X hits on this document

PDF document

The split-fovea model of visual word - page 54 / 78





54 / 78

Chapter 2. The Simulations


is therfor important for the pronunciation of the word. Since the fixation net sees the word most often as ’pin-t’ it might develop a preference to treat ’in’ as the bigram it has previously learned which would make it more difficult to eventually learn the correct pronunciation. Figure 2.3 shows the evolution of both networks during training. As can be seen from this graph, both networks train fairly quick due to the small size of their training corpus. Also, the control network learns slightly faster than the fixation net. This might be due to the errors introduced in the fixation net by position 1. Position 1 has such a low frequency compared to the other position that it will inevitably lead to more errors at that position than in the control network.

Mean Sum-Square Error

The evolution of the mean sum-square error over time at

all fixation positions can be seen in tables 2.4, 2.6 and 2.8 for the control net and in figures 2.5, 2.7 and 2.9. Tables 2.4 and 2.5 show the MSE for the respective nets. Both curves tend to zero quite fast, which is expected from the general results of the nets (table 2.3). It can be seen that the MSE curves for the fixation positions are all very similar for the control network. This comes from the fact that there is no preferred fixation position for this net and all the positions have the same frequency. However, for the fixation net, there is a noticeable difference in the MSE curves of the different


To make it easier to analyse the graphs, figures 2.6 and 2.9 show a part of the whole curve for the control network, from epoch 300 to 500 and 2000 to 2750 respectively. The same portion of the curve is shown by figures 2.7 and 2.9 for the fixation network. On these magnified curves, it can be clearly seen that the differences in the curves are larger for the fixation net. In figure 2.7 it can also be seen that the shape of the curves matches their respective frequencies. For example position 4 has the lowest MSE and the highest frequency and position 1 has the highest MSE and the lowest frequency. In the corresponding figure for the control net, there is a slight difference in MSE at epoch 300 but the curves move much closer together at epoch 500. As previously

Document info
Document views294
Page views308
Page last viewedFri Jan 20 02:19:48 UTC 2017