Memory - Computational Explorations in Cognitive Neuroscience - page 289

epc_ctr trial Event stm_err_on stm_err_off rmbr

1−

0 0 i0_i0_c0i0_ 0.0555556 0.0833333 1

Lures

0 1 i1_i1_c0i1_ 0.416667 0.555556 0

0.8−

0 2 i2_i2_c0i2_ 0.166667 0.555556 0

0 3 i3_i3_c0i3_ 0.111111 0.194444 1

0.6−

0 4 i4_i4_c0i4_ 0.0277778 0.0833333 1

0 5 i5_i5_c0i5_ 0.0833333 0.166667 1

0.4−

0 6 i6_i6_c0i6_ 0.0833333 0.222222 1

0 7 i7_i7_c0i7_ 0.0833333 0.138889 1

Trained

0 8 i8_i8_c0i8_ 0.277778 0.555556 0

0.2−

0 9 i9_i9_c0i9_ 0.0833333 0.25 1

1 0 i0_i10_c1i0_ 0 0 1

0−

1 1 i1_i11_c1i1_ 0.0555556 0.194444 1

1 2 i2_i12_c1i2_ 0.0277778 0.0555556 1

0 0.2 0.4 0.6 0.8 1

1 3 i3_i13_c1i3_ 0 0.0277778 1

stm_err_off

1 4 i4_i14_c1i4_ 0 0.0277778 1

1 5 i5_i15_c1i5_ 0 0.25 1

1 6 i6_i16_c1i6_ 0 0.194444 1

1 7 i7_i17_c1i7_ 0 0.166667 1

Figure 9.14: Composite graph log display for testing the
hippocampal network, showing two kinds of errors — units
that should be off but were erroneously on ( stm err on ),
and units that should be on but were erroneously off
( stm err off ). The training items have relatively few of
both types of errors, and the lure items are mostly inactive
(high stm err off .

1 8 i8_i18_c1i8_ 0.111111 0.333333 1

1 9 i9_i19_c1i9_ 0.0555556 0.0833333 1

2 0 i20_i20_c2i0_ 0.388889 0.777778 0

2 1 i21_i21_c2i1_ 0.472222 0.833333 0

2 2 i22_i22_c2i2_ 0.166667 0.916667 0

2 3 i23_i23_c2i3_ 0.444444 0.611111 0

2 4 i24_i24_c2i4_ 0.361111 0.777778 0

2 5 i25_i25_c2i5_ 0.416667 0.666667 0

2 6 i26_i26_c2i6_ 0.555556 0.722222 0

2 7 i27_i27_c2i7_ 0.666667 0.805556 0

2 8 i28_i28_c2i8_ 0.166667 0.916667 0

2 9 i29_i29_c2i9_ 0.527778 0.611111 0

Figure 9.13: Text log display for testing the hippocampal
network. Shows results after training on both AB and AC
lists, with 70 percent remember responses for the AB list and
100 percent for the AC list.

stim_err_on and stim_err_off . The first shows
the proportion of units that were erroneously activated
in EC_out (i.e., active but not present in the targ
pattern), and the second shows the proportion of units
that were erroneously not activated in EC_out (i.e.,
not active but present in the targ pattern). When
both of these measures are near zero, then the net-
work has correctly recalled the original pattern. A large
stim_err_on indicates that the network has confab-
ulated or otherwise recalled a different pattern than the
cued one. This is relatively rare in the model (O'Reilly
et al., 1998). A large stim_err_off indicates that
the network has failed to recall much of the probe pat-
tern. This is common, especially for the novel lure items
in the last testing set (as we will see).
Now, look at the graph log just to the right of this
testing text log ( Trial_1_GraphLog , figure 9.14).
This shows the stim_err_on plotted on the Y axis
against stim_err_off on the X axis, with each event
showing up as a dot at a particular location. To the ex-
tent that these dots are in the lower left hand corner, the
network is recalling accurately. This log is automati-
cally cleared before each new testing set, so you can be
sure all the dots are for one particular testing environ-
ment. To code discrete responses from the network, we
need to set thresholds for these statistics. The following
thresholds were somewhat arbitrarily chosen to provide

cycle basis. Then, let's Clear the testing text log
( Trial_1_TextLog ). Then, do StepTest .
You will see the first testing event in the EC_in layer,
which corresponds to a studied A stimulus, an empty
gap where the B stimulus would be, and a list context
representation for the AB list. Since this was studied,
it is likely that the network will be able to complete the
pattern, which you should be able to see visually as the
gap in the EC activation pattern gets filled in.
When the network has stopped settling on this pat-
tern, you can compare the activity pattern produced over
EC_out with the original stored activity pattern.

To do so, press the targ button in the network win-
dow.
You should see that there is a target pattern (which is
only used for our comparison purposes) for this layer,
and because the active units are selected, you can rela-
tively easily compare these two (alternatively, you can
compare by switching back and forth between act and
targ ).
Now, look at the first line of the testing text log
(figure 9.13).

, !

The two critical columns here are