A transfer of sequence function via equivalence in a connectionist network

Connectionist networks may provide useful models of stimulus equivalence an d transfer of function phenomena. Such models have been applied to a range of behavioral tasks and have demonstrated transfers of function via equival ence relations following appropriate training, with networks accurately sim ulating the behavior of human subjects. In the current study, a connectioni st network was pretrained on a series of equivalence and sequence tasks to simulate the preexperimental experience of an adult subject. It was then ex posed to the equivalent of six conditional discriminations, and was tested for the formation of three 3-member equivalence classes (corresponding to A 1-A2-A3, B1-B2-B3, C1-C2-C3). It was subsequently trained to produce a pair of four part sequences (corresponding to B1 --> B2 --> Ct1 --> B3 and B3 - -> B2 --> Ct2 --> B1, where Ct1 and Ct2 represented contextual cues) before being tested for transfer, through equivalence, of the sequence responses to the C stimuli. Following appropriate pretraining, the network showed the formation of three equivalence classes and a transfer of sequence function to the nontrained C stimuli (producing the novel sequences Cl --> C2 --> C t1 --> C3 and C3 --> C2 --> Ct2 --> C1). A control network, which was not e xposed to conditional discrimination training, failed to demonstrate equiva lence and the transfer of sequence function, as predicted by findings from experimental demonstrations with human participants. Network performance wa s analyzed as a function of amount of pretraining and a number of psycholog ically plausible training methods are presented. The data suggest that conn ectionist networks may provide accurate and plausible models of stimulus eq uivalence and transfer of function phenomena in natural language.