Abstract

Sequence learning is an important process involved in many
cognitive tasks, and is probably one of the most important processes governing music processing tasks. In this work we build and evaluate computational models addressed to solve a tonesequence learning task in a framework which simulates forcedchoice tasks experiments. The specific approach we have selected is that of Artificial Neural Networks in an on-line setting, which means the network weights are always updated when new events are presented. Here, we aim at simulating the findings obtained by Saffran, Johnson, Aslin, and Newport (1999). We propose a validation loop that follows the experimental setup that was used with human subjects, in order to characterize the networks’ accuracy to learn the statistical regularities of tone sequences. Tone-sequence encodings based on pitch class, pitch class intervals and melodic contour are considered
and compared. We simulate the forced-choice task by
selecting the attended tone-word which is best expected among a tone-word pair. The experimental setup is extended by introducing a pre-exposure forced-choice task, which makes it possible to detect an initial bias in the model population prior to exposure. Two distinct models (i.e. Simple Recurrent Network or a Feedforward Network with a time window of one event) lead to similar results. We obtain the best match with the ground truth using an encoding based on Pitch Classes, that is, based on an absolute pitch representation instead of intervals.
Furthermore, we highlight the impact of tone sequence
encoding in both initial model bias and post-exposure discrimination accuracy and suggest that melodic encoding should be further investigated in the modeling of psychological experiments involving musical sequences.