Changkyu Choi

Reducing bit-widths of activations and weights of deep networks makes it efficient to compute and store them in memory, which is crucial in their deployments to resource-limited devices, such as mobile phones. However, decreasing bit-widths with quantization generally yields drastically degraded accuracy. To tackle this problem, we propose to learn to quantize activations and weights via a trainable quantizer that transforms and discretizes them. Specifically, we parameterize the quantization intervals and obtain their optimal values by directly minimizing the task loss of the network. This quantization-interval-learning (QIL) allows the quantized networks to maintain the accuracy of the full-precision (32-bit) networks with bit-width as low as 4-bit and minimize the accuracy degeneration with further bit-width reduction (i.e., 3 and 2-bit). Moreover, our quantizer can be trained on a heterogeneous dataset, and thus can be used to quantize pretrained networks without access to their training data. We demonstrate the effectiveness of our trainable quantizer on ImageNet dataset with various network architectures such as ResNet-18, -34 and AlexNet, on which it outperforms existing methods to achieve the state-of-the-art accuracy.

Face recognition under viewpoint and illumination changes is a difficult problem, so many researchers have tried to solve this problem by producing the pose- and illumination- invariant feature. Zhu et al. [26] changed all arbitrary pose and illumination images to the frontal view image to use for the invariant feature. In this scheme, preserving identity while rotating pose image is a crucial issue. This paper proposes a new deep architecture based on a novel type of multitask learning, which can achieve superior performance in rotating to a target-pose face image from an arbitrary pose and illumination image while preserving identity. The target pose can be controlled by the user's intention. This novel type of multi-task model significantly improves identity preservation over the single task model. By using all the synthesized controlled pose images, called Controlled Pose Image (CPI), for the pose-illumination-invariant feature and voting among the multiple face recognition results, we clearly outperform the state-of-the-art algorithms by more than 4~6% on the MultiPIE dataset.

Using cross-country panel data, we found evidence that the Internet plays a positive and significant role in economic growth after investment ratio, government consumption ratio, and inflation were used as control variables in the growth equation.

A doubling of Internet usage in a country turned out to lead to a 2 to 4% increase in services trade. An increase in a country's Internet access will facilitate an increase in its service trade with other countries.

Abstract Using pooled Gini coefficient 1993 to 2002 data for 119 countries from World Development Indicators 2004, World Bank, we find that income inequality, defined as the Gini coefficient, increases as FDI stocks as a percentage of GDP increase. Increases in per capita GDP and real per capita GDP growth rate reduce income inequality in a country, whereas an increase in GDP deteriorates income distribution. Furthermore, Latin American and Caribbean countries proved to have a less equal income distribution. Notes 1 ASIA includes China, Hong Kong, Indonesia, Japan, Korea, Malaysia, the Philippines, Singapore and Thailand. 2 Per capita GDP squared is added to the independent variables, but Kuznets’ ‘inverted-U curve’ hypothesis, that inequality has an inverted-U curve relationship with development (Kuznets, Citation1955; Tsai, Citation1995; Thornton, Citation2001), does not hold from our analysis. 3 I also added the product of rich country dummy (per capita GDP > 20 000 dollars) with each type of INTENSITY variable, the literacy rate, trade openness and general government final consumption expenditures as a percentage of GDP from WDI as independent variables. I could not find any significant results in relation to the Gini coefficient when these variables were added and thus those estimation results are not reported in Table 2. 4 This result is contradictory to Tsai (Citation1995), where the effect of FDI on income inequality becomes invalid when Asia and Latin America dummies are included.

Of late, deep neural networks have become ubiquitous in mobile applications. As mobile devices generally require immediate response while maintaining user privacy, the demand for on-device machine learning technology is on the increase. Nevertheless, mobile devices suffer from restricted hardware resources, whereas deep neural networks involve considerable computation and communication. Therefore, the implementation of a neural-network specialized hardware accelerator, generally called neural processing unit (NPU), has started to gain attention for the mobile application processor (AP). However, NPUs for commercial mobile AP face two challenges that are difficult to realize simultaneously: execution of a wide range of applications and efficient performance.In this paper, we propose a flexible but efficient NPU architecture for a Samsung flagship mobile system-on-chip (SoC). To implement an efficient NPU, we design an energy-efficient inner-product engine that utilizes the input feature map sparsity. We propose a re-configurable MAC array to enhance the flexibility of the proposed NPU, dynamic internal memory port assignment to maximize on-chip memory bandwidth utilization, and efficient architecture to support mixed-precision arithmetic. We implement the proposed NPU using the Samsung 5nm library. Our silicon measurement experiments demonstrate that the proposed NPU achieves 290.7 FPS and 13.6 TOPS/W, when executing an 8-bit quantized Inception-v3 model [1] with a single NPU core. In addition, we analyze the proposed zero-skipping architecture in detail. Finally, we present the findings and lessons learned when implementing the commercial mobile NPU and interesting avenues for future work.

It has been suggested that drift loss to the magnetopause can be one of the major loss mechanisms contributing to relativistic electron flux dropouts. In this study, we examine details of relativistic electrons' drift physics to determine the extent to which the drift loss through the magnetopause is important to the total loss of the outer radiation belt. We have numerically computed drift paths of relativistic electrons' guiding center for various pitch angles, various measurement positions, and different solar wind conditions using the Tsyganenko T02 model. We specifically demonstrate how the drift loss effect depends on these various parameters. Most importantly, we present various estimates of relative changes of the omnidirectional flux of 1 MeV electrons between two different solar wind conditions based on a simple form of the directional flux function. For a change of the dynamic pressure from 4 nPa to 10 nPa with a fixed IMF B Z = 0 nT, our estimate indicates that after this increase in pressure, the equatorial omnidirectional flux at midnight near geosynchronous altitude decreases by ∼56 to ∼97%, depending on the specific pitch angle dependence of the directional flux. The effect rapidly decreases at regions earthward of geosynchronous orbit and shows a general trend of decrease away from midnight. For a change of the IMF B Z from 0 nT to −15 nT with a fixed dynamic pressure of 4 nPa, the relative decrease of the omnidirectional flux at geosynchronous altitude on the nightside is much smaller than that for the pressure increase, but its effect becomes substantial only beyond geosynchronous orbit. Possibilities exist that our results may change to some extent for a different magnetospheric model than the one used here.

Advanced High-Strength steels (AHSS) are increasingly used in industry. Thus, the springback prediction in bending of AHSS is important to maintain close geometric tolerances in deformed parts. The unique properties of AHSS: a non-constant Young׳s modulus and a stress–strain relation which does not follow a simple flow stress equation make the prediction of springback very difficult. In this study, an analytical model was developed to predict the springback in air-bending of AHSS, considering the special properties of these materials. A computer code which is developed based on classical bending theory was updated based on the new model. The finite element simulation which includes the properties of the AHSS is also presented in the paper. The comparison between the experimental results and predictions indicates that the detailed consideration of the properties of AHSS affects the accuracy of the springback prediction with the analytical method. It is hoped that this new analysis can be incorporated into the controller of a press brake to adjust the machine for compensating springback during bending.

As the expressive depth of an emotional face differs with individuals or expressions, recognizing an expression using a single facial image at a moment is difficult. A relative expression of a query face compared to a reference face might alleviate this difficulty. In this paper, we propose to utilize contrastive representation that embeds a distinctive expressive factor for a discriminative purpose. The contrastive representation is calculated at the embedding layer of deep networks by comparing a given (query) image with the reference image. We attempt to utilize a generative reference image that is estimated based on the given image. Consequently, we deploy deep neural networks that embed a combination of a generative model, a contrastive model, and a discriminative model with an end-to-end training manner. In our proposed networks, we attempt to disentangle a facial expressive factor in two steps including learning of a generator network and a contrastive encoder network. We conducted extensive experiments on publicly available face expression databases (CK+, MMI, Oulu-CASIA, and in-the-wild databases) that have been widely adopted in the recent literatures. The proposed method outperforms the known state-of-the art methods in terms of the recognition accuracy.

We examine the effect of the Internet on the relationship between R&D expenditure and economic growth. Data for 105 countries over the period 1994–2014 are used for panel data analysis. The effect of R&D expenditure on economic growth proves to be affected positively by the Internet and the effect of the Internet on the economic growth is positively strengthened by an increase in R&D expenditure.

This paper studies the effect of the Internet on the volume of inward foreign direct investment (FDI). The Internet is assumed to induce more FDI by improving productivity. Using bilateral FDI data from 14 source countries and 53 host countries, cross-country empirical regressions based on a gravity FDI equation are performed. We found by ordinary least-squares and weighted least-squares analysis that when the number of the Internet hosts or users in a host country increased by 10%, FDI inflows increased by more than 2%.

