Memory - Computational Explorations in Cognitive Neuroscience - page 309

ABC ABC

ence in shaping such weights has been demonstrated
elsewhere (Munakata, McClelland, Johnson, & Siegler,
1997).
The input representations for the model are based on
the idea that spatial location and object identity are pro-
cessed in separate pathways, as discussed in chapter 8.
Thus, one pathway represents the location of the hidden
toys, while another represents the identities of the toy(s)
and the lids that cover the toys when hidden. These
feed into a hidden layer that represents the locations,
but which also receives input from the object represen-
tations. This hidden layer, which has self-recurrent con-
nections to each representation of a given location, rep-
resents a prefrontal-like active memory system.
The model has two output layers (for reaching and
gaze/expectation); the single difference between them is
the frequency of their resp on ses (i.e., the updating of the
unit activity) during the AB task. The gaze/expectation
layer responds to every input, while the reaching layer
responds only to inputs corresponding to a stimulus
within “reaching distance.” This updating constraint on
the output layers is meant to capture the different fre-
quencies of infants' reaching and gaze/expectation dur-
ing the AB task. Reaching is permitted at only one
point during each trial - when the apparatus is moved to
within the infant's reach. In contrast, nothing prevents
infants from forming expectations (which may underlie
longer looking to impossible events) throughout each
trial. Similarly, al th ough infants' gaze is sometimes
restricted during AB experiments, infants nonetheless
have more opportunities to gaze than to reach. As we
shall see in the simulations, more frequent responses
can change the dynamic between active and weight-
based traces, resulting in dissociations between looking
and reaching.
The network's initial connectivity includes a bias to
respond appropriately to location information, e.g., to
look to location A if s om ething is presented there. In-
fants appear to enter AB experiments with such biases.

Gaze/
Expectation

Reach

ABC

Hidden

ABC C1 C2 T1 T2

Location Cover Toy

Figure 9.22: The AB network, with location, cover, and toy
inputs, and gaze/expectation and reach outputs. The internal
hidden layer maintains information over the delay while the
toy is hidden using recurrent self-connections, and represents
the prefrontal cortex (PFC) in this model.

cations A , B ,and C , represented in the location-based
units in the network. Also, there are two cover (lid) in-
put units corresponding to C1, the default cover type,
and C2, a different cover type, and two toy units corre-
sponding to T1, the default toy, and T2, a different toy
type.

Now, click on r.wt and observe the connectivity.
Each of the three input layers is fully connected to
the hidden layer, and the hidden layer is fully connected
to each of the two output layers. You can see that
there is an initial bias for the same locations to be more
strongly activating, with weights of .7, while other loca-
tions have only a .3 initial connection weight. Connec-
tions from the toy and cover units are relatively weak
at .3. The hidden and output layers have self-recurrent
excitatory connections back to each unit, which are ini-
tially of magnitude .1, but we can change this with
the rec_wts parameter in the control panel. Stronger
weights here will improve the network's ability to main-
tain active representations. We are starting out with rel-
atively weak ones to simulate a young infant that has
poor active maintenance abilities.
Now, we can examine the events that will be pre-
sented to the network.

9.6.2 Exploring the Model

Open the project ab.proj.gz in chapter_9 .
Let's examine the network first (figure 9.22). Notice
that there are three location units corresponding to lo-