Language - Computational Explorations in Cognitive Neuroscience - page 328

The network is trained by randomly selecting (on
each trial) one of the three main representational layers
to act as the “input” layer , and then setting the others as
targets. Thus, the network learns to take any one aspect
of a word's representation and map it onto the other cor-
responding representations. Standard parameters were
used, with 25 percent hidden unit activity.

name as the input is reflected by a zero in the sm_nm
column (a 1 here would indicate error when the two
names were not the same). As you step through the
patterns, you should monitor these values to make sure
the network is reading correctly. As you step through
the words, also pay attention in the network display to
the timing of the initial activation of the phonological
representations.

Continue to StepTest through the words.

10.3.5 Exploring the Model

, !

Open the project dyslex.proj.gz in

Question 10.1 (a) Do you think the initial phonolog-
ical activation is caused by the “direct” input via or-
thography or the “indirect” input via semantics? (b)
Check for any cases where this initial phonological pat-
tern is subsequently altered when the later input arrives,
and describe what you find. (c) Discuss what this might
imply about the behavior of the network under damage
to either of these pathways.

chapter_10 to begin.
You should see that the network is not constructed —
it is stored as a skeleton to keep the project file relatively
small. Because the network takes some time to train (for
250 epochs), we will just load in a pre-trained network
to begin with.

Do LoadNet in the overall dyslex_ctrl control
panel.

, !

Reading with Complete Pathway Lesions

Normal Reading Performance

We next explore the network's ability to read with one
of the two pathways to phonology removed from action.
This relatively simple manipulation provides some in-
sight into the network's behavior, and can be mapped
onto two of the three dyslexias. Specifically, when we
remove the semantic pathway, leaving an intact direct
pathway, we reproduce the characteristics of surface
dyslexia, where words can be read but access to se-
mantic representations is impaired and visual errors are
made. When we remove the direct pathway, reading
must go through the semantic pathway, and we repro-
duce the effects of deep dyslexia by finding both seman-
tic and visual errors. Note that phonological dyslexia is
a milder form of deep dyslexia, which we explore when
we perform incremental amounts of partial damage in-
stead of lesioning entire pathways.
We begin by lesioning the semantic pathway.

For our initial exploration, we will just observe the be-
havior of the network as it “reads” the words presented
to the orthographic input layer. Note that the letters in
the input are ordered left-to-right from the bottom to the
top.

Press View on the overall control panel and
select TEST_LOGS , to bring up two testing logs
( Trial_2_TextLog , and Epoch_2_TextLog ). Then
do StepTest to test the first item.
You will see the activation flow through the network,
and it should settle into the correct pronunciation and
semantics for the first word, “tart.”
In the large square Trial_2_TextLog , you
should see a record of the network's performance on
this word, including the name of the input word (in col-
umn Event ), and the name of the word that was closest
to the phonemic output produced by the network (in col-
umn ev_nm ). In this case, this should be the same as
the input word, because it pronounced it correctly. The
distance between the actual network phonemic output
and this closest event is shown in the column just be-
fore ev_nm (column min_Phon_dst , which should
show a 0). The fact that this closest output is the same

, !

Set lesion_path on the overall control panel to
SEMANTICS , and Apply .
This does not actually remove any units or other
network structure; it just flips a “lesion” flag that (re-
versibly) deactivates an entire layer. Note that by re-
moving an entire pathway, we make the network rely
on the intact one. This means that the errors one would

, !