In warm forging, die life is affected by abrasive wear and plastic deformation and may be shortened considerably due to thermal softening of the die surface caused by forging temperature and pressure. In this study, a methodology is developed for estimating abrasive die wear and plastic deformation in a warm forging operation, using a tempering parameter. This methodology consists of: (a) determination of the steady state die temperatures using multiple FE simulations, (b) using the calculated temperatures to predict the plastic deformation of the dies, (c) measuring the surface profile of the worn dies using a Coordinate Measuring Machine (CMM), (d) identifying the wear profiles caused by abrasive wear and plastic deformation, and (e) determining the abrasive wear parameters that can be used for future estimation of die wear in warm forging. The predictions have been compared with experimental results and it was concluded that this method can be used for estimation of die failure (wear and plastic deformation) in hot/warm forging.

In this paper, we present a novel attentive and immersive user interface based on gaze and hand gestures for interactive large-scale displays. The combination of gaze and hand gestures provide more interesting and immersive ways to manipulate 3D information.

Accurate age estimation from a facial image is quite challenging, since physical age and apparent age can be quite different, and this difference is dependent on gender, ethnicity, and many other factors. Multitask deep learning is one of the approach to improve age estimation by employing auxiliary tasks, such as gender recognition, that are related to the primary task. However, in traditional multitask learning for age estimation, the relationship between the primary and auxiliary tasks is difficult to describe; how the auxiliary tasks enhance the model for the primary objective is ambiguous. In this letter, we propose a conditional multitask learning method that architecturally factorizes an age variable into gender-conditioned age probabilities in a deep neural network. The lack of accurate training labels with discrete age values is another critical limitation to training age estimation models. Therefore, we propose a label expansion method that increases the number of accurate labels from weakly supervised categorical labels. To verify the generality of the proposed method, we perform intensive experiments on the publicly available MORPH-II and FG-NET datasets. The proposed methods outperform state-of-the art methods in both age estimation and gender recognition accuracy. These performance gains are verified on well-known deep network architectures-VGG-16, CASIA-WebFace, and Alexnet-to confirm the proposed methods generality.

Gaze estimation for ordinary smart phone, e.g. estimating where the user is looking at on the phone screen, can be applied in various applications. However, the widely used appearance-based CNN methods still have two issues for practical adoption. First, due to the limited dataset, gaze estimation is very likely to suffer from over-fitting, leading to poor accuracy at run time. Second, the current methods are usually not robust, i.e. their prediction results having notable jitters even when the user is performing gaze fixation, which degrades user experience greatly. For the first issue, we propose a new tolerant and talented (TAT) training scheme, which is an iterative random knowledge distillation framework enhanced with cosine similarity pruning and aligned orthogonal initialization. The knowledge distillation is a tolerant teaching process providing diverse and informative supervision. The enhanced pruning and initialization is a talented learning process prompting the network to escape from the local minima and re-born from a better start. For the second issue, we define a new metric to measure the robustness of gaze estimator, and propose an adversarial training based Disturbance with Ordinal loss (DwO) method to improve it. The experimental results show that our TAT method achieves state-of-the-art performance on GazeCapture dataset, and that our DwO method improves the robustness while keeping comparable accuracy.

Abstract The role of foreign direct investment (FDI) in the convergence of income level and growth has been investigated by panel data regressions. Bilateral FDI data from OECD from 1982 to 1997 is used. Income level and growth gaps between source and host countries turn out to decrease as bilateral FDI increases. It is also found that geographical closeness and common language play an important role in convergence in income level and growth. Acknowledgements I am grateful to Shang-Jin Wei for providing bilateral FDI data for this research and to participants in Macroeconomics Research Group Seminar in Korea. All remaining errors are mine. This paper was supported by Myongji University, 2003 Research Grant.

In this article, we examine the short-term impact of COVID-19 on daily stock market performance in Korea. We find strong evidence of asymmetric effects: although an increase in the number of confirmed cases negatively affected stock returns, returns were unaffected by a decline in the number of cases. We also found that the pandemic’s second wave further decreased stock returns, particularly in the food & beverage sector. Finally, we confirm that a rising number of confirmed cases increases the volatility of the returns. As Korea entered the second and third waves of infection, however, such effects diminished. Policymakers may find our empirical findings useful in their efforts to bolster financial stability, especially when an outbreak of COVID-19 erupts or a COVID-19 style event reoccurs in the future.

The hypothesis that the Internet improves productivity and thus will reduce inflation is tested by pooled OLS and random effects model using cross-country panel data from 1991 to 2000. We found that when the ratio of the Internet users to total population increases by 1%, the inflation drops by 0.04264% point to 0.13193% point.

Abstract We extend Portes et al. (2001) by introducing the Internet as a variable, and we test the model empirically by using cross-country panel data on portfolio flows between the United States and other countries from 1990 to 2008. Asymmetric information accounts for the strong negative relationship between international asset transactions and distance. The Internet plays an important role in mitigating information asymmetry between countries and increases the volume of cross-border portfolio flows.

Metric learning has attracted significant attentions in the past decades, for the appealing advances in various realworld applications such as person re-identification and face recognition.Traditional supervised metric learning attempts to seek a discriminative metric, which could minimize the pairwise distance of within-class data samples, while maximizing the pairwise distance of data samples from various classes.However, it is still a challenge to build a robust and discriminative metric, especially for corrupted data in the real-world application.In this paper, we propose a Robust Discriminative Metric Learning algorithm (RDML) via fast low-rank representation and denoising strategy.To be specific, the metric learning problem is guided by a discriminative regularization by incorporating the pair-wise or class-wise information.Moreover, low-rank basis learning is jointly optimized with the metric to better uncover the global data structure and remove noise.Furthermore, fast low-rank representation is implemented to mitigate the computational burden and make sure the scalability on large-scale datasets.Finally, we evaluate our learned metric on several challenging tasks, e.g., face recognition/verification, object recognition, and image clustering.The experimental results verify the effectiveness of the proposed algorithm by comparing to many metric learning algorithms, even deep learning ones.

Metric learning has attracted increasing attentions recently, because of its promising performance in many visual analysis applications. General supervised metric learning methods are designed to learn a discriminative metric that can pull all the within-class data points close enough, while pushing all the data points with different class labels far away. In this paper, we propose a Discriminative Low-rank Metric Learning method (DLML), where the metric matrix and data representation coefficients are both constrained to be low-rank. Therefore, our approach can not only dig out the redundant features with a low-rank metric, but also discover the global data structure by seeking a low-rank representation. Furthermore, we introduce a supervised regularizer to preserve more discriminative information. Different from traditional metric learning methods, our approach aims to seek low-rank metric matrix and low-rank representation in a discriminative low-dimensional subspace at the same time. Two scenarios of experiments, (e.g. face verification and face identification) are conducted to evaluate our algorithm. Experimental results on two challenging face datasets, e.g. CMU-PIE face dataset and Labeled Faces in the Wild (LFW), reveal the effectiveness of our proposed method by comparing with other metric learning algorithms.

Abstract Anisotropic velocity distributions of protons have long been considered as free energy sources for exciting electromagnetic ion cyclotron (EMIC) waves in the Earth's magnetosphere. Here we rigorously calculated the proton anisotropy parameter using proton data obtained from Van Allen Probe‐A observations. The calculations are performed for times during EMIC wave events (distinguishing the times immediately after and before EMIC wave onsets) and for times exhibiting no EMIC waves. We find that the anisotropy values are often larger immediately after EMIC wave onsets than the times just before EMIC wave onsets and the non‐EMIC wave times. The increase in anisotropy immediately after the EMIC wave onsets is rather small but discernible, such that the average increase is by ~15% relative to the anisotropy values during the non‐EMIC wave times and ~8% compared to those just before the EMIC wave onsets. Based on the calculated anisotropy values, we test the criterion for ion cyclotron instability suggested by Kennel and Petschek (1966, https://doi.org/10.1029/JZ071i001p00001 ) by applying it to the EMIC wave events. We find that despite the weak increase in anisotropy, the majority of the EMIC wave events satisfy the instability criterion. We suggest that the proton distributions often remain close to the marginal state to ion cyclotron instability, and consequently, the proton anisotropy values should often be observed near threshold values for ion cyclotron instability. Additionally, we demonstrate the usefulness and limitation of the instability criteria expressed in the form of an inverse relation between the anisotropy and plasma beta.

Linder posed a hypothesis in 1961 that the closer the preference structure between two countries is, the bigger the trade volume becomes. The empirical results using pooled trade data from 63 countries for 1970, 1980, 1990, and 1992 are in support of the Linder hypothesis. It is also found that the Linder hypothesis gained strength in the 1990s. Recent globalization may have strengthened the Linder hypothesis.

