Information Technology Reference
In-Depth Information
reasonable performance: if
stim_err_on
is less than
.2, and
stim_err_off
is less than .4, then we say
that the hippocampus has remembered the item. This
is shown in the
rmbr
column in the text log, with a 1
for successful remembering. Ignore the other columns
at this point, because they are not relevant to this simu-
lation.
Then
Run
the process. Monitor the training and testing
text logs as the network is trained and tested.
Question 9.7 (a)
Again report the total number of
rmbr
responses from your testing text log for the first,
second, and third testing environments (
Test_AB
,
Test_AC
, and
Lure
, respectively) after each epoch
of training on AC.
(b)
Do you find evidence of any in-
terference from learning AC on the testing results for
AB?
(c)
Compare and contrast the performance of
this hippocampus model with that of the cortical model
and the human data of this same basic task from sec-
tion 9.2.2, paying particular attention to both interfer-
ence and number of epochs necessary to learn.
Continue to
StepTest
for a few more patterns.
Make sure you understand the relationship between
the network's performance and the statistics in the logs.
,
!
Then, turn
test_updt
back to
NO_UPDT
(or
TRIAL_UPDT
if you want to see the end state of recall
for each event), and continue to step until all 10 studied
items have been tested. Then,
StepTest
again.
You will see that the
epc_ctr
field in the testing
text log increments from 0 to 1. Now we are presenting
the very same βAβ stimulus, but with the list context for
the second (AC) pairing of items (which has yet to be
trained). Thus, we expect that the network will not suc-
cessfully recall these items, because it has not learned
of them yet.
Go to the
PDP++Root
window. To continue on to
the next simulation, close this project first by selecting
.projects/Remove/Project_0
. Or, if you wish to
stop now, quit by selecting
Object/Quit
.
9.3.4
Summary and Discussion
StepTest
for the next 10 items, the AC list.
The network should not
rmbr
any of them, but
you might notice that the
stim_err_off
values are
somewhat smaller than you might expect for items that
the network has not been trained on, which is caused by
the similarity of these probes to the studied items (they
overlap 50%). The next set of testing events is a com-
pletely novel set of lure items (i.e., βDEβ), to which the
network should give high
stim_err_off
values in-
dicating lack of prior exposure to any aspect of these
items.
We have seen that the pattern separation capabilities of
the hippocampal formation as captured in our model en-
able it to rapidly and sequentially learn arbitrary infor-
mation (e.g., the AB-AC paired associate lists) with-
out suffering from massive levels of interference. This
unique ability suggests that the hippocampal system
plays a complementary role to the slow-learning cor-
tical system that forms richly overlapping distributed
representations based on task and model learning prin-
ciples. Importantly, although this story is supported and
motivated by relevant behavioral and neuroscience data,
it is founded on basic computational principles.
For example, we can understand interference in terms
of the shared use of units and weights for different asso-
ciations, which makes it clear that pattern separation is
critical for avoiding interference. However, pattern sep-
aration is incompatible with the pattern overlap in dis-
tributed representations that is useful for generalization,
similarity-based reasoning, and efficient encoding of
high-dimensional, complex environments. This tradeoff
between sparse, pattern-separated representations and
overlapping distributed representations, in addition to
the constraint that cortical learning requires slow, inter-
leaved learning, motivates the idea that there are two
StepTest
for the next 10 items, the Lure list.
Question 9.6
Report the total number of
rmbr
re-
sponses from your testing text log for the first, second,
and third testing environments (
Test_AB
,
Test_AC
,
and
Lure
, respectively).
Now, having trained on the AB associates, and tested
the resulting performance, we will train the network for
two epochs on the AC associates, and then examine the
testing log as the network is automatically tested after
this training epoch, looking for signs of interference.
Set the
train_epcs
field to 5 instead of 3. Then,
set
train_env
to
TRAIN_AC
instead of
TRAIN_AB
.
Search WWH ::
Custom Search