Chapter 2. The Simulations
Control Network Fixation Network
Table 2.8: Errors made by the network in the different slots of the output pattern. The
output is subdivided into the onset, nucleus and coda positions. Each of these positions
has two slots associated with it.
So a word like ’back’, whose phoneme is ’b&k’ will have a target output b.|&.|.k, where ’.’ indicates an empty slot and ’|’ separates onset|nucleus|coda. The reason for left aligning the the onset but right aligning the coda is to have a clear mapping be- tween the exterior letters and the exterior phonetic features. In this way, the first letter of the word (if it is a consonant) will always correspond to the first phonetic feature and the last letter (if it is a consonant) will always correspond to the last phonetic feature. Right aligning the coda in this way gives more structure to the input-output mapping. If for example the coda was left aligned the last letter would sometimes be mapped to the first coda slot (when there are two consonants in the coda) and sometimes to the second coda slot (when there is only one consonant in the coda).
Again, the results are not conclusive as to a difference between the two networks. The fact that the fixation network was trained on more words fixated to the right of centre and hence with most of the word in the left visual field and the right hemisphere