Chapter 2. The Simulations
model that could differentiate between two different contexts or meanings and so the network is unable to learn which pronunciation to apply.
The overall results for both networks can be seen in table 2.1. The table shows the networks’ performance after 2750 epochs (∼4 million words). The upper part of the table shows the statistics relating to the number of events that the network was trained on. In overall performance, both networks are very similar, with the control network getting 98.58 percent of the events in the training corpus correct and the fixation net- work getting 98.61 percent correct. This very slight difference of 0.03 percent (or three words) between the networks is so small that it is negligible and one can say that both networks performed the task equally well. The fact that the control network only made three errors more than the fixation network can not be considered as evidence that the fixation network performed better than the control. This first result is not surprising as both networks were trained on exactly the same training corpus and with exactly the same initial conditions, with only a very small difference in fixation position frequen- cies.
The bottom part of table 2.1 shows how many words the networks got correct in every position. The statistics in the upper part of the table relates to the individual events and so includes each word five times, once in each of the different possible fixation positions. The bottom part on the other hand only considers a word to be correct if the network computed the right output for each of the fixation positions. Hence, for example the control network only made a mistake on the word ’lewd’ when it was fixated at the second position (network: ’lUUd’ - Correct: ljUUd). So ’lewd’ is counted as an error in the bottom part of the table since it has not been learned correctly at all of the fixation positions. Again, it can be seen from the table that both networks are closely matched in their overall performance. However there is a small difference in that the fixation network gets 8 words more wrong here than the control.