Self-Organization Driven by Pattern Generators: Expressing Genetics Through Learning

James A. Bednar and Risto Miikkulainen

How can a system as complex as the human brain be specified and constructed? Recent discoveries of widespread spontaneous activity in the nervous system suggest a simple yet powerful explanation: genetic information may be expressed as internally generated training patterns for an environment-driven self-organizing system.

We are exploring this idea in a computational model of cortical development, using the visual system as a test case. The simulations show how capabilities such as orientation and face processing can be specified in terms of simple training patterns, and how the prespecification interacts with learning from the environment.

Most of our results on this topic are presented in my 2002 dissertation. Simulations of neonatal face processing appear in our 2003 Neural Computation paper. Simulations of postnatal face processing were presented at the 2002 CogSci conference. Earlier and preliminary work addressed prenatal orientation perception (CNS*97) and face perception (AAAI-00). Detailed evolutionary and functional arguments supporting the pattern generation approach are presented in chapter three of my Master's thesis. Specific arguments about the role of internal pattern generation in sleep and memory were published as a commentary on the 2000 BBS Special Issue on Sleep and Dreaming.

Upcoming work will examine how the idea of pattern generation can be applied to the design of complex systems in general. I hope to demonstrate that the combination of a general self-organizing learning system with an internal pattern generator constitutes an efficient way to specify, develop, and maintain a complex adaptive system. Unlike current computing systems, such a system can seamlessly integrate information from internal and environmental sources. The goal is to use computational simulations as a tool to understand brain development, and at the same time to use brain development as a tool to understand how a complex information-processing system can be designed and implemented.