15,834 research outputs found

    The anisotropic noise in stochastic gradient descent: Its behavior of escaping from sharp minima and regularization effects

    No full text
    Understanding the behavior of stochastic gradient descent (SGD) in the context of deep neural networks has raised lots of concerns recently. Along this line, we study a general form of gradient based optimization dynamics with unbiased noise, which unifies SGD and standard Langevin dynamics. Through investigating this general optimization dynamics, we analyze the behavior of SGD on escaping from minima and its regularization effects. A novel indicator is derived to characterize the efficiency of escaping from minima through measuring the alignment of noise covariance and the curvature of loss function. Based on this indicator, two conditions arc established to show which type of noise structure is superior to isotropic noise in term of escaping efficiency. We further show that the anisotropic noise in SGD satisfies the two conditions, and thus helps to escape from sharp and poor minima effectively, towards more stable and flat minima that typically generalize well. We systematically design various experiments to verify the benefits of the anisotropic noise, compared with full gradient descent plus isotropic diffusion (i.e. Langevin dynamics)

    Determination of lead isotope ratios in seawater by quadrupole inductively coupled plasma mass spectrometry after Mg(OH)2 co-precipitation

    No full text
    A low blank Mg(OH)2 pre-concentration method was evaluated for the determination of lead isotope ratios (208Pb/206Pb, 207Pb/206Pb) in seawater using a quadrupole ICP-MS VG Plasma Quad II+. Possible matrix effects derived from the Mg(OH)2 co-precipitate were assessed by spiking lead-free seawater (PbFS) and 1% (v/v) HNO3 with the certified common lead standard NBS 981 to give solutions with concentrations in the lower picogram per millilitre range. The standard curves for all three masses were linear in both matrices with minor signal loss (18%) in the Mg matrix. Mass fractionation showed similar mass biases (&lt;2% frac./amu) for 208Pb/206Pb and 207Pb/206Pb in both seawater and 1% (v/v) HNO3, indicating that there is no significant matrix influence on the isotope ratio determination. Using the Mg(OH)2 precipitation method, real seawater samples were pre-concentrated approximately 28-fold, and 1 ml of 5% (v/v) HNO3 end volume was used for the subsequent measurement. The data acquisition parameters dwell time, replicates per analysis, and acquisition time were first varied to optimize analytical precision and accuracy of the ICP-MS measurements. On the basis of these results, approximately 0.5 ml of pre-concentrated solution was finally used employing a low-flow, CETAC microconcentric nebulizer to minimize sample consumption and extend acquisition time. Analyzing 18 seawater samples from the North Atlantic, the average 1? external precision of triplicate measurements was approximately 0.3% for both ratios, 208Pb/206Pb and 207Pb/206Pb, at the level of approximately 20 pg ml?1 Pb. Blanks accounted for less than 3% of total lead analyzed for each sample. Six samples were also measured with TIMS and agreed in average within 0.26% for 207Pb/206Pb and 0.37% for 208Pb/206Pb. Three surface water samples from the Sargasso Sea, collected in 1989, showed ratios in line with previous published ratios from the western North Atlantic.<br/

    Tangent-normal adversarial regularization for semi-supervised learning

    No full text
    Compared with standard supervised learning, the key difficulty in semi-supervised learning is how to make full use of the unlabeled data. A recently proposed method, virtual adversarial training (VAT), smartly performs adversarial training without label information to impose a local smoothness on the classifier, which is especially beneficial to semi-supervised learning. In this work, we propose tangent-normal adversarial regularization (TNAR) as an extension of VAT by taking the data manifold into consideration. The proposed TNAR is composed by two complementary parts, the tangent adversarial regularization (TAR) and the normal adversarial regularization (NAR). In TAR, VAT is applied along the tangent space of the data manifold, aiming to enforce local invariance of the classifier on the manifold, while in NAR, VAT is performed on the normal space orthogonal to the tangent space, intending to impose robustness on the classifier against the noise causing the observed data deviating from the underlying data manifold. Demonstrated by experiments on both artificial and practical datasets, our proposed TAR and NAR complement with each other, and jointly outperforms other state-of-the-art methods for semi-supervised learning.</p

