Cognitive modeling of symbolic-like relationships with the adaptive neural network associator (ANNA)

In their influential articles (Science, 283, 1999), Marcus, Vijayan, Bandi Rao and Vishton (pp. 77-80) and Pinker (pp. 40-41) argue that a prominent model of associative learning, the simple recurrent network, SRN, would fail to simulate rule-learning by seven-month-old infants. Furthermore, the authors argue for the consideration of rule based, symbolic explanations. Subsequently, several authors proposed variations of the simple recurrent network that were better able to model the infant data, but Marcus argued that these models would themselves implement (hidden) rule-mechanisms. Moreover, he was able to show in a further experimental test that one of these models predicted exactly the opposite of what was found in an infant learning experiment. The paper proposes ANNA, a new recurrent neural network architecture that is fully based on associative mechanisms, i.e. ANNA does not implement rules. ANNA succeeds in simulating the infant results. This includes Marcus´ recent experimental tests. The author therefore argues that the results of Marcus do not necessarily prove that infants make use of rules (though they might apply rules). Moreover, ANNA shows that rule/symbolic-like relationships can at least sometimes arise out of associations.