Wyniki wyszukiwania - BazTech

1

Implications of Pooling Strategies in Convolutional Neural Networks: A Deep Insight

Sharma Shallu, Mehra Rejesh

Foundations of Computing and Decision Sciences

|

2019

|

Vol. 44, No. 3

303--330

EN

Convolutional neural networks (CNN) is a contemporary technique for computer vision applications, where pooling implies as an integral part of the deep CNN. Besides, pooling provides the ability to learn invariant features and also acts as a regularizer to further reduce the problem of overfitting. Additionally, the pooling techniques significantly reduce the computational cost and training time of networks which are equally important to consider. Here, the performances of pooling strategies on different datasets are analyzed and discussed qualitatively. This study presents a detailed review of the conventional and the latest strategies which would help in appraising the readers with the upsides and downsides of each strategy. Also, we have identified four fundamental factors namely network architecture, activation function, overlapping and regularization approaches which immensely affect the performance of pooling operations. It is believed that this work would help in extending the scope of understanding the significance of CNN along with pooling regimes for solving computer vision problems.

2

A comparison of deep convolutional neural networksfor image-based detection of concrete surface cracks

Słoński Marek

Computer Assisted Methods in Engineering and Science

|

2019

|

Vol. 26, no. 2

105--112

EN

The aim of this paper is to compare the performance of four deep convolutional neural networks in theproblem of image-based automated detection of concrete surface cracks in the case of a small dataset. Thiscrack detection problem is treated as a binary classification problem, and it is solved by training a deepconvolutional neural network on the small dataset. In this context, overfitting during training was the mainissue to cope with and various techniques were applied to overcome this issue. The results of the experi-ments suggest that the best approach for this problem is to use the pretrained convolutional base of a largepretrained convolutional neural network as an automatic feature extraction method and adding a new bi-nary classifier on top of the convolutional base. Then, at the training the new classifier and fine-tuningthe last few layers of the pretrained network take place at the same time. The classification accuracy of thebest deep convolutional neural network on the testing set is about 94%.

3

Kinetics of flotation. Order of process, rate constant distribution and ultimate recovery

Bu X., Xie G., Peng Y., Ge L., Ni C.

Physicochemical Problems of Mineral Processing

|

2017

|

Vol. 53, iss. 1

342--365

EN

Kinetic models can be used to characterize the flotation process. In this paper, three primary parameters, namely, distribution of flotation rate constant f(K), order of flotation process n and ultimate recovery R∞ are presented to perform analysis of flotation kinetics. The flotation rate constant f(K) is a function of both the size and hydrophobicity of particles. Though the more commonly used distributions are Delta function as well as Rectangular, Kelsall and Gamma models, there is no agreement in the literature as to which distribution function better characterize the floatability distribution. The first-order models can be used to describe most mineral flotation processes, while there is also evidence that the non-integral-order equation is capable of representing the kinetic characteristics of the batch flotation process. The order is lower than 1 in the initial moments of the flotation process. The solution of ultimate recovery calculated by the least squares method is greater than 100% (R∞ >100%). An empirical model was proposed to avoid the improper phenomenon in the solution of ultimate recovery, which can improve the availability and validity of kinetic models. Finally, more attention should be paid to the overfitting resulting from the increase in the number of parameters in the statistical analysis of kinetic models.

4

Portfolio Inputs Selection from Imprecise Training Data

Raudys S., Raudys A., Biziuleviciene G., Pabarskaite Z.

Schedae Informaticae

|

2016

|

Vol. 25

177--188

EN

This paper explores very acute problem of portfolio secondary overfitting. We examined the financial portfolio inputs random selection optimization model and derived the equation to calculate the mean Sharpe ratio in dependence of the number of portfolio inputs, the sample size L used to estimate Sharpe ratios of each particular subset of inputs and the number of times the portfolio inputs were generated randomly. It was demonstrated that with the increase in portfolio complexity, and complexity of optimization procedure we can observe the over-fitting phenomena. Theoretically based conclusions were confirmed by experiments with artificial and real world 60,000-dimensional 12 years financial data.

5

Spectral Mapping Using Kernel Principal Components Regression for Voice Conversion

Song P., Zhao L., Bao Y.

Archives of Acoustics

|

2013

|

Vol. 38, No. 1

39--45

EN

The Gaussian mixture model (GMM) method is popular and efficient for voice conversion (VC), but it is often subject to overfitting. In this paper, the principal component regression (PCR) method is adopted for the spectral mapping between source speech and target speech, and the numbers of principal components are adjusted properly to prevent the overfitting. Then, in order to better model the nonlinear relationships between the source speech and target speech, the kernel principal component regression (KPCR) method is also proposed. Moreover, a KPCR combined with GMM method is further proposed to improve the accuracy of conversion. In addition, the discontinuity and oversmoothing problems of the traditional GMM method are also addressed. On the one hand, in order to solve the discontinuity problem, the adaptive median filter is adopted to smooth the posterior probabilities. On the other hand, the two mixture components with higher posterior probabilities for each frame are chosen for VC to reduce the oversmoothing problem. Finally, the objective and subjective experiments are carried out, and the results demonstrate that the proposed approach shows greatly better performance than the GMM method. In the objective tests, the proposed method shows lower cepstral distances and higher identification rates than the GMM method. While in the subjective tests, the proposed method obtains higher scores of preference and perceptual quality.