Rule learning by Habituation can be Simulated in Neural Networks

Contrary to a recent claim that neural network models are unable to account for data on infant habituation to artificial language sentences, the present simulations show successful coverage with cascade-correlation networks using analog encoding. The results demonstrate that a symbolic rule-based account is not required by the infant data. One of the fundamental issues of cognitive science continues to revolve around which type of theoretical model better accounts for human cognition -a symbolic rulebased account or a sub-symbolic neural network account. A recent study of infant habituation to expressions in an artificial language claims to have struck a damaging blow to the neural network approach (Marcus, Vijayan, Rao, & Vishton, 1999). The results of their study show that 7month-old infants attend longer to sentences with unfamiliar structures than to sentences with familiar structures. Because of certain features of their experimental design and their own unsuccessful neural network models, Marcus et al. conclude that neural networks cannot simulate these results and that infants possess a rule-learning capability unavailable to neural networks. A companion article suggests that rule learning is an innately provided capacity of the human mind, distinct from associative learning mechanisms like those in neural networks (Pinker, 1999). My paper presents neural network simulations of the key features of the Marcus et al. (1999) experiment, thus showing that their infant data do not uniquely support a rule-based account. Psychological Evidence and One Interpretation Marcus et al. (1999) present experiments in which 7-monthold infants habituate to three-word sentences in an artificial language and are then tested on novel sentences that are either consistent or inconsistent with those to which the infant has habituated. In one experiment, illustrated in the first three columns of Table 1, infants habituated to sentences exhibiting an ABA pattern, for example, ga ti ga or li na li. There were 16 of these ABA sentences, created by combining four A words (ga, li, ni, and ta) with four B words (ti, na, gi, and la). Then the infants were presented with two novel sentences that were consistent with the ABA pattern (wo fe wo, and de ko de) and two novel sentences that were inconsistent with ABA because they followed an ABB pattern (wo fe fe, and de ko ko). A second, control condition habituated infants to sentences with an ABB pattern, for example, ga ti ti and ga na na. Again, 16 such sentences were created by combining the four A words with the four B words. The test sentences were the same in this second condition, but here the novel ABB sentences were consistent and the novel ABA sentences were inconsistent with the habituated ABB pattern. Table 1: Conditions and error in simulation of Experiment 1 Procedure Condition 1 Condition 2 Mean SE Habituate ABA ABB Consistent ABA ABB 0.649 0.107 Inconsistent ABB ABA 1.577 0.088 The dependent measure was looking time. During the test phase, if the infant looked at a flashing light to her left or right, a test sentence was played from a speaker near that light. A test sentence was played over and over until the infant either looked away or until 15 s elapsed. Infants attended more to inconsistent novel sentences than to consistent novel sentences, indicating that they were sensitive to grammatical differences between the sentences. Marcus et al. designed another experiment, described in the first three columns of Table 2, that contrasted habituation to ABB sentences with AAB sentences. The idea was to rule out the possibility that infants might have used the presence or absence of duplicated words to distinguish grammatical types in their other experiments. For example, ABA sentences duplicate no words, but ABB sentences do (by duplicating B). In this Experiment 3, both grammatical sequences have duplicated words. Table 2: Conditions and error in simulation of Experiment 3 Procedure Condition 1 Condition 2 Mean SE Habituate ABB AAB Consistent ABB AAB 0.570 0.100 Inconsistent AAB ABB 1.491 0.072 Infants performed in a similar fashion in both experiments, i.e., they attended more to inconsistent than to consistent novel sentences. All infants except one showed the predicted preference for inconsistent over consistent test sentences. The issue is the proper theoretical account of this grammatical knowledge -is it based on rules or on connections? Marcus et al. argue that these simple grammars could not be learned by a computational system that is sensitive only to transitional probabilities or event frequencies.