A tilt-controlled photo browsing method for small mobile devices is presented. The implementation uses continuous inputs from an accelerometer, and a multimodal (visual, audio and vibrotactile) display coupled with the states of this model. The model is based on a simple physical model, with its characteristics shaped to enhance usability. We show how the dynamics of the physical model can be shaped to make the handling qualities of the mobile device fit the browsing task. We implemented the proposed algorithm on Samsung MITs PDA with tri-axis accelerometer and a vibrotactile motor. The experiment used seven novice users browsing from 100 photos. We compare a tilt-based interaction method with a button-based browser and an iPod wheel. We discuss the usability performance and contrast this with subjective experience from the users. The iPod wheel has significantly poorer performance than button pushing or tilt interaction, despite its commercial popularity.

Reducing bit-widths of activations and weights of deep networks makes it efficient to compute and store them in memory, which is crucial in their deployments to resource-limited devices, such as mobile phones. However, decreasing bit-widths with quantization generally yields drastically degraded accuracy. To tackle this problem, we propose to learn to quantize activations and weights via a trainable quantizer that transforms and discretizes them. Specifically, we parameterize the quantization intervals and obtain their optimal values by directly minimizing the task loss of the network. This quantization-interval-learning (QIL) allows the quantized networks to maintain the accuracy of the full-precision (32-bit) networks with bit-width as low as 4-bit and minimize the accuracy degeneration with further bit-width reduction (i.e., 3 and 2-bit). Moreover, our quantizer can be trained on a heterogeneous dataset, and thus can be used to quantize pretrained networks without access to their training data. We demonstrate the effectiveness of our trainable quantizer on ImageNet dataset with various network architectures such as ResNet-18, -34 and AlexNet, on which it outperforms existing methods to achieve the state-of-the-art accuracy.

Reducing bit-widths of activations and weights of deep networks makes it efficient to compute and store them in memory, which is crucial in their deployments to resource-limited devices, such as mobile phones. However, decreasing bit-widths with quantization generally yields drastically degraded accuracy. To tackle this problem, we propose to learn to quantize activations and weights via a trainable quantizer that transforms and discretizes them. Specifically, we parameterize the quantization intervals and obtain their optimal values by directly minimizing the task loss of the network. This quantization-interval-learning (QIL) allows the quantized networks to maintain the accuracy of the full-precision (32-bit) networks with bit-width as low as 4-bit and minimize the accuracy degeneration with further bit-width reduction (i.e., 3 and 2-bit). Moreover, our quantizer can be trained on a heterogeneous dataset, and thus can be used to quantize pretrained networks without access to their training data. We demonstrate the effectiveness of our trainable quantizer on ImageNet dataset with various network architectures such as ResNet-18, -34 and AlexNet, on which it outperforms existing methods to achieve the state-of-the-art accuracy.

Abstract Electromagnetic ion cyclotron (EMIC) waves have long been considered major plasma waves that can cause the atmospheric precipitation of electrons. In this work, we perform test particle calculations of the interactions between relativistic electrons and EMIC waves with oblique wave normal angle (WNA) and noncircular polarization. We demonstrate advective and diffusive changes in electron pitch angle (PA) under a uniform background magnetic field for broad ranges of kinetic energy (KE), PA, WNA, and ellipticity. First, we find that the direction of advective PA changes can be of either sign, being positive (negative) for a lower (higher) initial PA. Second, an oblique WNA can enhance the advection effects on PA changes, particularly for KE ~2.5–4 MeV and a broad range of initial PAs. In contrast, a more linear polarization can decrease advection for a broad range of KE and initial PAs. Third, phase space trajectories depend strongly on specific values of WNA, ellipticity, KE, and PA. Sometimes a phase‐averaged PA change is a net result of a time‐dependent mixture of the phase trapping, untrapping, and bunching effects. Lastly, the overall effect of finite WNA and ellipticity on diffusive PA changes is less systematic than on advection. Nevertheless, we identify a range of KE and PA where the diffusion tends to increase for a larger WNA and decrease for a more linear polarization. The results in this work emphasize a strong necessity for incorporating precise information on WNA and ellipticity into evaluations of electron scattering by EMIC waves.

This study analyzes the relationship between the proportion of the economically active population aged 15–64 to total population and housing prices. A panel of 31 provinces in China from 2002 to 2014 is used in our analysis. We find empirical evidence that the impact of the population structure on housing-price growth increases as the population growth rates rise. This observation suggests that, to understand provincial housing price movements in China, one should consider the ratio of working-age population to total population in a province. The main policy implication is that Chinese policymakers need to ensure a moderated population growth to effectively promote stability in housing prices and the economy.

This article examines the link between the population structure and housing prices. We use a panel of 23 countries from 1976 to 2013 in our empirical analysis. We find statistically significant impacts of the proportion of the economically active population aged 15–64 to the total population on housing-price growth. Our study supports a policy for stable population growth to moderate housing-price growth and economic cycles.

Speech signal contains intrinsic and extrinsic variations such as accent, emotion, dialect, phoneme, speaking manner, noise, music, and reverberation. Some of these variations are unnecessary and are unspecified factors of variation. These factors lead to increased variability in speaker representation. In this paper, we assume that unspecified factors of variation exist in speaker representations, and we attempt to minimize variability in speaker representation. The key idea is that a primal speaker representation can be decomposed into orthogonal vectors and these vectors are recombined by using deep neural networks (DNN) to reduce speaker representation variability, yielding performance improvement for speaker verification (SV). The experimental results show that our proposed approach produces a relative equal error rate (EER) reduction of 47.1% compared to the use of the same convolutional neural network (CNN) architecture on the Vox-Celeb dataset. Furthermore, our proposed method provides significant improvement for short utterances.

Abstract Electromagnetic ion cyclotron (EMIC) waves can cause relativistic electron scattering and atmospheric precipitation, primarily via cyclotron resonant interactions in the Earth's radiation belts. However, the conventional quasilinear resonance theory suggests that the cyclotron resonance condition is not satisfied for 90° pitch angle (PA) electrons, which constitute the majority of electrons in the outer radiation belt, such that scattering mainly affects low‐PA electrons. In contrast to this theory, using test particle calculations, we demonstrate that even exactly 90° PA electrons can be significantly scattered by large‐amplitude EMIC waves. The finite wave force results in the parallel transport of 90° PA electrons away from the equator, corresponding to intrinsically nonresonant scattering. This can lead to parallel velocity that meets cyclotron resonance conditions as local PA deviates from 90°. Different types of resonance are identified depending on the wave normal angle, that is, first‐ and second‐order resonances for parallel and oblique waves, respectively.

Effects of dust charges on the nonlinear ion acoustic solitary waves in a fully relativistic dusty plasma for both cases of negative and positive dusts are numerically studied based on the pseudopotential method. In the presence of dusty particles, it is found that various types of nonlinear acoustic waves exist in forms which can be viewed as sequential combinations of three kinds of elementary solitary waves: bump, dip, and kink-type solitary waves. The number and the sequence of the constituent elementary solitary waves in a given nonlinear waves depend more sensitively on dust particle density than any other parameters. For negatively charged dust particles of low density, the nonlinear wave is in the shape of bumpy solitary wave. For a somewhat higher density, the wave changes into a form which can be viewed as a combination of bump and dip-type solitary waves. As the density is increased further, a more complex nonlinear wave composed of bump, kink, and dip-type solitary waves emerges. For a much higher density of dust particle, the nonlinear wave can have a shape that can be considered as a combination of bump and kink-type solitary waves. For the case of positively charged dust particles, two kinds of nonlinear waves can exist: bump-type solitary wave and a combination of bump and kink solitary waves. For both cases of negative and positive dust particles, it is found that single dip-type solitary wave does not exist. It is also found that as dust particle density increases, the signature of the elementary waves becomes less prominent.

In this study, the estimation of shear force for blind shear ram type blowout preventer was investigated by using Finite Element Method (FEM).So, the effect of the blowout preventer working condition on shear force requirement for shear operation could be accurately approximated by simulating the entire process, and ram geometry could be optimized to reduce force and energy used to shear the tube by plastic deformation.The results of FEM analyzes was compared with blowout preventer manufacturer shear force information.Comparisons show that forces evaluated by using FEM (Deform 3D) simulations provided fairly accurate results for actual shear force.Also, it was found that by using Finite Element simulations the effect of the blowout preventer working condition on shearing operation can be estimated and ram geometries can be optimized.Therefore, FEM analyses could be used to design more reliable and efficient ram type blowout preventers.

The 2007/2008 US financial crisis is related to the securitization of mortgage loans and the housing-price boom and bust. In this article, we test the hypothesis that housing-price change is related to the development of the financial system. Using panel data for 23 countries from 1988 to 2012, we have found that the housing-price growth rate increases as the financial system moves a bank orientation to a market orientation. The policy implication is that the government should beware sudden increases in the capital market relative to the banking sector. Especially, more sophisticated financial supervision with respect to housing-price movement is required when a bank-based financial system progresses quickly to a market-oriented financial system.

