6533b830fe1ef96bd1296fb5

RESEARCH PRODUCT

RECURRENT SELF-ORGANIZATION OF SENSORY SIGNALS IN THE AUDITORY DOMAIN

subject

description

In this study, a psychoacoustical and connectionist modeling framework is proposed for the investigation of musical cognition. It is suggested that music perception involves the manipulation of 1) sensory representations that have correlations with psychoacoustical features of the stimulus, and 2) abstract representations of the statistical regularities underlying a particular musical syntax. In the implicit learning domain, sensory features have been shown to interact with the processes involved in the extraction of the regularities governing musical events combinations in a stream [e.g., 1]. Furthermore, in a more ecological context, it is well known that traditional Western tonal system has sought a great convergence between sensory and syntactic factors. The present research aims at investigating the effects of the sensory coding simulated by an auditory model of pitch perception [2] on the representations of the sequential regularities developed in a recurrent connectionist model. According to Arbib, the brain can be described as “layered, somatotopic, distributed computer”. The auditory cortex provides an excellent example of a somatotopic processing array, as it shows tonotopic (pitch-dependent) organization in multiple processing stages. In an effort to model the somatotopic maps found in the cerebral cortex, Kohonen[3] developed the Self-Organizing Map (SOM). Although it has produced very good results with static inputs, it is often pointed out in the literature [4,5] that the standard SOM is not designed for time-domain processing. Yet, music, like language, is a highly structured domain, in which a set of principles governs the combination of discrete structural elements into sequences. These combinatorial principles can be observed at multiple levels, such as the formation of chords, chord progressions and keys, and are at the origin of various temporal dependencies between elements. Extensions of the SOM that learn temporal dynamics have been proposed [4,5]. These models contain a useful idea: recurrent, temporal feedback in addition to the purely spatial recurrent excitation/inhibition used in the conventional SOM. Each incoming signal is thus associated with a contextual signal which reflects the current state of the map. Hence, these models show the ability to maintain state and memory based on past input while engaging in self-organizing learning and context recognition, making them strong candidates for modeling processes involved in music cognition. Through a series of simulations, I will first investigate this type of connectionist network as a model of human sequential learning. I will then show that sensory signals shape the representations of sequences developed by these recurrent models. The strength with which sensory signals and contextual signals interact during learning determines the type of topology realized in the topographic maps (i.e., spatially or temporally defined signal topology). More specifically, the interactions between bottom-up representations built by an auditory model of pitch perception and an emergent syntactic-like knowledge in recurrent self-organizing networks working on these sensory signals can give an account of some experimental results in music cognition literature. Furthermore, the general learning dynamic of the maps could explain the developmental interactions of sensory and syntactical characteristics of the musical environment. When units become more and more specialized in the time-domain, sensory effects tends to become weaker, compared to syntactical effects.