Figure 11.

Lack of hyper-generalization of remote fear. (I) Hyper-generalization of remote fear fails to occur if the test context is familiar. If there is already a representation of the test context, BaconREM does not initially mis-identify the test context as the conditioned one but instead activates the test context’s own representation. Under these circumstances, there is somewhat more generalization when both contexts are remote than when both are recent because the generalization function (Figure 7) has a higher value for any given degree of attribute overlap. However, the extreme hyper-generalization seen when the test context has no representation of its own does not occur. (II) Hypo-generalization of remote fear can occur if (unfamiliar) test contexts are in a different category from the conditioned one. In this simulation, the overlap of the particular attributes of contexts A and C (of different categories) was 0.75. As the bar graph shows, there was less generalization in the Remote than in the Recent case. In the Recent case, Rep A stopped being activated a little less than halfway through the session. In the Remote case, Rep A and the established representation that was in the same category competed for activation for about the first 3/4 of the session. However, because systems consolidation had caused additional categorical attributes that differentiated context A from C to get added to Rep A, B_Rep was lower when Rep A was active in the Remote than in the Recent case. Thus, even when Rep A was active in the Remote case, B_Rep was very low, and hence, fear expression was slight, as can be seen by comparing the first half of the Recent and Remote generalization tests. When, in both the Recent and Remote sessions, Rep C was created, fear became very low. (III) Hyper-generalization of remote fear to an unfamiliar context fails to occur if a great deal is known about the conditioned context. That is because, as discussed with respect to Figure 5, if the number of attributes associated with the hippocampal version of a representation is especially large, a greater proportion of them tends to get copied to the cortical version. This is illustrated in this experiment where BaconREM was pre-exposed to context A until Z_cur = 98, so a great deal would be known about it. As the bar graph shows, there is slightly more generalization in the remote than in the recent case, but nothing as extreme as in Figure 9.