Face Parsing assigns every pixel in a facial image with a semantic label, which could be applied in various applications including face recognition, facial beautification, affective computing and animation. While lots of progress have been made in this field, current state-of-the-art methods still fail to extract real effective feature and restore accurate score map, especially for those facial parts which have large variations of deformation and fairly similar appearance, e.g. mouth, eyes and thin eyebrows. In this paper, we propose a novel pixel-wise face parsing method called Residual Encoder Decoder Network (RED-Net), which combines a feature-rich encoder-decoder framework with adaptive prior mechanism. Our encoder-decoder framework extracts feature with ResNet and decodes the feature by elaborately fusing the residual architectures in to deconvolution. This framework learns more effective feature comparing to that learnt by decoding with interpolation or classic deconvolution operations. To overcome the appearance ambiguity between facial parts, an adaptive prior mechanism is proposed in term of the decoder prediction confidence, allowing refining the final result. The experimental results on two public datasets demonstrate that our method outperforms the state-of-the-arts significantly, achieving improvements of F-measure from 0.854 to 0.905 on Helen dataset, and pixel accuracy from 95.12% to 97.59% on the LFW dataset. In particular, convincing qualitative examples show that our method parses eye, eyebrow, and lip regins more accurately.

Softmax-based learning methods have shown state-of-the-art performances on large-scale face recognition tasks. In this paper, we discover an important issue of softmax-based approaches: the sample features around the corresponding class weight are similarly penalized in the training phase even though their directions are different from each other. This directional discrepancy, i.e., process discrepancy leads to performance degradation at the evaluation phase. To mitigate the issue, we propose a novel training scheme, called minimum discrepancy learning that enforces directions of intra-class sample features to be aligned toward an optimal direction by using a single learnable basis. Furthermore, the single learnable basis facilitates disentangling the so-called class-invariant vectors from sample features, such that they are effective to train under class-imbalanced datasets.

In this paper, an adaptive microphone array system with a two-stage adaptation mode controller (AMC) is proposed for high-quality speech acquisition in real environments. The proposed system includes an adaptive array algorithm, a time-delay estimator and a newly proposed AMC. To ensure proper adaptation of the adaptive array algorithm, the proposed AMC uses not only temporal information, but also spatial information. The proposed AMC is constructed with two processing stages: an initialization stage and a running stage. In the initialization stage, a sound source localization technique is adopted, and a signal correlation characteristic is used in the running stage. For the adaptive array algorithm, a generalized sidelobe canceller with an adaptive blocking matrix is used. The proposed algorithm is implemented as a real-time man-machine interface module of a home-agent robot. Simulation results show 13 dB SINR improvement with the speaker sitting 2 m distance from the home-agent robot. The speech recognition rate is also enhanced by 32% when compared to the single channel acquisition system.

Speech recognition using circular microphone array is addressed in this paper. Eight microphones are located around the service robot to form a 2D microphone array. To enhance the speech quality, a novel adaptive beamformer composed of a delay-and-sum beamformer, adaptive blocking filters (ABFs) and adaptive cancelling filters (ACFs) is proposed. While the adaptive generalized sidelobe canceller (AGSC) connects the ABF and the ACF in feedforward, the proposed adaptive beamformer has them in feedback. The advantages of the proposed structure are the robustness to the steering vector errors and cross-talks and the reduced number of filter taps that gives the same speech quality compared to the AGSC with a larger number of filter taps. The experimental results show that the proposed structure is superior to the AGSC in objective and subjective evaluations. Speech recognition result shows that the proposed robust adaptive beamformer guarantees the recognition performance even in a low SNR and highly reverberant environment.

Through panel-data regressions, we found that both per capita income level and growth turn out to converge when the trade intensity ratio increases between the countries. Geographical proximity and language similarities also turn out to be associated with convergence in both income level and growth.

The paper presents a novel user interface combining bare hands and the line of sight (LoS) by using a depth camera from far distance without any handheld devices; as well as a 3D GUI providing both stereoscopy and motion parallax for smart displays. The proposed user interface provides a precise and convenient manipulation which is applicable to browsing thousands of channels andor media files. Especially, the combined interaction methods of the two modalities achieve 120(x) × 70(y) × 5(z) manipulation resolution. And then various user tasks were performed so as to assess the proposed user interface.

Forming of AHSS creates several challenges because these materials have higher strength and lower formability compared to low carbon steels. One of these challenges is springback that leads to dimensional inaccuracy in the formed part. In the present study, the effect of unloading apparent modulus (E- modulus) variation with strain on the accuracy of springback prediction in V-die bending and U-bending of DP 780 is investigated. A reliable methodology to measure springback which is very important is developed. Load-unload tensile tests were performed to obtain unloading apparent modulus variation. Springback in V-die bending and U bending was estimated by using FEA and a variable E-modulus. Compared with experimental data, the predictions gave reasonably accurate results.

The goal of Multi-label learning is to predict multiple labels of each single instance. This is a challenging problem since the training data is limited, long-tail label distribution, and complicated label correlations. Generally, more training samples and label correlation knowledge would benefit the learning performance. However, it is difficult to obtain large-scale well-labeled datasets, and building such a label correlation map requires sophisticated semantic knowledge. To this end, we propose an end-to-end Generative Correlation Discovery Network (GCDN) method for multi-label learning in this paper. GCDN captures the existing data distribution, and synthesizes diverse data to enlarge the diversity of the training features; meanwhile, it also learns the label correlations based on a specifically-designed, simple but effective correlation discovery network to automatically discover the label correlations and considerately improve the label prediction accuracy. Extensive experiments on several benchmarks are provided to demonstrate the effectiveness, efficiency, and high accuracy of our approach.

In this paper, we use the new classification of exchange-rate arrangements developed by Reinhart and Rogoff (2004) to test whether reserve holdings decrease with increasing exchange-rate flexibility. Using pooled data for 127 countries over the period 1980-2000, we find several new results. First, the degree of exchange-rate flexibility has an inverted-U relationship with the countryi?½i?½s reserve holdings. Exchange-rate regimes with intermediate flexibility need more reserves than polar regimes (hard pegs and freely floating). Second, reserve holdings are smaller under hard pegs than under freely floating, implying that current large stockpiles of reserves in East Asian countries can be significantly reduced if they adopt a single currency. Finally, per capita GDP and reserve holdings have an inverted-U relationship, too, reflecting that their correlation would be negative for industrial countries, but positive for developing countries.

Abstract In this paper, we describe a novel multi‐touch LCD display architecture with hover sensing capability based on infrared sensor array. The proposed display architecture has an advantage over existing multi‐touch LCD displays in that it maintains a slim form‐factor of LCD displays without loss of a display quality while it is possible to sense multiple touches and hovers simultaneously.

A classical bidirectional associative memory (BAM) suffers from low storage capacity and abundance of spurious memories though it has the properties of good generalization and noise immunity. In this paper, Hamming distance in recall procedure of usual asymmetrical BAM is replaced with modified Hamming distance by introducing weighting matrix into connection matrix. This generalization is validated to increase storage capacity, to lessen spurious memories, to enhance noise immunity, and to enable multiple association using simulation work.

Abstract In this paper, we describe a novel LCD display system with depth‐sensing capability based on coded aperture imaging. By invisibly displaying coded aperture patterns in an LCD screen, the depth of an object can be estimated from refocused images at various depths with reconstructed multi‐view images.

Purpose There have been many studies on the euro's impact on trade volume, foreign direct investment (FDI) and the integration of European financial markets. Previous research has tried to find empirical evidence for convergence of real estate securities markets. However, less attention has been paid to the euro's effect on FDI in the real estate industry. The purpose of this paper, therefore, is to analyze the euro's effect on FDI in the real estate industry between Germany and European partner countries. Design/methodology/approach It is hypothesised that the adoption of the euro will increase the volume of FDI flows in the real estate industry between Germany and European partner countries. To estimate the euro's effect on FDI in the real estate industry, a modified gravity equation is adopted. Pooled OLS and random effects models are utilised. Findings Results from panel data from 34 countries between 1986 and 2009 suggest that the euro contributed to the increase in the German bilateral FDI in the real estate industry to and from European partner countries. However, it is interesting that the euro's effects were only significant in FDI inflows under a random effects model. Originality/value The paper's findings provide original evidence for the positive impact of the euro on FDI in the real estate industry between Germany and European partner countries.

Generation of coherent ion acoustic solitary waves (IASWs) in inhomogeneous plasmas by an odd eigenmode (OEM) of electron holes (EHs) is investigated using 1D electrostatic particle-in-cell (PIC) simulations. The OEM oscillates at a frequency comparable to the trapped electron bouncing frequency, as also demonstrated by Lewis' theoretical formalism about the linear eigenmode in Bernstein-Greene-Kruskal (BGK) equilibrium. The density gradient in the inhomogeneous plasmas causes asymmetry in the EH potential structure associated with the OEM, whose amplitude grows rapidly as it propagates through the density gradient region. As the ions interact with this asymmetric potential, which oscillates slowly enough for the ions to respond, they are ejected to the lower density side with a larger potential amplitude, forming a chain of IASWs coherently with the oscillation of the OEM.

