Training the NN

We used a three-layer feedforward architecture. The number of input neurons in the input layer of NN is equal to the number of selected parameters (five). There is only one output of the NN, which is the corresponding MOS score. All the values (inputs and output are normalized to achieve better performance) The training algorithm for ANN is the well-known Backpropagation algorithm. For RNN, we used Gelenbe's algorithm [47]. In addition we used the set of training algorithms that we developed for RNN. The precision is the same, but using our algorithms, the required iterations to achieve the same error conditions is very small compared to Gelenbe's algorithm (see Section 10.5.4 for numerical data). The number of hidden neurons in the hidden layer is variable as it depends on the complexity of the problem (inputs, outputs, training set, and the global precision needed). There is a trade-off between the number of hidden neurons and the ability of the NN to generalize (i.e. to produce good results for inputs not seen during the training phase). Hence, the number of hidden neurons should not be too large. In our case we chose 5 neurons for RNN and 4 for ANN. After carrying out the MOS experiment for the 94 samples, we divided the database into two parts: one to train the NN containing 80 samples (the first in Table 6.2), and the other to test the performance of the NN, containing other 14 samples. After training the NN using the first database and comparing the training set against the values predicted by the NN, we got a correlation coefficient of 0.9801, and a mean square error of 0.108; that is, the NN model fits quite well the way in which humans rated video quality. The result is shown in Figures 6.3 and 6.5(a). It should be noted that for the optimal number of hidden neurons for both ANN and RNN, the obtained performances are almost the same. However, ANN can be overtrained easily if care is not taken. In this case, the performance on the training phase is very good but it cannot generalize well for the non-seen samples during the training. As previously mentioned in Chapter 4, we varied the number of hidden neurons for both ANN and RNN. We used them to evaluate the quality for each case when varying the Bit Rate and fixing the other parameters. As we can see from Figure 4.10(a) and Figure 4.10(b) that for the optimal number of hidden neurons, they give the same performance. But ANN cannot generalize for the other values. However, RNN almost give the same results for different values of hidden neurons. In addition, for the maximum value of Bit Rate, RNN gives an MOS score of almost 9 (as expected).

Figure 6.3: Actual and Predicted MOS scores for the training database. The NN learned the way by which the group of human subjects rated video quality for the video samples with very good precision.