    Control and Filtering for Discrete Linear Repetitive Processes with H infty and ell 2--ell infty Performance

    No full text
    Repetitive processes are characterized by a series of sweeps, termed passes, through a set of dynamics defined over a finite duration known as the pass length. On each pass an output, termed the pass profile, is produced which acts as a forcing function on, and hence contributes to, the dynamics of the next pass profile. This can lead to oscillations which increase in amplitude in the pass to pass direction and cannot be controlled by standard control laws. Here we give new results on the design of physically based control laws for the sub-class of so-called discrete linear repetitive processes which arise in applications areas such as iterative learning control. The main contribution is to show how control law design can be undertaken within the framework of a general robust filtering problem with guaranteed levels of performance. In particular, we develop algorithms for the design of an H? and 2\ell_{2}–\ell_{\infty} dynamic output feedback controller and filter which guarantees that the resulting controlled (filtering error) process, respectively, is stable along the pass and has prescribed disturbance attenuation performance as measured by HH_{\infty} and 2\ell_{2}\ell_{\infty} norms

    Acoustic radiation due to scattering of T-S wave by the mean-flow distortion induced by steady local suction

    No full text
    Substantial sound waves can be generated by boundary-layer instability modes when the latter are scattered by a rapid mean-flow distortion. This is a rather generic mechanism and operates when an oncoming T-S wave is scattered by a steady local suction slot. This paper focuses on this problem by extending a recently developed Local Scattering Theory (Wu & Dong, J. Fluid Mech. submitted), where a so-called transmission coefficient, defined as the ratio of the T-S wave amplitude downstream of the scatter to that upstream, is introduced to characterize the effect of a local scatter on boundary-layer instability and transition. As in the earlier work, the mathematical formulation is based on triple-deck formulism, but in order to accommodate the acoustic far field, which was not considered in the paper mentioned, the unsteady terms in the upper deck, which play a leading-order role in radiation, are retained, and the influence of the radiated sound on the near-wall perturbation is included. The upper deck equation for the pressure is the Helmholtz equation rather than the Laplace equation. This leads to a modified pressure-displacement relation, which is coupled with the linearized boundary-layer equations in the lower deck. Discretization of the whole system formulates a generalized eigenvalue problem, which is solved numerically. It is found that suction suppresses oncoming T-S waves, and this effect increases with the suction velocity and the slot width. The directivity is ndependent of the flow parameters only when the Mach number is low. The intensity of the radiated sound in general increases with the frequency, the suction velocity and the width of the suction slot. Interestingly, for O(1) suction velocities, the radiated sound is very weak, indicating that the gain of stabilizing effect does not cause aeroacoustic penalty

    Physiological and environmental control on ecosystem water use efficiency in response to drought across the northern hemisphere

    No full text
    Drought, a natural hydrometeorological phenomenon, has been more frequent and more widespread due to climate change. Water availability strongly regulates the coupling (or trade-off) between carbon uptake via photosynthesis and water loss through transpiration, known as water-use efficiency (WUE). Understanding the effects of drought on WUE across different vegetation types and along the wet to dry gradient is paramount to achieving better understanding of ecosystem functioning in response to climate change. We explored the physiological and environmental control on ecosystem WUE in response to drought using observations for 44 eddy covariance flux sites in the Northern Hemisphere. We quantified the response of WUE to drought and the relative contributions of gross primary production (GPP) and evapotranspiration (ET) to the variations of WUE. We also examined the control of physiological and environmental factors on monthly WUE under different moisture conditions. Cropland had a peak WUE value under moderate drought conditions, while grassland, deciduous broadleaf forest (DBF), evergreen broadleaf forest (EBF), and evergreen needleleaf forest (ENF) had peak WUE under slight drought conditions. WUE was mainly driven by GPP for cropland, grassland, DBF, and ENF but was mainly driven by ET for EBF. Vapor pressure deficit (VPD) and canopy conductance (Gc) were the most important factors regulating WUE. Moreover, WUE had negative responses to air temperature, precipitation, and VPD but had a positive response to Gc and ecosystem respiration. Our findings highlight the different effects of biotic and abiotic factors on WUE among different vegetation types and the important roles of VPD and Gc in controlling ecosystem WUE in response to drought

