X hits on this document

PDF document

The split-fovea model of visual word - page 60 / 78





60 / 78

Chapter 3. Discussion and Conclusion


This validates the networks as models of visual word recognition. Having established this, the next step is to look at the differences between the two networks that were trained.

Differences in the networks.

The main part of this paper was to investigate whether

there was a significant difference between the control network and the fixation network. Several tests were made to detect any such differences. In both simulations, there was no significant difference found in the overall performance of both networks. The main finding of these tests were that the performance (as measured by the number of errors as well as the mean sum-square error over all the words in the corpus) of the control network was best when the words presented were fixated slightly left of centre. For the fixation net, this tendency was turned around and performance was better for words

presented slightly right of centre.

The behaviour of the control network corresponds to the optimal viewing location during single word recognition, where a word at a time is presented to human subjects and the fixation positions are measured (e.g. Nazir (2000)). This is quite interesting since there was no initial bias for the network to evolve in this way. It seems to arise simply out of the structure of English words, where the beginning of a word is usually more informative than the end of the word. A fixation position slightly left of centre will thus ensure that an equal amount of information will be sent into both of the hidden layers, agreeing with Shillcock et al. (2000)’s claim that the optimal viewing location is beneficial to word reading. Note that the same result was obtained in the second simulation with less words in the training corpus, which helps to establish that the previous result was not just a random occurrence.

The preferred fixation position for the fixation net on the other hand is slightly right of centre. This is not surprising since in this network the fixation position with the highest frequency was position 4, just right of the word centre. Thus, during training, the net-

Document info
Document views284
Page views298
Page last viewedWed Jan 18 23:09:05 UTC 2017