Edward W. Large
Dept. of Comput. and Info. Sci., Ohio State Univ., 2036 Neil Ave., Columbus, OH 43210
According to music psychologists, listeners internalize characteristic patterns of rhythm, pitch, harmony, and so forth, which are then used to recode and organize sequentially presented musical stimuli. In a previous paper, a connectionist model was proposed for performing the recoding task in music [Large et al., Proc. of the 13th Conf. of the Cog. Sci. Soc. (1991)]. A compressor network recursively recodes musical patterns into fixed-width neural vectors. A reconstructor network decodes the vectors produced by the compressor into facsimiles of the original musical patterns. These mechanisms are trained simultaneously by linking their training sets together using an auto-associative form of backpropagation. It has been shown that the reconstructions produced by the network agree both with empirical data from studies of improvisational music performance, and with reductionist theories of mental representation for music [F. Lerdahl and R. Jackendoff, A Generative Theory of Tonal Music (MIT, Cambridge, MA, 1983)]. Here, the behavior of the neural network is analyzed from a dynamical systems perspective. The analysis helps explain how the model computes reductions of the musical surface, while minimizing loss of information. It also provides insight into the generalization capabilities of this architecture.