Learning of Neural Network with Reduced Interference – An Ensemble Approach

Abstract

Abstract-This paper focuses on reducing the interference effect among input attributes. When training different attributes together, there may exist negative effect among them due to interference. To reduce the interference, input attributes are placed into different groups such that attributes with no interference with each other are placed in the same group. Two types of grouping strategies are examined in this paper, i.e. non-overlapping and overlapping. To further enhance the performance, multiple learners are employed to tackle different groups. Three integration methods i.e. voting, weighting and result-integration network (RIN) are examined.It turns out that the result-integration network has the best performance, followed by weighting and then voting. The ensemble approach can improve the performance of neural-network learning. Such an approach also can be employed with feature selection to further enhance the performance.

A hyperbolic divergence based nonparametric test for two‐sample multivariate distributions

“…In this section, we present the numerical performance of the proposed hyperbolic divergence test for distribution testing and compare it with other competing tests such as energy distance (ED; Székely & Rizzo, 2004), maximum mean discrepancy (MMD; Gretton et al, 2012), ball divergence (BD; Pan et al, 2018), and projective ensemble (PE; Li & Zhang, 2020). The projection‐averaging strategy (Kim, Balakrishnan & Wasserman, 2020) was not used here because the power performance of the projection‐averaging test method is equivalent to that of the PE approach but requires more computation (Li & Zhang, 2020). In our simulations, we employed the Gaussian kernel with the median heuristic for the MMD, and set for the proposed method.…”

“…The metric in a unit hypersphere is the inverse cosine function of the inner product of two unit vectors. Recently, Li & Zhang (2020) and Kim, Balakrishnan & Wasserman (2020) investigated nonparametric two‐sample tests in which Euclidean data were expressed with the spherical model. A natural question is whether and how we may portray Euclidean data using the hyperbolic model.…”