'Schema Abstraction" in a Multiple-Trace Memory Model

ion With traces of category exemplars stored in SM, what is the effect of a probe with a category name? To find out, 3, 6, and 9 high-level distortions of three category prototypes were generated. Each was combined with its category name, and one copy was stored in SM, as just described, for a total of 18 memory traces. Testing was then carried out using each of the three category names as probes (features j = 1 ... 10 were filled in and j = 11 ... 23 set at 0). Twenty subjects were simulated by carrying out the entire procedure, beginning with the generation of new prototypes, 20 times. An example of the outcome is shown in Figure 3. Here, the name of the 9-item category was used as the probe. The top panel shows the histogram of the original prototype of the category, and the bottom panel shows the histogram of the echo content across the 13 stimulus features, that is, the values of C(l 1) ... G(23). Clearly the echo is not an exact copy of the prototype, but it does strongly resemble it. The correlation between prototype and echo is r =• .82. Moreover, the echo resembles the prototype more than it does any of the category exemplars that were stored. We know this because each exemplar was generated from the prototype by a rule ensuring that exactly four feature values were reversed in sign, and in the echo only one feature is opposite in sign from the prototype. This shows that MINERVA 2 is capable of retrieving an "abstract idea," even though an abstraction as such was