X hits on this document

PDF document

The split-fovea model of visual word - page 31 / 78

216 views

0 shares

0 downloads

0 comments

31 / 78

Chapter 2. The Simulations

26

Control Network

Fixation Network

Total Events

9800

Correct

9661

Errors

139

9800

9664

136

Percentage Correct

98.58

Total Words

1960

Correct

1894

Errors

66

Percentage Correct

96.63

98.61

1960

1886

74

96.22

Table 2.1: Overall results of the two networks after

2750 epochs (or about 4 million

words) of training. The upper table is for all the events

so that each word is represented

five times: once per fixation position. The lower table

is only for individual words.

2.2

Results

This section presents the results of the networks after training for 2750 epochs. It will start off by discussing the overall results about of the networks’ performance and then go on to list more detailed results in order to compare the two networks.

2.2.1

General

In this section some general results about both networks will be discussed, showing that there is virtually no difference in the general overall performance of both networks. A more detailed analysis will follow in subsequent sections. Note that all of the following results only include non-homographic words. An analysis of the 14 homographs (70 events) in the training corpus follows in a separate section. The homographs were left out from the current (and subsequent) analysis because the networks will inevitable get one of the possible pronunciations wrong. This is because there is nothing in the

Document info
Document views216
Page views225
Page last viewedThu Dec 08 02:17:19 UTC 2016
Pages78
Paragraphs3087
Words17414

Comments