    Measurement of dwell times of spin polarized rubidium atoms on octadecyltrichlorosilane- and paraffin-coated surfaces

    No full text
    We report the measurement of dwell times of spin polarized Rb atoms on octadecyltrichlorosilane OTS- and paraffin-coated surfaces. We find that at a cell temperature of 72 °C the dwell times for OTS- and paraffin-coated surfaces are 0.9+-0.1 microsecond and 1.8 +-0.2 microsecond , respectively. Since the relaxation probability on paraffin is almost one order of magnitude smaller than that on OTS, the longer dwell time for paraffin indicates that the average strength of the interactions experienced by Rb atoms while they are inside paraffin is much weaker than while they are inside OTS.Peer reviewe

    [[alternative]]A Study on History of Shin-Wu Elementary School at Tao-Yuan(1905-2003)

    No full text
    [[abstract]]To explore the school cultural development in Shin-Wu Elementary school in the last hundred years and the significance of school history research in educational history, this study examines documents and files in the school history room of Shin-Wu Elementary school and has oral history interviews as well. The main findings of this study are as follows. 1. In late Meiji, most of the teachers were Japanese or Taiwanese from the nearby common schools, however, the turnover was high. At that time, only affluent students attended schools while later, more students were from poor families. Yet, the drop-out rate was relatively high. 2. As for the age, most students went to school in their teens, for they had to share responsibilities in home chores. 3. In Taisyo period, there were more and more teachers who had been Shin-Wu graduates and students were from different sources. 4. Owing to the air raid during the war, schools got closed very often and the content of education was mainly about laboring. 5. Post-war education, compared with that before, varied little in educational contents and forms because of the position-retaining teachers and staff. 6. For the last hundred years, nearly one-third of Shin-Wu teachers were their graduates and principals and teachers of branch schools came form Shin-Wu Elementary School, too. Therefore, Shin-Wu Elementary School had played an influential role in the elementary education and the local people cultivation in Shin-Wu area. 7. Shin-Wu Elementary School had always followed the national educational policy, seldom did they have local standpoints, which meant, they had to strengthen their autonomy and capacity when faced with new challenges especially after the lifting of Martial Law. 8. Through elaborative school cultural analysis, relationship among education, politics, economy and society, in particular, individual area development will be more carefully investigated.

    Fighting Inequalities: Impact of Urban-Rural Health Insurance Integration in Rural China

    No full text
    Integrating social health insurance, which helps unify the administration, policies, and funds of various health insurance programs, is expected to combat health inequalities. This paper provides the first micro-level evidence from China on the impact of health insurance integration on health care. We exploit the China Health and Retirement Longitudinal Study (CHARLS) dataset and combine it with original city-level data on social health insurance policy. We employ difference-in-difference models with matching to estimate the effects of integrating the Urban Resident Basic Medical Insurance (URBMI) and the New Rural Cooperative Medical Insurance (NRCMI) on individuals’ health care utilization (including inpatient and outpatient care) and health outcomes. We find that the health insurance integration has significantly increased inpatient and outpatient care utilization. Moreover, we provide evidence of a possible mechanism that drives this relationship: the inpatient reimbursement rate significantly increased after the integration. We further find that after the integration, people traveled longer distances to seek inpatient and outpatient care, and their health, as measured by self-assessed health status, has improved. Nonetheless, these results are found in the rural sample only, and we observe no significant impact of the health insurance integration on total hospital charges or patients’ out-of-pocket payments.Peer reviewe
    corecore