Language - Computational Explorations in Cognitive Neuroscience

Information Technology Reference

In-Depth Information

Code

Meaning

Example

Now we will go through the contents of the

text log (figure 10.15). The word input to the

network is shown in the Event column (e.g.,

best_best_bbbestt_HRC ), and has four compo-

nents: the orthography ( best ), the phonemes ( best ),

the repeated consonant phonology ( bbbestt ), and a

special status code ( HRC , see table 10.5). The pronun-

ciation that the network actually produced is shown in

the ph_out column, which should show that the cor-

rect output was produced. Note that this output may oc-

casionally contain an “X,” indicating that the phoneme

in this position did not exactly match one of the valid

phoneme patterns. Also shown are the number of cy-

cles it took for the network to settle, the training epoch,

and the sum_se error.

Let's continue to observe the network's reading per-

formance, observing specifically the translation invari-

ance property.

HRC

High freq regular consistent

best

HRI

High freq regular inconsistent

bone (c.f., done)

HAM

High freq ambiguous

brown (c.f., blown)

HEX

High freq exception

both (c.f., cloth)

LRC

Low freq regular consistent

beam

LRI

Low freq regular inconsistent

brood (c.f., blood)

LAM

Low freq ambiguous

blown (c.f., brown)

LEX

Low freq exception

bowl (c.f., growl)

Table 10.5: Codes for the probe words used to test the net-

work and its settling times. Regular consistent means there

are no other inconsistent pronunciations, whereas regular in-

consistent words have other inconsistent versions (examples

given). Ambiguous is where there is no clear regularity, and

exceptions are the exceptions to the regularities.

env

trial

Event

best_best_bbbestt_HRC

big_big_bbbiggg_HRC

came_kAm_kkkAmmm_HRC

class_kl@s_kll@sss_HRC

dark_dark_dddarkk_HRC

sum_se

lst_cycles

ph_out

bbbestt

bbbiggg

kkkAmmm

kll@sss

dddarkk

epoch

410

StepTest several more times.

You should notice that the “best” input appears in

successively more rightward positions in the input. De-

spite these differences in input location, the network

produces the correct output. This spatial invariance cod-

ing, like the one we explored in chapter 8, requires the

network to both maintain some information about the

local ordering of the letters (so it pronounces “best” in-

stead of “steb,” for example), but also treat the entire

pattern the same regardless of where it appears. We will

see in a moment that this network developed the same

general solution to this problem as the object recogni-

tion network, using a combination of locally spatially

invariant and yet conjunctive encoding.

You can continue to observe the network's perfor-

mance, and speed up the process by controlling the rate

at which the network display is updated.

Figure 10.15: Trial 1 TextLog showing the output of

the network as it reads a series of words in different locations.

Click on the t button in the consonants window, and

then scroll down to the s button (it's number 11), and

select it with the middle button (or shift-left button).

The patterns should match those produced by the net-

work. Next, we can examine the vowel /e/.

To switch the network updating to only update

after each trial, instead of each cycle, set net_updt

to TRIAL_UPDT instead of CYCLE_UPDT . Continue to

StepTest .

Although you may observe an occasional error (espe-

cially as the items get lower in frequency and more ir-

regular), the network should pronounce most words cor-

rectly — no small feat itself given that there are nearly

3,000 words presented in as many as 4 different loca-

tions each!

Select View , VOWELS . Then click on the e button.

You should see that this pattern matches the central

vowel slot pattern.

Fortunately, you don't need to go through this pattern

matching process every time, because the simulator can

do it for you, with the results shown in a text log.

, !

Iconify the consonant and vowel windows, and

then do View , TEST_LOG , which will bring up the

Trial_1_TextLog window. Do StepTest again.

, !

Computational Explorations in Cognitive Neuroscience

Search WWH ::

Custom Search

Home