Memory - Computational Explorations in Cognitive Neuroscience - page 310

ter each event. The other columns show the weights,
which can be somewhat more difficult to interpret at this
point, so we will postpone that.
You should notice that when the toy is presented dur-
ing the p-toy-pres events, the corresponding hid-
den A location is also activated and the network “looks”
toward this location in the gaze/expectation layer. Be-
cause Hebbian learning is taking place after each trial,
those units that are coactive experience weight in-
creases, which in this case increases the propensity of
the network to activate the A location representations.

Click on View , EVENTS in the ab_ctrl control
panel.
There are three types of trials, represented by the
three rows of events that you see in this window: pre-
trials (corresponding to the “practice” trials provided at
the start of an experiment to induce infants to reach to
A ), A trials, and B trials. Each of these trial types can
be repeated multiple times, as can the events w ithin the
trial. In the version we will be running, the AB task will
consists of four pretrials, two A trials, and one B trial.
Each trial consists of four segments , corresponding to
the experimental segments of an AB trial as follows:
(1) start , covers sit in place on the apparatus before the
experimenter draws infant's attention to a particular lo-
cation; (2) presentation , experimenter draws the infant's
attention to one location in apparatus; (3) delay ,theap-
paratus sits with covers in place; and (4) choice ,theex-
perimenter presents the apparatus with covers in place
for the infant's response (reaching permitted only dur-
ing this segment). During each segment, patterns of ac-
tivity are presented to the input units corresponding to
the visible aspects of the stimulus event; all other input
units have no activity. The levels of input activity rep-
resent the salience of aspects of the stimulus, with more
salient aspects (e.g. a toy that the experimenter waves)
producing more activity.
Now, let's run the network. It is much easier to tell
what is going on in the network by looking at a grid
log display, rather than viewing each trial of activation
separately.

Next press ViewA .
You will see the A testing trials, where the network's
tendency to reach to the A location is assessed. Note
that as a result of the Hebbian learning, the hidden and
output units are more active here than in the pretrials.

Now press ViewB .

Question 9.13 (a) Describe what happens to the
network's internal representations and output (gaze,
reach) responses over the delay and choic e trials. You
should observe the network making the AB error. (b)
Explain why the network is performing as it is in terms
of the interactions between the weights learned from
prior experience on A trials, and the recurrent activity
within each representation.

Now increase the rec_wts parameter to .75 from
the default of .3, and Run and ViewB again.

Question 9.14 (a) Describe how the network responds
(i.e., in the gaze and reach outputs) this time. (b) Given
that the experience-based weights learned during the
experiment are basically the same in this and the previ-
ous case, explain why the network performs differently.

Iconify the events, then do View , GRID_LOG . Then
press the Run button (which does a ReInit and Run of
the training process).
When yo u do this, the network will run through
an entire AB experiment, and record the activations
and weights in the grid log (figure 9.23), which is
not updated dynamically during processing since the
Display button is turned off.
We can simply jump to specific points in the experi-
ment using the buttons in the control panel.

Now decrease the rec_wts parameter to an inter-
mediate value of .47, and Run and ViewB again.

Question 9.15 (a) What happens on the B trial this
time? (b) Explain why the network exhibits these dif-
ferent responses, simulating the dissociation observed
in infants.

Press ViewPre to see the pretrials.
The main columns of interest for now are the Event
column, which tells you which event is being presented,
and the next column over from that, which shows a
miniature rendition of the activations in the network af-