Studying the effects of memory size on melody recognition using a neural network simulation of cohort theory

Dalla Bella, Peretz, and Aronoff (2003) compared performance between musicians and nonmusicians in the perception and recognition of known, isolated melodies. They used a gating task to identify three events in the time course of melody recognition. These events are the familiarity emergence point (FEP), the isolation point (IP), and the recognition point (RP). Dalla Bella et al. explain their results using the cohort theory. Inspired by the cohort theory, this research explains the development of a simulation based on a sequence recognition neural network to model the hypothesized cognitive processes underlying the occurrence of these three events. Specifically, the IP is modeled using the core sequence recognition network which implements a winner-take-all mechanism. The FEP and RP are modeled by introducing meta-level readout units that monitor the dynamically evolving state of the core network. Two separate networks with different stored memory sizes are used to model musicians and nonmusicians. The simulations presented in this research capture the qualitative results of Dalla Bella et al. and show how stored memory size may affect the melody recognition process. The unit that models the FEP plays a causal role in determining the RP. It was not previously known whether the core network contained enough information to allow simple meta-level readout of dynamic cognitive states. The simulations showed that such readout is in fact possible. The modeling framework is then extended to illustrate melody frequency effects by using trace strengths within the network, and tested for robustness using two additional corpora. A method for the acquisition of new melodies is introduced. Preliminary studies incorporating hierarchy into the existing framework are also presented.