Search

We study the effect of normalization on the layers of deep neural networks of feed-forward type. A given layer i with Ni hidden units is allowed to be normalized by 1=Nγii with γi ∊ [1=2, 1] and we study the effect of the choice of the i on the statistical behavior of the neural network's output (such as variance) as well as on the test accuracy on the MNIST data set. We find that in terms of variance of the neural network's output and test accuracy the best choice is to choose the i's to be equal to one, which is the mean-field scaling. We also find that this is particularly true for the outer layer, in that the neural network's behavior is more sensitive in the scaling of the outer layer as opposed to the scaling of the inner layers. The mechanism for the mathematical analysis is an asymptotic expansion for the neural network's output. An important practical consequence of the analysis is that it provides a systematic and mathematically informed way to choose the learning rate hyperparameters. Such a choice guarantees that the neural network behaves in a statistically robust way as the Ni grow to infinity. [ABSTRACT FROM AUTHOR]

Background: Yoga therapy is widely applied to the maintenance of health and to treatment of various illnesses. Previous researches indicate the involvement of autonomic control in its effects, although the general agreement has not been reached regarding the acute modulation of autonomic function. Aim: The present study aimed at revealing the acute effect of yoga on the autonomic activity using heart rate variability (HRV) measurement. Methods: Twenty-seven healthy controls participated in the present study. Fifteen of them (39.5 ± 8.5 years old) were naïve and 12 (45.1 ± 7.0 years old) were experienced in yoga. Yoga skills included breath awareness, two types of asana, and two types of pranayama. HRV was measured at the baseline, during yoga, and at the resting state after yoga. Results: In both yoga-naïve and experienced participants, the changes in low-frequency (LF) component of HRV and its ratio to high-frequency (HF) component (LF/HF) after yoga were found to be correlated negatively with the baseline data. The changes in LF after yoga were also correlated with LF during yoga. The changes in HF as well as the raw HRV data after yoga were not related to the baseline HRV or the HRV during yoga. Conclusion: The results indicate that yoga leads to an increase in LF when LF is low and leads to a decrease in LF when it is high at the baseline. This normalization of LF is dependent on the autonomic modulation during yoga and may underlie the clinical effectiveness of yoga therapy both in yoga-naïve and experienced subjects. [ABSTRACT FROM AUTHOR]

In Gamma Knife forward treatment planning, normalization effect may be observed when multiple shots are used for treating large lesions. This effect can reduce the proportion of coverage of high-value isodose lines within targets. The aim of this study was to evaluate the performance of forward treatment planning techniques using the Leksell Gamma Knife for the normalization effect reduction. We adjusted the shot positions and weightings to optimize the dose distribution and reduce the overlap of high-value isodose lines from each shot, thereby mitigating the normalization effect during treatment planning. The new collimation system, Leksell Gamma Knife Perfexion, which contains eight movable sectors, provides an additional means to reduce the normalization effect by using composite shots. We propose different techniques in forward treatment planning that can reduce the normalization effect. Reducing the normalization effect increases the coverage proportion of higher isodose lines within targets, making the high-dose region within targets more uniform and increasing the mean dose to targets. Because of the increase in the mean dose to the target after reducing the normalization effect, we can set the prescribed marginal dose at a higher isodose level and reduce the maximum dose, thereby lowering the risk of complications. [ABSTRACT FROM AUTHOR]

Field normalization is a necessary step in a fair cross-field comparison of citation impact. In practice, mean-based method ( m-score) is the most popular method for field normalization. However, considering that mean-based method only utilizes the central tendency of citation distribution in the normalization procedure and dispersion is also a significant characteristic, an open and important issue is whether alternative normalization methods which take both central tendency and variability into account perform better than mean-based method. With the aim of collapsing citation distributions of different fields into a universal distribution, this study compares the normalization effect of m-score and z-score based on 236 Web of Science (WoS) subject categories. The results show that both m-score and z-score have remarkable normalization effect as compared with raw citations, but neither of them can realize the ideal goal of 'universality of citation distributions'. The results also suggest that m-score is generally preferable to z-score. The essential cause that m-score has an edge over z-score as a whole has a direct relationship with the characteristics of skewed citation distributions in which case m-score is more applicable than z-score. [ABSTRACT FROM AUTHOR]

J. J. Gibson has noted that during prolonged viewing a line perceptually rotates towards the nearest vertical or horizontal meridian. This is known as the normalization effect, but the phenomenon remains poorly investigated. According to our experimental results, the adapting line perceptually rotates to the nearest of three orientations: vertical, horizontal or diagonal. The orientation of these three lines does not change during prolonged viewing. Furthermore, the orientation of lines tilted by either 22.5° or 67.5° does not change subjectively, either. Any changes in the orientation of these lines cause subjective drift towards the nearest vertical, diagonal (oriented by 45°) or horizontal line. [ABSTRACT FROM AUTHOR]