Information Technology Reference
In-Depth Information
This will bring up two text logs to display the results
of testing. We will perform two different kinds of tests.
First, we will assess any biases that may exist to re-
spond to the input patterns with either the
a
or
b
out-
puts. This assessment of the
baseline
responses is done
with learning turned off. Second, we will turn learning
on and train the network to alternately produce the
a
output (i.e., present the
a
output in the plus phase and
adjust the weights just as was done during training) and
then the
b
output. When compared with the baseline,
we will be able to see if one trial of learning to produce
a particular output will have a substantial effect on the
probability of producing that output.
First, let's collect the baseline values without learn-
ing. The
wt_update
parameter in the control panel,
which should be set to
TEST
, determines that we will
not do any learning at this point.
output pattern), and a 1 otherwise. You should observe
that roughly 50 percent of the time it produces the same
name as the current event. The smaller log shows you
this result by summing the
sm_nm
column, and also
summing the
both_err
column just to monitor the
actual accuracy of the network in producing one of the
two correct outputs. Note also that the
dist
statistic
is not perfect, but most of the errors are less than 1 in
magnitude.
Look at the summary statistics in the smaller text
log.
You should have observed that the network always
produced a valid response (
a
or
b
) and produced the
a
response roughly 50 percent of the time. Given that
the network was trained equally often to produce the
a
and
b
responses, this is just the kind of baseline perfor-
mance that we would expect.
Now, let's turn learning on, and see if there are any
obvious effects of one trial of learning on a given input
pattern for subsequent performance on that same input.
Do
Test
in the control to present one epoch of
events.
You can see in the
Event
column of the larger text
log that the events were presented in sequential order,
with all the
a
outputs presented first and then all the
b
outputs. Because we are only testing now, the out-
puts were not actually presented (i.e., no plus phase was
run, only a minus phase where only the input pattern is
presented), which means that the first half of the list is
basically the same as the second half, because it is the
same input patterns in both halves. This order of event
presentation will make more sense when we turn on the
training. Any differences in the two halves at this point
are due to the noise added to the unit membrane poten-
tials.
The other columns to pay attention to in the larger
text log are
ev_nm
and
Outp_sm_nm
. The first is the
closest event name for the actual output produced by
the network as described previously. Because it will
generally be producing one of the two correct outputs,
the critical issue is whether it is an
a
or
b
response.
You should see that there is a fairly random set of out-
put responses. To determine how many times the net-
work produced the output pattern corresponding to the
particular target output for the current event, you can
look at the
Outp_sm_nm
column (which we will re-
fer to as simply
sm_nm
). There is a 0 every time the
same name as the event was produced (i.e., the same
Set
wt_update
to
ON_LINE
in the control panel, to
learn (update the weights) after each event (and
Apply
that change).
Then, press
Test
again in the process
control panel.
Now, because we are learning after each event, there
is a plus phase following the minus phase, where the
a
output is trained in the first half of the inputs, and
the
b
output is trained in the second half. Note that
the evaluation of the network's output is all done in the
minus phase,
before
this plus phase training is provided,
so it reflects the prior state of the network.
As a result of having seen the
a
output associated
with a particular input pattern in the first half of the list
of events, we would expect that the network would be
more likely to produce this
a
output when it comes to
that input pattern again in the second half. Thus, unlike
in the baseline case, you should observe a systematic
difference in the responses of the network in the two
halves of the patterns (figure 9.3).
Question 9.1 (a)
What do you notice about the proba-
bility of the network producing the
a
output in the sec-
ond half of the events (i.e., the second time through the
same inputs patterns)?
(b)
How does this affect the
sm_nm
statistic (i.e., what is the summary value for
Search WWH ::
Custom Search