Distributed Computation Using Algebraic Elements

Abstract

Information processing in the nervous system is subject to a number of constraints. Because the number of neurons far exceeds the number of genes expressed in the brain, a computing architecture based on multiple copies of relatively simple modules seems necessary. Since most neuronal processes are short, the connections between modules have to be primarily local. Processing in neurons is slow compared to the complexity of the tasks that have to be carried out, so distributed rather than sequential computation is required. There does not seem to be a master “clock” regulating transfer of data from module to module; nor is there any rigorous distinction between signal transmission and processing. Signals are transformed even while propagating through the brain. One might call such an architecture “diffusional” rather than “parallel”. Diffusional architecture creates still another problem for the nervous system, however, because the probabilistic nature of propagation in dendrites and small-caliber axons makes repeated signal copying and transformation subject to a high error rate. Error-correction of some kind must be built into the system.