A Connectionist Approach to Algorithmic Composition

Abstract

With the advent of von Neumann-style computers, widespread exploration of new methods of music composition became possible. For the first time, complex sequences of carefully specified symbolic operations could be performed in a rapid fashion. Composers could develop algorithms embodying the compositional rules they were interested in and then use a computer to carry out these algorithms. In this way, composers could soon tell whether the results of their rules held artistic merit. This approach to algorithmic composition, based on the wedding between von Neumann computing machinery and rule-based software systems, has been prevalent for the past thirty years. The arrival of a new paradigm for computing has made a different approach to algorithmic composition possible. This new computing paradigm is called parallel distributed processing (PDP), also known as connectionism. Computation is performed by a collection of several simple processing units connected in a network and acting in cooperation (Rumelhart and McClelland 1986). This is in stark contrast to the single powerful central processor used in the von Neumann architecture. One of the major features of the PDP approach is that it replaces strict rule-following behavior with regularity-learning and generalization (Dolson 1989). This fundamental shift allows the development of new algorithmic composition methods that rely on learning the structure of existing musical examples and generalizing from these learned structures to compose new pieces. These methods contrast greatly with the majority of older schemes that simply follow a previously assembled set of compositional rules, resulting in brittle systems typically unable to appropriately handle unexpected musical situations.

Música e semântica incorporada

“…Enfim, seguindo em outra direção, proponho aqui considerar os produtos gerados na superfície linguística apenas como sinalizações dos dispositivos e estratégias cognitivas operadas pelos usuários de música, sejam eles autores de textos musicais-escritos ou sonoros-ou realizadores desses textos-performers e ouvintes em geral-, no nível pré-conceitual do nosso entendimento, o nível das ações "indizíveis" referido por Kramer. O empreendimento teórico musical sobre o qual desenvolvo este estudo surgiu no mesmo período de emergência da nova musicologia; trata-se de uma musicologia cognitiva enquanto ramo da musicologia sistemática que abriu, nos anos 1980, uma nova frente de produção teórica em Música. Contudo é necessário aqui distinguir a corrente teórica que subjaz o desenvolvimento teórico a que refiro, da densa literatura musical em estudos cognitivos, surgida nos anos 1980, seja esta no âm-bito da psicoacústica, da neurofisiologia, de teorias da percepção, teorias gerativas ou teorias conexionistas e computacionais das chamadas ciências cognitivas, que pode ser reunida num contexto teórico admitido como um "estruturalismo cognitivo" (McAdams y Bregman, 1979;Krumhansl y Shepard, 1979;Shepard, 1982;Deutsch, 1982;Krumhansl y Kessler, 1982;Lerdahl y Jackendoff, 1983;Terhardt, 1984;McAdams, 1984;Sloboda, 1985;Todd, 1985Todd, , 1991Krumhansl, 1985;Rosner y Meyer, 1986;Clarke, 1987;Bharucha, 1987;Deliège, 1987;Lerdahl, 1988;Zatorre, 1988;Leman, 1989Leman, , 1990Krumhansl, 1990;Bregman, 1990). O presente trabalho, ao invés, está apoiado no paradigma não representacionista das ciências cognitivas contemporâneas, uma ciência cognitiva "incorporada": a corrente enacionista.…”

Deep Creations: Intellectual Property and the Automata

“…As one of the most promising early machine-learning techniques, neural network architectures were, thus, soon assimilated by computer artists. Using data to learn musical structure, neural networks were first used as a mechanism to elaborate a composition by learning note-to-note transitions (Todd, 1989;Mozer, 1994). Other attempts learned to improvise beats and melodies as part of a jazz band (Nishijima and Watanabe, 1993), or to select synthetic images based on an evaluation of their esthetic value (Machado and Cardoso, 1997), allowing the learning set as much as the neural network training protocols became constitutive elements of the artistic product.…”

Reduced memory representations for music

“…Other connectionist researchers in music have assumed that knowledge about constituent structure should be extracted from a training set using general learning algorithms, and then explicitly or implicitly dealt with by a network memory coding mechanism (Mozer, 1992;Todd, 1991). Our assumption is based on the theory that the constituent structures most relevant to music cognition are closely related to rhythmic structure (Lerdahl & Jackendoff, 1983).…”

“…Other connectionist researchers have explored issues of sequential structure in music with discrete-time recurrent network architectures (Bharucha & Todd, 1989;Mozer, 1991;Todd, 1991). Recurrent networks are often trained to predict the next event of a sequence, given a memory of past events in the same sequence (Mozer, 1993), rather than being explicitly trained to develop a representation for an entire sequence.…”