Wyniki wyszukiwania - BazTech

1

Data Selection for Neural Networks

Kordos M.

Schedae Informaticae

|

2016

|

Vol. 25

153--164

EN

Several approaches to joined feature and instance selection in neural network leaning are discussed and experimentally evaluated in respect to classification accuracy and dataset compression, considering also their computational complexity. These include various versions of feature and instance selection prior to the network learning, the selection embedded in the neural network and hybrid approaches, including solutions developed by us. The advantages and disadvantages of each approach are discussed and some possible improvements are proposed.

2

Effectiveness of Unsupervised Training in Deep Learning Neural Networks

Rusiecki A., Kordos M.

Schedae Informaticae

|

2015

|

Vol. 24

41--51

EN

Deep learning is a field of research attracting nowadays much attention, mainly because deep architectures help in obtaining outstanding results on many vision, speech and natural language processing – related tasks. To make deep learning effective, very often an unsupervised pretraining phase is applied. In this article, we present experimental study evaluating usefulness of such approach, testing on several benchmarks and different percentages of labeled data, how Contrastive Divergence (CD), one of the most popular pretraining methods, influences network generalization.

3

Neural network-based prediction of additives in the steel refinement process

Wieczorek T., Kordos M.

Computer Methods in Materials Science

|

2010

|

Vol. 10, No. 1

16-24

EN

To make fine adjustments to the steel composition alloy additions are added to the crude steel to adjust composition for the grade of steel being manufactured. The prediction of hot metal composition is a crucial factor in the economy of ladle furnace operation. Usually it is made by calculations based on the equilibrium of chemical reactions in molten steel. In the paper the problem of prediction of alloy additions has been solved using Artificial Neural Nets (ANN’s) and the SVM algorithm. A brief state of the art review of the application of computational intelligence (CI) in secondary steelmaking has been made. The prediction system used by authors has been introduced. Problems with data preparation have been presented. Details of the neural network prediction and other approaches to the prediction problem, i.e. Support Vector Regression Module and Multivariate Linear Regression have been introduced. Experimental results and the final conclusions and recommendations have been presented.

PL

Dla uzyskania żądanego składu chemicznego stali wprowadzane są do niej w procesie produkcji dodatki stopowe. Precyzyjne przewidywanie końcowego składu chemicznego produkowanej stali jest niezwykle ważne dla ekonomicznego prowadzenia procesu w piecu kadziowym. Zwykle obliczenia są prowadzone w oparciu teorię procesów chemicznych zachodzących w ciekłej stali w warunkach równowagi. W pracy problem przewidywania masy dodatków stopowych koniecznych do uzyskania żądanego składu został rozwiązany przy użyciu sztucznych sieci neuronowych i algorytmu maszyny wektorów podpierających (SVM). Opisano opracowany model inteligentny obliczania dodatków. Przedstawiono zbudowane moduły tego systemu: sieci neuronowe, SVM, regresję liniową wielu zmiennych. Omówiono zagadnienia preprocesingu danych do uczenia modułów inteligentnych. Przedstawiono uzyskane wyniki, wnioski wyciągnięte z badań oraz efekty uzyskane w rzeczywistych warunkach przemysłowych.

4

Search-based approach to multilayer perceptron training

Kordos M.

Studia Informatica

|

2005

|

Vol. 26, nr 1

29-43

EN

The paper presents two search-based algorithms for MLP training; numerical gradient and yariable step search algorithm. The advantages of the methods comparing to analytical gradient-based algorithms include Iow memory reąuirements, the algorithm simplicity and determining morę optimal next step direction by direct access to the influence of hidden layer weights on nerwork error.

PL

Artykuł przedstawia dwie metody uczenia sieci MLP oparte na . przeszukiwaniu: gradient numeryczny i metodę zmiennego kroku przeszukiwania (VSS). Zaletą tych metod są małe wymagania pamięciowe, prosta budowa algorytmu t v oraz dokładniejsze określenie kierunku następnego kroku przez bezpos'rednie określanie wpływu wartości wag warstwy ukrytej na błąd sieci.

5

A survey of factors influencing MLP error surface

Kordos M., Duch W.

Control and Cybernetics

|

2004

|

Vol. 33, no 4

611-631

EN

Visualization of neural network error surfaces and learning trajectories helps to understand the influence of numerous factors on the neural learning process. This understanding can be used to improve training and design of MLP networks. The following topics are discussed using a few benchmark datasets for illustration: general error surface properties including local minima, plateaus and narrow funnels, their dependence on network structure, input data, transfer and error functions, consequences of weight initialization, and interesting directions in the weight space. The error surfaces are shown in 3-dimensional PCA-based projections. Finally a possibility of effective weight number reduction is discussed.