Deploying deep convolutional neural networks on ultra-low power systems is challenging due to the extremely limited resources. Especially, the memory becomes a bottleneck as the systems put a hard limit on the size of on-chip memory. Because peak memory explosion in the lower layers is critical even in tiny models, the size of an input image should be reduced with sacrifice in accuracy. To overcome this drawback, we propose a novel Raster-Scanning Network, named RaScaNet, inspired by raster-scanning in image sensors. RaScaNet reads only a few rows of pixels at a time using a convolutional neural network and then sequentially learns the representation of the whole image using a recurrent neural network. The proposed method operates on an ultra-low power system without input size reduction; it requires 15.9–24.3× smaller peak memory and 5.3–12.9× smaller weight memory than the state-of-the-art tiny models. Moreover, RaScaNet fully exploits on-chip SRAM and cache memory of the system as the sum of the peak memory and the weight memory does not exceed 60 KB, improving the power efficiency of the system. In our experiments, we demonstrate the binary classification performance of RaScaNet on Visual Wake Words and Pascal VOC datasets.

A dynamical local path-planning algorithm of an autonomous mobile robot available for moving obstacle avoidance as well as stationary obstacle avoidance using artificial pressure and nonlinear friction is described. The dynamical path-planning algorithm is considered to adequately accommodate the mobile robot to a dynamic situation of a path-planning nature. Artificial pressure is just a conceptual idea and a mimicry of the real physical pressure. It can be thought of as a density gradient in the neighborhood of the mobile robot. Together with the previous virtual force field (VFF) method, the path of the mobile robot is a solution of a path-planning equation. Local minima problems in stationary environments are solved by introducing nonlinear friction into the chaotic neuron. Due to the nonlinear friction, the proposed path-planner reveals chaotic dynamics in some parameter regions. This new path-planner is feasible in guiding, on real-time, the mobile robot to avoid stationary obstacles and reach the goal. Computer simulations are presented to show the effectiveness of the proposed algorithm.

Abstract This paper proposes a method which analyzes an expression from a facial image using a three‐dimensional facial model and then extracts the facial expression. First, the head motion and the facial actions (such as those of eyebrows, eyes, and lips) are separated from the facial image. This is realized by estimating the three‐dimensional motion of the face based on the three‐dimensional facial model and by compensating the motion. Next, the expression information is extracted from the separated facial actions in two ways. One is the method to extract successively the facial expressions considering the characteristics of the facial actions based on the facial muscles. The other is the method to estimate the facial expression as a whole using the least‐square method and regarding the facial actions by the facial muscles as a vector. Those methods are combined with the expression synthesis rules. This makes it possible to reconstruct the original expression from the extracted facial expression parameters. Finally, the result of the analysis of the facial expression from the actual image is compared to the result of evaluation by a psychologist to demonstrate the usefulness of the proposed method. The image reconstructed from the result of analysis also is compared with the original image.

Despite an abundance of research outcomes of blind source separation (BSS) in many types of simulated environments, their performances are still not satisfiable to apply to the real environments. The major obstacle may seem the finite filter length of the assumed mixing model and the nonlinear sensor noises. This paper presents a two-step speech enhancement method with stereo microphone inputs. The first step performs a frequency-domain BSS algorithm with no prior knowledge of the mixed source signals and generates stereo outputs. The second step further removes the remaining cross-channel interference by a spectral cancellation approach using a probabilistic source absence/presence detection technique. The desired primary source is detected every frame of the signal, and the secondary source is estimated in the powerspectral domain using the other BSS output as a reference interference source. Then the secondary source is subtracted to remove the cross-channel interference. Our experimental results show good separation enhancement performances on the real recordings of speech and music signals compared to the conventional BSS methods.

In this paper, we propose a novel mobile LCD device for simultaneous touch and hover sensing. A transparent infrared image sensor panel allows the device to display contents and to detect touch and hover actions alternately. The proposed system has an advantage over touch screen devices in that a user interacts with a mobile device by not only touching the display screen but also hovering over the display screen.

Ion acoustic solitary wave in a plasma consisting of electrons and ions with an external magnetic field is reinvestigated using the Sagdeev's potential method. Although the Sagdeev potential has a singularity for n < 1, where n is the ion number density, we obtain new solitary wave solutions by expanding the Sagdeev potential up to <TEX>${\delta}n^4$</TEX> near n = 1. They are compressiv (rarefactive) waves and shock type solitary waves. These waves can exist all together as a superposed wave which may be used to explain what would be observed in the solar wind plasma. We compared our theoretical results with the data of the Freja satellite in the study of Wu et al. (1996). Also it is shown that these solitary waves propagate with a subsonic speed.

Abstract In this paper, we propose a novel OLED display system by adding depth‐sensing capability to the display system. Coded apertures formed in OLED pixels with transmissive windows encode the radiation from a scene and the scene is reconstructed by an appropriate decoding pattern. Further, depth information is estimated from multiple coded images. In order to prove the feasibility of the proposed system, a 19” imaging system with transmissive windows within pixels was constructed and tested. Experimental results confirm that the proposed system can reconstruct the scene and accurately estimate the depth of an object in front of the display system.

Binaural blind source separation algorithm for noisy mixtures is proposed. We consider ambient background noise signals and nonstationary target source signals. The proposed blind source separation combines signal estimation from noisy observations with source identication through mixing parameter estimation. The sparseness property of target source signals enables the proposed noisy, underdetermined, binaural blind source separation. A minimum mean square error estimator in frequency domain is implemented to estimate the signal spectra from noisy observations. The K-means clustering algorithm is utilized to identify the sources. With the help of calculating the signal absence probability for each frequency bin, noises are effectively eliminated from the target source signals and the mixing parameter estimation becomes more accurate in noisy environments.

AbstractThe parameters of a chaotic system play an important role, whose variation will lead to completely different dynamics. Sometimes, we want to design a controller which is optimal in a certain sense. In this chapter, we focus on two kinds of methods of suppressing chaos: the adaptive control method and the inverse optimal control method. We develop two new methods of parametric adaptive control for a class of discrete-time chaotic systems and a class of continuous-time chaotic systems with multiple parameters, respectively. The systems are first assumed to be linear with respect to parameters. Then, systems with nonlinear distributed parameters and uncertain noise are considered. Finally, we apply the inverse optimal control method to stabilize a four-dimensional chaotic system.KeywordsFeedback ControlChaotic SystemAdaptive ControlReference ModelExternal DisturbanceThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

This paper investigates nonlinear evolutions of large amplitude oblique whistler waves (LOWWs) and the interaction with electrons using one-dimensional electromagnetic kinetic simulations. The present research is motivated by recent studies about the nonlinear phenomena of LOWWs. When the propagation angle is not close to the resonance cone angle, the trapping of electrons in the electric potential of LOWWs leads to a moderate damping and a mild acceleration of the electrons via the O'Neil-type damping. In contrast, when the propagation angle of LOWWs is close to the resonance cone angle, the LOWWs undergo a heavy damping accompanied by the stochastic thermalization of the electrons, especially in the perpendicular direction. It is found that the stochastic parameter S, defined by S=16k∥2(eme)Φ0 |J0(k⊥ρ)|ωce2, is a crucial factor determining the damping process. This result demonstrates the importance of self-consistent electron kinetic effects, which are not included in the previous single-particle or fluid approach. The implications of the present findings are discussed.

The human face includes a great deal of information related to individual features. If it were possible to describe facial images by means of parameters, one could think of a variety of applications. In this paper, wireframe models are applied to facial images, and orthogonal facial bases are found using principal-component analysis, while shape-related data and texture-related data are treated separately. By means of these facial bases, effective parametric description of facial images is achieved. The proposed method supports not only analysis of facial images but also image synthesis using obtained parameters. © 1998 Scripta Technica, Syst Comp Jpn, 28(10): 77–83, 1997

Genetic algorithtiis are getting more popular nowadadvs because of their sirtiplicity and robustness. Genetic algorithm are global search techniques for optimizations and many other problems. A feed-forward neural network that is widely used in control applications usually learns by back propagation afgorithm(BP). However, when there exist certain constraints, BP cannot be applied. We apply a genetic algorithtti to such a case. To ittiprove hill-climbing capability and speed up the convergence, we propose a tnodified genetic algorifhnr(i2fGA). The validity and efficiency of the proposed algorithtti, hlG.4 are shown by various sitnulation examples of systetti identification and nonlinear svsteiti control such as cart-pole system and robot manipulators

The paper proposes a novel recognition system for hand shapes at a distance for smart TV. Two types of hand shapes are selected for the needs of TV browsing. The proposed method provides robust recognition performance on various hand orientations and guarantees real-time computation.

