Self-organization in connectionist models: Associative memory, dissipative structures, and thermodynamic law

Abstract

In the development of connectionist models it is popular to rely on the concept of self-organization and to employ analogies from thermodynamics. Here we review some aspects of self-organization and thermodynamic law. We conclude that they do, indeed, have much to offer the modeling of human action. However, we further conclude that connectionists have failed to exploit the full potential of the properties inherent in a thermodynamic model of self-organization. Their use of self-organization lacks the imperatives of physical theorists or biologists who have written extensively on the topic. The use of computational temperature as an ordering principle for associative memory is analyzed. The more common approach in connectionism, to seek order through cooling, has less potential to explain the emergence of new behavioral properties than an approach that seeks order through heating. Thermodynamics as a source of analogies is also seen as limiting and we question the value of analogy as a basis for a scientific endeavor. An appeal to the constructive role of the Second Law as it operates on open systems can account for important features of organized activity. In this view the Second Law does not offer analogies; it is a law that describes the causal basis of human action.