech analysis tool for robust speech recognition in noisy environments. In this paper, improvement in recognition rate of ZCPA model is addressed by incorporating time-derivative features with several di erent time-derivative window lengths. Experimental results show that ZCPA has relatively higher sensitivity to derivative window length than conventional feature extraction algorithms. Also, experimental comparisons with several front-ends including some auditorylike schemes in real-world noisy environments demonstrate the robustness of ZCPA model. ZCPA model shows superior performance compared with other frontends especially in noisy condition corrupted by white Gaussian noise.

This paper presents a method for solving inverse mapping of a continuous function learned by a multilayer feedforward mapping network. The method is based on the iterative update of input vector toward a solution, while escaping from local minima. The input vector update is determined by the pseudo-inverse of the gradient of Lyapunov function, and, should an optimal solution be searched for, the projection of the gradient of a performance index on the null space of the gradient of Lyapunov function. The update rule is allowed to detect an input vector approaching local minima through a phenomenon called "update explosion". At or near local minima, the input vector is guided by an escape trajectory generated based on "global information", where global information is referred to here as predefined or known information on forward mapping; or the input vector is relocated to a new position based on the probability density function (PDF) constructed over the input vector space by Parzen estimate. The constructed PDF reflects the history of local minima detected during the search process, and represents the probability that a particular input vector can lead to a solution based on the update rule. The proposed method has a substantial advantage in computational complexity as well as convergence property over the conventional methods based on Jacobian pseudo-inverse or Jacobian transpose.

The authors present a novel method of accomplishing robot kinematic control based on a bidirectional mapping neural network (BMNN). The BMNN constructed is composed of a multilayer feedforward network with hidden units having sinusoidal activation functions and a feedback network forming a recurrent loop around the feedforward network. The feedforward network can be trained to accurately represent the forward kinematic equations of a robot arm. The feedback network iteratively generates joint-angle updates based on a Lyapunov function to modify the current joint angles in such a way that the output of the forward network converges to the desired Cartesian position and orientation. The proposed BMNN offers the following advantages over the conventional approaches: (1) the accurate computation of robot forward and inverse kinematic solutions with simple training; (2) the ability to handle one-too-many inverse mapping required for redundant arm kinematics solutions; and (3) the automatic generation of arm trajectories. Simulation results are shown