How Well does the NN Perform?

In order to address the question ``How well does the NN perform?'', the NN was applied to the testing set (which contains samples that never have been seen during the training process). The results were correlation coefficient of 0.9821 and a mean square error of 0.07. Once again the performance of the NN was excellent, as can be observed in Figure 6.4. We show also a scatter plot for this case in Figure 6.5(b). From Figures 6.3, 6.4, and 6.5, it can be observed that video quality scores generated by the NN fits nicely the nonlinear model built by the subjects participating in the MOS experiment. Also, by looking at Figure 6.4, it can be established that learning algorithms give NN the advantage of high adaptability, which allows them to self optimize their performance when functioning under a dynamical environment (that is, reacting to inputs never seen during the training phase).

Figure 6.4: Actual and Predicted MOS scores for the testing database. The trained NN is able to evaluate video quality for new samples that never have been seen during the training, with good correlation with the results obtained by human subjects

Figure 6.5: Scatter plots showing the correlation between Actual and Predicted MOS scores.

[Training DB] [Testing DB]