Page 257 - Deep Learning

240 Adaptation

Table 7.4. A run by the HS counting model depicted in Tables 7.2 and 7.3. Each trial rep-
resents one attempt at counting one and the same set of objects. In the right-hand column,
the rule that was revised is underlined.

Trial number Cycles before first Constraint violated Rule revised on
constraint violation each trial
1 1 A 1 2 3 4 5 6
2 1 D 1 2 3 4 5 6
3 2 D 1 2 3 4 5 6
4 2 C 1 2 3 4 5 6
5 2 A 1 2 3 4 5 6
6 3 C 1 2 3 4 5 6
7 3 D 1 2 3 4 5 6
8 3 B 1 2 3 4 5 6
9 4 D 1 2 3 4 5 6
10 4 D 1 2 3 4 5 6
11 4 E 1 2 3 4 5 6
12 4 B 1 2 3 4 5 6
13 4 E 1 2 3 4 5 6
14 5 D 1 2 3 4 5 6
15 5 D 1 2 3 4 5 6
16 5 A 1 2 3 4 5 6
17 6 D 1 2 3 4 5 6
18 6 D 1 2 3 4 5 6
19 7 D 1 2 3 4 5 6
20 7 D 1 2 3 4 5 6
21 9 C 1 2 3 4 5 6
22 10 E 1 2 3 4 5 6
23 11 None None
Based on Ohlsson and Rees, 1991a, Table 5.

its actions. Nor was there a one-to-one mapping between constraints and rules
such that each constraint informed the revision of a single rule. The fourth col-
umn in Table 7.4 shows the rule that was revised on each trial. Some constraints
led to improvements in multiple rules. The two violations of constraint D in tri-
als 7 and 9 led to improvements in rules 4 and 2, respectively. Some rules violated
more than one constraint. For example, rule 2 violated constraints A (trial 5), D
(trial 9) and E (trial 22). There was no one-to-one mapping between constraints
and rules such that each constraint was translated or transformed into a partic-
ular rule. Some rules, like rule 2, required multiple revisions, while others, like
rules 1 and 6, required only two revisions each.