Techniques for enhancing neuronal evolvability

Abstract

High dimensionality and interactional complexity, appropriately introduced, can enhance the evolvability of a pattern processing network. We describe a processor, referred to as the cytomatrix module, that can be used to investigate the requisite conditions for such enhancement. The processor is characterized by multiplicity of component types, graded interactions among components, separation of signal integration dynamics from the readout mechanisms that interpret these dynamics, and multiplicity of parameters open to evolution (including component connectivity). The adaptation procedure is mediated by a multiparameter variation–selection algorithm that acts on the various parameters in an alternating (i.e., phasic) manner. Experiments with both structured and unstructured learning tasks, as well as with difficult parity problems, demonstrate that opening more parameters to evolution increases the flexibility exhibited by the processor in response to evolutionary pressure, essentially by loosening the coupling between the local and global aspects of the response. The cytomatrix processor can be thought of as a highly abstracted representation of signal integration within single neurons; alternatively, it can be viewed as a collection of cells in a multicellular organization.