This paper proposes a novel algorithm for robust finger blob detection in a multi-touch tabletop display where an array of multiple cameras is placed behind the display screen. The proposed algorithm reliably separates true finger touch blobs from false blobs introduced by palms and hovering hands in the presence of non-uniform illumination. The algorithm is based on majority quadrant search (MQS) which is designed to find a hand direction which leads to the removal of false blobs. Experimental results show that the proposed method reliably extract true finger touch blobs by rejecting the false blobs from a real interactive tabletop display with a multi-camera set-up.

In real-world applications, a single instance could have more than one label. To solve this task, multi-label learning methods emerged in recent years. It is a more challenging problem for many reasons, such as complex label correlation, long-tail label distribution, and data shortage. In general, overcoming these challenges and bettering learning performance could be achieved by utilizing more training samples and including label correlations. However, these solutions are expensive and inflexible. Large-scale, well-labeled datasets are difficult to obtain, and building label correlation maps requires task-specific semantic information as prior knowledge. To address these limitations, we propose a general and compact Multi-Label Correlation Learning (MUCO) framework. MUCO explicitly and effectively learns the latent label correlations by updating a label correlation tensor, which provides highly accurate and interpretable prediction results. In addition, a multi-label generative strategy is deployed to handle the long-tail label distribution challenge. It borrows the visual clues from limited samples and synthesizes more diverse samples. All networks in our model are optimized simultaneously. Extensive experiments illustrate the effectiveness and efficiency of MUCO. Ablation studies further prove the effectiveness of all the modules.

This study investigated the short-term effects of partisan control by state governments on employment during the 2020 COVID-19 pandemic in the United States. Specifically, we examined whether Republican control of the state government was associated with lower unemployment rates and higher employment-to-population ratios. Our results revealed that party control exerted a weak effect but that the interaction with the number of pandemic-related deaths had a strong negative (positive) impact on unemployment rates (employment-to-population ratios). The moderation effect of our state partisan control variables supports the conclusion that Republican-led states produced better employment outcomes against the pandemic.

In home environments, demands for a robot to serve a user are on the increase, such as cleaning rooms, bringing something to the user, and so on. To achieve these tasks, it is essential for developing a natural way of human-robot interaction (HRI). One of the most natural ways is that the robot approaches the user to do some tasks after recognizing the user's call and localizing its position. In this case, user localization becomes a key technology. In this paper, we propose a novel audio visual user localization system. It consists of a microphone array with eight sensors and a video camera. Estimating calling direction is achieved by the spectral subtraction of the spatial spectra. In particular, a novel beam forming method is proposed to suppress the nonstationary audio noises where they always occur in a real world. Furthermore, a robust method for face detection is proposed to double check the user based on an Adaboost classifier. It is improved to reduce the false alarms remarkably through a new postprocessing on face candidates. Successful results in a real home environment show its efficacy and feasibility. The implementation issues, limitations, and their possible solutions are also discussed.

This paper examines the importance of portfolio-flow volatility as a determinant of the demand for intemational reserves over the 1980-99 period. Using panel data. we find that portfolio-flow volatility significantly raises the level of reserve holdings. Especially reserve accumulation is most sensitive to the volatility of portfolio balance (net flows). Capital account liberalization has increased uncertainty in the world economy, thereby making open economies more vulnerable to international financial crises. The regression results imply that monetary authorities have accumulated more precautionary reserve balances against increased uncertainty in portfolio flows as capital account Iiberalization progresses. As in previous studies, real openness is an important explanatory factor in determining the demand for reserves.

Speaker segmentation is an important task in multi-party conversations. Overlapping speech poses a serious problem in segmenting audio into speaker turns. We propose an audio-visual speech separation system consisting of an array microphone with eight sensors and an omnidirectional color camera. Multiple concurrent speeches are segmented by fusing the two heterogeneous sensors. Each segmented speech is further enhanced by a linearly constrained minimum variance beamformer. Regardless of co-existing wide-band sound sources and pictures of human in a reverberant environment the proposed system effectively separates multiple target speeches.

A new face detector with very high accuracy and real-time speed for mobile phone is introduced. The method achieves the fastest speed and the highest accuracy compared with other similar methods. A series of ideas are proposed in order to accelerate detection speed of the traditional Adaboost detector. First of all, the threshold for weak classifier is learned based on a new multiple instance pruning method regarding not only positive samples but also negative samples, by which, weak classifier is able to reject background more efficiently. Then, a coarse-to-fine scan is applied. Coarse scan is used to find possible face location, and the fine scan refines the face location and rejects false alarms. We further improve the speed of multi-scale face detection by introducing two different template sizes for detector training. By which, the smaller faces can be rapidly detected and the high performance is kept for larger faces. The proposed method is evaluated on public dataset FDDB, the result shows competitive performance against all Adaboost based methods. The method has been implemented on mobile phone and the speed is superior to all competitors.

This paper describes a dynamical local path-planning algorithm of an autonomous mobile robot available for stationary obstacle avoidance using nonlinear friction. Dynamical path-planning algorithm is considered to accommodate the mobile robot to the dynamic situation of the path-planning nature. Together with the previous virtual force field method, the path of the mobile robot is a solution of a path-planning equation. Local minima problems in stationary environments are solved by introducing nonlinear friction into the chaotic neuron. Because of the nonlinear friction, the proposed path-planner reveals chaotic dynamics in some parameter regions. This new path-planner is feasible to guide, in real-time, the mobile robot to avoid stationary obstacles and to reach the goal. Computer simulations are presented to show the effectiveness of the proposed algorithm.

Abstract— In this paper, a novel multi‐touch LCD architecture with hover‐sensing capability is described. To detect multiple touch points and hover points simultaneously, a sensitive backlight, which is a backlight integrated with an IR sensor array, is introduced. The sensitive backlight uses visible light to display contents on a display screen and is also used to detect reflected IR light from objects on or near the display screen. The captured image from the sensitive backlight is used to extract touch and hover information. The proposed display architecture maintains the slim form factor of an LCD with no loss of display quality, while making it possible to sense multiple touches and hovers simultaneously.

In this paper, we propose an efficient synthetic refocusing method from multiple coded aperture images for 3D user interaction. The proposed method is applied to a flat panel display with a sensor panel which forms lens-less multi-view cameras. To capture the scene in front of the display, the modified uniformly redundant arrays (MURA) patterns are displayed on the LCD screen without the backlight. Through the imaging patterns on the LCD screen, MURA coded images are captured in the sensor panel. Instead of decoding all coded images to synthetically generate a refocused image, the proposed method only decodes one coded image corresponding to the refocusing image at a certain distance after circularly shifting and averaging all coded images. Further, based on the proposed refocusing method, the depth of an object in front of the display is estimated by finding the most focused image for each pixel through a stack of the refocused images at different depth levels. Experimental results show that the proposed method captures an object in front of the display, generates refocused images at different depth levels, and accurately determines the depth of an object including real human hands near the display

In this paper, we propose a novel depth estimation method from multiple coded apertures for 3D interaction. A flat panel display is transformed into lens-less multi-view cameras which consist of multiple coded apertures. The sensor panel behind the display captures the scene in front of the display through the imaging pattern of the modified uniformly redundant arrays (MURA) on the display panel. To estimate the depth of an object in the scene, we first generate a stack of synthetically refocused images at various distances by using the shifting and averaging approach for the captured coded images. And then, an initial depth map is obtained by applying a focus operator to a stack of the refocused images for each pixel. Finally, the depth is refined by fitting a parametric focus model to the response curves near the initial depth estimates. To demonstrate the effectiveness of the proposed algorithm, we construct an imaging system to capture the scene in front of the display. The system consists of a display screen and an x-ray detector without a scintillator layer so as to act as a visible sensor panel. Experimental results confirm that the proposed method accurately determines the depth of an object including a human hand in front of the display by capturing multiple MURA coded images, generating refocused images at different depth levels, and refining the initial depth estimates.

Abstract : Armoured vehicles are primarily designed to provide protection against blast and ballistic events, however there is pressure to reduce the weight of vehicles in order to achieve improvements in range and maneouvrability combined with reductions in operating cost. To obtain improved performance without compromising blast and ballistic properties, a range of tougher, lighter and harder materials have been investigated. This work looks at the possible use of ultra-fine grain (UFG) materials to achieve the desired properties. From a fundamental viewpoint, as indicated by the Hall-Petch equation, UFG is an ideal means for hardening and strengthening a metal without changing its chemical composition and without compromising ductility. The work described here is a microstructural investigation of TWIP steel that has been subjected to blast loading. It is found that the pre-blast technique reduced the grain size of the TWIP steel significantly and that the dominant deformation mechanism of the grain refined material was dislocation slips. The reduction in grain size resulted in a considerable increase in material hardness.

