48 G. P. Tsialiamanis et al. Table 5.5 Confusion Matrix of neural network classifier trained on the transformed data of the second dataset, test set, total accuracy: 96.96% Predicted panel 3 6 Missing panel 3 65 1 Missing panel 6 3 63 Fig. 5.6 Principal components of original features of the first dataset (a) and transformed features (b) Fig. 5.7 Principal components of original features of the second dataset (a) and transformed features (b) The training histories of the two models can be seen in Fig. 5.8. It is clear that the loss history of the model with transformed data (blue and cyan lines) converges faster, especially in the initial part of the training, and it also reaches a lower minimum value for the loss function in the same number of training epochs. This can be explained by looking at the effect of the learnt transformation on the data. In Figs. 5.6 and 5.7 this effect is illustrated. (Note that the points are different from those in Fig. 5.4, because principal component analysis was performed this time on the normalised data in the interval [−1, 1] for the neural network training). The transformation spreads out the points of the original problem (first dataset) in order to make their separation by the decision layer easier; however, it is clear that it also accomplishes the same result on the second dataset. The points in Fig. 5.7b are spread out compared to the initial points and thus, their separation by the single layer neural network is easier. Furthermore, the points lay further away from the required decision boundary and this explains
RkJQdWJsaXNoZXIy MTMzNzEzMQ==