Unsupervised Multi-net Simulation: An Application to Child Development

We propose a gated multi-net system that simulates the transition of child language from the one-word stage to the two-word stage. In this system, a modified counterpropagation network simulates child language at the one-word stage, whilst an unsupervised temporal neural network simulates child language at the two-word stage. During simulation, inputs encoding the perceptual and conceptual features of events a child is likely to speak about are simultaneously applied to the two networks. Initially, the system is predisposed to outputting one-word utterances, but as the simulation proceeds, a gating mechanism predisposes the system to output two-word utterances. The system exhibits a gradual transition from one-word utterances to two-word utterances similar to that observed in children undergoing the same transitional phase.