X hits on this document

PDF document

The split-fovea model of visual word - page 13 / 78





13 / 78

Chapter 1. Introduction and Background


to the hidden units as in figure 1.3. The assumption here was that a feed-forward archi- tecture was enough for a net that only has to learn how to pronounce words and which does not have to worry about the semantics and the context in which the words appear.

For the orthographic as well as the phonological representations in the network, Sei- denberg and McClelland (1989) chose a Wickelfeature scheme (Wickelgren, 1969). In this scheme the representation is of a word is made in terms of triples. For example the orthographic form of the word PINT would be represented by PI, PIN, INT, NT . This representation was necessary in order to avoid the problem that words like TUB and BUT would have the same representation since the same letters are activated for both of them. Note that these orthographic and phonological representations of the words are distributed activation patterns of the orthographic units and as such are fundamentally different from the orthographic lexicon as employed in Coltheart’s Dual-Route model. There is no such lexicon in this model but rather the orthographic representation of a word is distributed over several units. When a word is presented to the network, activation spreads to the hidden layer and from there to the phonological layer. For example the activation of a hidden unit is determined by the activation and weights of all the input units connected to it.

The network was trained on 2897 monosyllabic English words for 250 runs, with words being presented at random to the network at each run. The words were ran- domly selected according to the frequency with which they appear in the English lan- guage. However, rather than using the actual frequencies, logarithmic frequencies were used to reduce the time of training1. Learning was done by the standard error back- propagation algorithm (Rumelhart et al., 1986; O’Reilly and Munakata, 2000). This algorithm compares the computed output to the the correct target output and uses the

distance between the two as an error score. 1For further details on logarithmic frequency see chapter 2

Document info
Document views272
Page views286
Page last viewedTue Jan 17 09:37:49 UTC 2017