Chapter 2. The Simulations
gives the mean sum-squared error for the whole training set at each of the fixation positions.
errors if the word is fixated slightly left of centre than it does when the word is fixated right of centre. According to this table, the best fixation positions at which to present the word to the network would be positions 2 or 3 - fixated before the 2nd and 3rd letter respectively. Comparing this to the mean sum-square errors given in table 2.10 one can see that once again, the error score for fixation position 2 is lowest of all. However, for example fixation position 5 has the 2nd lowest mse, although the network makes the
most errors at this position.
Taken together, these results (lowest error score and lowest mse at fixation position 2) point to a preferred viewing location on or slightly left of centre. At a first glance this is surprising since the words were presented with the same frequency at each of the positions. If there is a preferred position, one would rather expect that it is easier for the network to learn words that are entirely in one of the two hemispheres. This is because in such a case all of the information needed to identify the word is transmitted to the corresponding hidden layer. Thus in theory there would not need to be a correspondence with the other hidden layer. One would expect that if the word is divided into the two layers, the process of ’re-uniting’ the word would be a harder process.
However, a preferred viewing position slightly left of centre corresponds exactly to Nazir (2000) and Shillcock et al. (2000) as described in section 2.3. It is interesting that this preferred viewing position should emerge from the network without any initial conditions that would bias it towards this behaviour. It suggests that, at least in the network, this behaviour emerges solely from the structure of English words and seems to support Shillcock et al. (2000) in that a fixation position slightly left of centre is