|NITAAI-Veda.nyf > Soul Science God Philosophy > Science and Spiritual Quest > Section 4 Towards a New Biology > MATHEMATICAL TECHNIQUES FOR STUDY OF EEG > 4. Discussion|
From the above Tables 1-3, we can observe the following:The features fractal dimension and complexity measure play a significant role in the classification problem.
1.The obtained accuracies for the back propagation network, RBF networks units and support vector machine with RBF kernel are higher than for k-neighbors. This is because in the case of data used in this work, different units bring different amount of information about presence of meditation state where as k-neighbors algorithm compute the Euclidean distances between classified vector and some other vectors. Therefore, this algorithm cannot take into consideration the fact that different inputs bring different amount of information, which is the natural feature of neural networks.
2.The features fractal dimension and complexity measure play a significant role in the classification problem. They were able to detect the changes in the background activity. In some cases (like four category problem), frequency features are also significant.
3.KNNC is not at all suitable for four-category case.
4.RBF network, in general, gives better performance than MLP network since it can generalize more.
5.SVM gives better performance than MLP and RBF networks since we are using RBF kernel in SVM.
6.In some cases part of the total features gives almost equal performance that is obtained with total features. This is due to the fact that these features alone are significant and adding more features to them is not increasing the performance significantly.
7.Very high classification accuracy of meditation problem with neural networks was due to features computed for the windows of the signal,delivered to the network's inputs. The features express the characteristic features of different stages of meditation, which distinguish from each other.