Using GMM estimation with the US data from January 1967 to April 2003, the precautionary saving hypothesis is tested using time-varying consumption uncertainty and a nonexpected-utility model of intertemporal optimal consumption. Overidentifying restrictions of the model specification are also tested for both expected and nonexpected utility using Hansen's J-statistics. It was found that the precautionary saving hypothesis did not hold under expected-utility preferences but did hold partly under nonexpected-utility preferences.

Magnetic flux ropes, often observed during intervals of interplanetary coronal mass ejections, have long been recognized to be critical in space weather. In this work, we focus on magnetic flux rope structure but on a much smaller scale, and not necessarily related to interplanetary coronal mass ejections. Using near-Earth solar wind advanced composition explorer (ACE) observations from 1998 to 2016, we identified a total of 309 small-scale magnetic flux ropes (SMFRs). We compared the characteristics of identified SMFR events with those of normal magnetic cloud (MC) events available from the existing literature. First, most of the MCs and SMFRs have similar values of accompanying solar wind speed and proton densities. However, the average magnetic field intensity of SMFRs is weaker (~7.4 nT) than that of MCs (~10.6 nT). Also, the average duration time and expansion speed of SMFRs are ~2.5 hr and 2.6 km/s, respectively, both of which are smaller by a factor of ~10 than those of MCs. In addition, we examined the geoeffectiveness of SMFR events by checking their correlation with magnetic storms and substorms. Based on the criteria Sym-H &lt; -50 nT (for identification of storm occurrence) and AL &lt; -200 nT (for identification of substorm occurrence), we found that for 88 SMFR events (corresponding to 28.5 % of the total SMFR events), substorms occurred after the impact of SMFRs, implying a possible triggering of substorms by SMFRs. In contrast, we found only two SMFRs that triggered storms. We emphasize that, based on a much larger database than used in previous studies, all these previously known features are now firmly confirmed by the current work. Accordingly, the results emphasize the significance of SMFRs from the viewpoint of possible triggering of substorms.

In the progress towards automatic cancer pre-screening, it is essential to achieve successful image segmentation prior to cancer pre-screening. A medical image segmenter system, which obtains local information from artificial life region labelling and global information using mathematical morphology is described. Also, we describe progress towards an automatic segmentation system that employs objective segmentation measures.

Abstract The automatic synthesis of the facial expression by computer is one of the most basic techniques to be applied in various fields. This paper describes a method to synthesize various facial expressions by modification of the 3‐D model of the face. The 3‐D facial model is obtained by adjusting the prepared 3‐D facial shape model to the frontal view of the object neutral face image, and by projecting the gray‐level informations. This paper proposes first a hierarchical model of facial description units for the hierarchical description of the facial expressions. The action unit (AU) of the FACS (facial action coding system), which is known as a systematic method of describing the facial expression, is considered as a layer in the proposed hierarchical model. AU of FACS is realized on a computer, and the automatic synthesis of the expression is executed. This procedure is based on the muscular structure of the face and allows cancellation of individual differences. At present, 34 kinds of AUs are implemented. The graylevel information for wrinkles and teeth, which are important in the synthesis of expressions, is not contained in the 3‐D facial model obtained from the neutral expression. Thus, such information is extracted from other face images, and registered as auxiliary information of the 3‐D facial model. This leads to synthesis of natural expressions. Finally, numerous examples for the facial expression synthesis are presented to show the usefulness of the method.

이 논문에서는 2000년부터 2004년까지 발생한 지자기 폭풍과 Sawtooth 진동 현상의 통계적 관계에 대해 연구하였다. 먼지 이 시기에 발생된 154건의 지자기 폭풍을 Dst 지수를 이용하여 선별하였으며 특히 선별된 지자기 폭풍이 코로나 물질 분출(Coronal Mass Ejection; CME), Corotating Interaction Region(CIR) 등 어떤 유도체에 의해 발생되었는지 구분하였다. 또한 같은 <TEX>$2000{\sim}2004$</TEX>년 기간에 대해 정지궤도 고에너지 대전 입자 플럭스 자료를 통해 Sawtooth 진동 현상 사례 48건을 선별하였다. 이 두 종류의 현상에 대한 통계적 상관관계를 분석한 결과, 총 154건의 지자기 폭풍 중에서 47건(약 30%)이 Sawtooth 진동 현상을 동반하는 지자기 폭풍이었다. 또한 총 48건의 Sawtooth 진동 현상 사건 중 단 1건의 경우를 제외하고 모든 Sawtooth 현상이 지자기 폭풍 기간 동안 발견되었다. 그리고 Sawtooth 진동을 동반하는 지자기 폭풍은 그 유도체가 CIR인 경우(약 30%) 보다는 CME인 경우(약 62%)가 더 많았다. 이외에도 Sawtooth 진동 현상은 CME에 의한 지자기 폭풍의 경우에는 주로(약 82%) 주상기간(Main Phase)에 발생하였지만 CIR에 의한 지자기 폭풍의 경우에는 주로(약 78%) 회복기간(Recovery Phase)에 발생하였다. 다음으로 지자기 폭풍을 유발하는데 중요한 요소인 행성간 자기장 IMF (Interplanetary Magnetic Field)의 남쪽 방향 성분 Bz 및 태양풍의 속도가 Sawtooth 진돌 발생기간 중 어떤 평균적인 특징을 갖는지 조사하였다. 대부분의 Sawtooth 진동 현상은 IMF Ba가 -15nT에서 0 사이이고, 태양풍 속도가 <TEX>$400{\sim}700km/s$</TEX>인 상태에 해당한다. 또한 IMF Bz의 강도는 Sawtooth 진동 기간 동안에 대전 입자 플럭스 증가의 횟수와 약한 상관관계가 있음을 발견하였다. We have investigated a statistical relationship between sawtooth oscillations and geomagnetic storms during 2000-2004. First of all we selected a total of 154 geomagnetic storms based on the Dst index, and distinguished between different drivers such as Coronal Mass Ejection (CME) and Co-rotating Interaction Region (CIR). Also, we identified a total of 48 sawtooth oscillation events based on geosynchronous energetic particle data for the same 2000-2004 period. We found that out of the 154 storms identified, 47 storms indicated the presence of sawtooth oscillations. Also, all but one sawtooth event identified occurred during a geomagnetic storm interval. It was also found that sawtooth oscillation events occur more frequently for storms driven by CME <TEX>$({\sim}62%)$</TEX> than for storms driven by CIR <TEX>$({\sim}30%)$</TEX>. In addition, sawtooth oscillations occurred mainly <TEX>$({\sim}82%)$</TEX> in the main phase of storms for CME-driven storms while they occurred mostly <TEX>$({\sim}78%)$</TEX> during the storm recovery phase for CIR-driven storms. Next we have examined the average characteristics of the Bz component of IMF, and solar wind speed, which were the main components for driving geomagnetic storm. We found that for most of the sawtooth events, the IMF Bz corresponds to -15 to 0 nT and the solar wind speed was in the range of <TEX>$400{\sim}700km/s$</TEX>. We found that there was a weak tendency that the number of teeth for a given sawtooth event interval was proportional to the southward IMF Bz magnitude.

Purpose The present study compared physical fitness, metabolic syndrome risk factors, and resting metabolic rate (RMR) according to body mass index (BMI) and percent body fat (%BF) in 20s females. Methods Fifty-one women in their 20s were recruited and assigned into three groups, i.e., normal group (n=18), normal weight obesity (NWO) group (n=18), and obesity group (n=15) according to BMI and %BF. Physical fitness, metabolic syndrome risk factors, and RMR were measured and compared among three groups. Results Main results were as follows: 1) Physical fitness were not significantly different among three groups. 2) Regarding 1-RM, arm curl and leg extension were significantly lower in normal group and NWO group than obesity group. Leg press was significantly lower in normal group than obesity group. 3) Regarding metabolic syndrome risk factors, there were significant differences in waist circumference, ordering from low to high such as normal, NWO, and obesity groups. Systolic blood pressure and diastolic blood pressure were significantly lower in normal group and NWO group than obesity group, while HDL-C was significantly higher in normal group than NWO group and obesity group. 4) Regarding RMR, absolute values of RMR such as VO2(㎖·min-1), RMR (Kcal·min-1), RMR (KJ·min-1), and RMR (Kcal·day-1) were significantly lower in normal group and NWO group than obesity group. On the other hand, relative value of RMR such as RMR (KJ·kg-1FW·h-1) was significantly higher in normal group than NWO group and obesity group. Conclusions It was concluded that obese women showed increased risk of metabolic syndrome and low relative RMR level, and NWO had similar problems. Active health management through physical activity and dietary control should be committed to NWO individuals because the NWO has possibility of high risk of metabolic syndrome and reduction of metabolic rate from 20s even though there was no problem in their external appearance.

Reducing bit-widths of activations and weights of deep networks makes it efficient to compute and store them in memory, which is crucial in their deployments to resource-limited devices, such as mobile phones. However, decreasing bit-widths with quantization generally yields drastically degraded accuracy. To tackle this problem, we propose to learn to quantize activations and weights via a trainable quantizer that transforms and discretizes them. Specifically, we parameterize the quantization intervals and obtain their optimal values by directly minimizing the task loss of the network. This quantization-interval-learning (QIL) allows the quantized networks to maintain the accuracy of the full-precision (32-bit) networks with bit-width as low as 4-bit and minimize the accuracy degeneration with further bit-width reduction (i.e., 3 and 2-bit). Moreover, our quantizer can be trained on a heterogeneous dataset, and thus can be used to quantize pretrained networks without access to their training data. We demonstrate the effectiveness of our trainable quantizer on ImageNet dataset with various network architectures such as ResNet-18, -34 and AlexNet, on which it outperforms existing methods to achieve the state-of-the-art accuracy.

Post-training quantization is a representative technique for compressing neural networks, making them smaller and more efficient for deployment on edge devices. However, an inaccessible user dataset often makes it difficult to ensure the quality of the quantized neural network in practice. In addition, existing approaches may use a single uniform bit-width across the network, resulting in significant accuracy degradation at extremely low bit-widths. To utilize multiple bit-width, sensitivity metric plays a key role in balancing accuracy and compression. In this paper, we propose a novel sensitivity metric that considers the effect of quantization error on task loss and interaction with other layers. Moreover, we develop labeled data generation methods that are not dependent on a specific operation of the neural network. Our experiments show that the proposed metric better represents quantization sensitivity, and generated data are more feasible to be applied to mixed-precision quantization.

Post-training quantization is a representative technique for compressing neural networks, making them smaller and more efficient for deployment on edge devices. However, an inaccessible user dataset often makes it difficult to ensure the quality of the quantized neural network in practice. In addition, existing approaches may use a single uniform bit-width across the network, resulting in significant accuracy degradation at extremely low bit-widths. To utilize multiple bit-width, sensitivity metric plays a key role in balancing accuracy and compression. In this paper, we propose a novel sensitivity metric that considers the effect of quantization error on task loss and interaction with other layers. Moreover, we develop labeled data generation methods that are not dependent on a specific operation of the neural network. Our experiments show that the proposed metric better represents quantization sensitivity, and generated data are more feasible to be applied to mixed-precision quantization.

In recent years, lectures on control systems have focused on hands-on experience using actual control equipment than before. However, the online education triggered compulsorily by the Covid-19 pandemic poses has restricted the construction of an education environment for hands-on experience. In this study, we proposed an economical and compact experimental environment for control education that enables hands-on experience even through online education. To this end, we utilized a light weight rapid control prototyping (LW-RCP), which is a lab-built RCP environment, and the environment was constructed using 3D printing. In the proposed environment, LW-RCP enabled students to focus on the learning and application of related control concepts, without the inconvenience of manual C-coding and the possibility of debugging errors. In addition, the proposed control equipment, which was manufactured using 3D printing, is an inexpensive equipment with a sufficiently small size that can be placed on a desk. Owing to the low cost and small size of the proposed environment, each student can have his/her own experimental equipment, which will enable a hands-on experience even through online education. Online education is expected to expand more and more in the future as it exhibits various advantages and potentials from traditional face-to-face classes. The proposed educational environment is expected to play a meaningful role in satisfying the demands of hands-on experience for control-related lectures.

In this paper, the local minima free search algorithm using chaos is proposed for an unstructured search space. The problem is that given the quality function, find the value of a configuration that minimizes the quality function. The proposed algorithm started basically from the gradient search technique but at the prescribed points, that is, local minimum points, which are to be automatically detected the chaotic jump is introduced by the dynamics of a chaotic neuron. Chaotic motions are mainly because of the Gaussian function having a hysteresis as a refractoriness. In order to enhance the probability of finding the global minimum, a parallel search strategy is also given. The validity of the proposed method wil be verified in simulation examples of the function minimization problem and the motion planning problem of a mobile robot.

A classical bidirectional associative memory (BAM) suffers from low storage capacity and abundance of spurious memories though it has the properties of good generalization and noise immunity. In this paper, Ham- ming distance in recall procedure of usual asymmetrical BAM is replaced with modified Hamming distance by introducing weighting matrix into connection matrix. This generalization is validated to increase storage capacity, to lessen spurious memories, and to enhance noise immunity using simulation work.

Abstract:The integration of eurozone financial markets since the advent of the euro in 1999 has been the center of attention in policy debates and academic research. We analyze the bank assets of monetary financial institutions in Germany vis-a-vis nonresidents. The financial institutions of the eurozone countries have tended to invest in assets of other eurozone countries substantially more since the introduction of the euro. The euro effect is especially stronger in the weaker eurozone economies than in the stronger eurozone economies. Furthermore, the impact of the euro has been even greater in securities than in loans. In this paper, we use Bundesbank balance-of-payment statistics to analyze the euro’s effects on the asset portfolios of German banks vis-a-vis nonresidents.

This paper presents a novel method which combines global shape parameters and scalable local descriptors for accurate body parts recognition from a single depth image in real-time. Human poses are of extremely large variation in aspects of visual shapes, because human can take poses from daily activities to gymnastic actions. In order to cover wide-range of the human poses, the proposed algorithm employs a unified structure which combines pose clustering and body parts classification. We name the proposed method Randomized Decision Bush (RDB). Specifically, global shape parameters which can discriminate coarse level shapes are utilized for pose clustering while scalable local shape descriptors are employed for accurate classification. RDB splits the various human poses into multiple clusters which contain similar shapes of the poses. As a result, it provides robust clustering which enables fine level classification within the cluster. The experimental results show improvements on recognizing body parts due to the pose clustering and classification with scalable local descriptors. Additionally, we significantly reduce the complexity of training a large number of human shapes.

Sewage treatment facilities can purify sewage enough to be send to river or sea water, that discharged from human life and industrial activities. In the sewage treatment process, we can get large amount of by-product energy resources and materials such as heat of sewage, digester gas and purified water etc., it can be utilized by applying various technologies thereby we can reduce energy consumption in the process. In this paper, it was analyzed using the data collected from the operational case study for the energy savings effect that can be obtained when using the digester gas, one of the by-product materials of sewage treatment process, for electric power generation. Cost of 623million won is an annual reduction of 4,032MWh corresponding 9% of the annual electricity consumption of the sewage treatment plant, such an alternative power generation using digester gas was proposed in this paper has been verified the feasibility of the proposed reduction of energy.

This paper presents a method for tracking human poses in real-time from depth image sequences. The key idea is to adopt recognition for generating the model to be tracked. In contrast to traditional methods utilizing a single-typed 3D body model, we directly define the human body model based on the body part recognition result of the captured depth image, which leads to the reliable tracking regardless of users' appearances. Moreover, the proposed method has the ability to efficiently reduce the tracking drift by exploiting the joint information inserted into our body model. Experimental results on real-world environments show that the proposed method is effective for estimating various human poses in real-time.

This paper presents a human pose recognition method which simultaneously reconstructs a human volume based on ensemble of voxel classifiers from a single depth image in real-time. The human pose recognition is a difficult task since a single depth camera can capture only visible surfaces of a human body. In order to recognize invisible (self-occluded) surfaces of a human body, the proposed algorithm employs voxel classifiers trained with multi-layered synthetic voxels. Specifically, ray-casting onto a volumetric human model generates a synthetic voxel, where voxel consists of a 3D position and ID corresponding to the body part. The synthesized volumetric data which contain both visible and invisible body voxels are utilized to train the voxel classifiers. As a result, the voxel classifiers not only identify the visible voxels but also reconstruct the 3D positions and the IDs of the invisible voxels. The experimental results show improved performance on estimating the human poses due to the capability of inferring the invisible human body voxels. It is expected that the proposed algorithm can be applied to many fields such as telepresence, gaming, virtual fitting, wellness business, and real 3D contents control on real 3D displays.

In this paper, we present a real-time prototype of a robot boxing game based on a novel interaction method which provides a simpler control for a miniature humanoid. Specifically, an upper body of the robot mimics a user's upper body motion, while a lower body of the robot moves autonomously towards a target object (an opponent robot). To the best of our knowledge this is the first robot boxing game which provides semi-autonomous control based on natural human motions. Questionnaire interview shows the users feel immersive gaming experience and companionship with the robot in the living room.

Fragments are essential for the realisation of my musical ideas. This commentary explains the musical, aesthetical and philosophical backgrounds to my work, and also the detailed compositional processes employed. Chapter one is a brief overview of Sigimsae and main-tone as found in traditional Korean music, which subsequently become transformed in my composition through the integration with contemporary Western musical elements. Chapter two analyses six compositions in this folio to clarify how unique sounds and forms are developed and achieved through the ever-changing fragment.