High Performance Computing Applications in Neurobiological Research; Muriel D. Ross, NASA Ames Research Center, Moffett Field, CA 94035; Kevin Montgomery, Sterling Software, Palo Alto, CA 94303; David G. Doshay, Sterling Software, Palo Alto, CA 94303; Thomas C. Chimento, Sterling Software, Palo Alto, CA 94303; Bruce R. Parnas, National Research Council Research Associate, Biocomputation Center, NASA Ames Research Center, Moffett Field, CA 94035

The human nervous system is a massively parallel processor of information. The vast numbers of neurons, synapses and circuits is daunting to those seeking to understand the neural basis of consciousness and intellect. Pervading obstacles are lack of knowledge of the detailed, three-dimensional (3-D) organization of even a simple neural system and the paucity of large scale, biologically relevant computer simulations. We use high performance graphics workstations and supercomputers to study the 3-D organization of gravity sensors as a prototypic architecture foreshadowing more complex systems. Scaled-down simulations run on a Silicon Graphics workstation and scaled-up, three-dimensional versions run on the Cray Y-MP and CM5 supercomputers.

To assist this research, we developed generalized computer-based methods for semiautomated, 3-D reconstruction of this tissue from transmission electron microscope (TEM) serial sections and for simulations of the reconstructed neurons and circuits. Sections are digitized directly from the TEM. Contours of objects are traced on the computer screen. Mosaicking images into sections, registration and visualization are automated. The same grids generated to connect contours for viewing objects provide tesselated surfaces for 1-D, 2-D and 3-D simulations of neuronal functioning. Finite element analysis of prism or segment volumes and color coding are used to track current spread after synapse activation. The biologically accurate simulation is reducible to a symbolic model that mimics the flow of information processing. Discharge patterns are displayed as spike trains. The symbolic model can be converted to an electronic circuit for potential implementation as a chip. The reconstructions can also be rendered in visual, sonic and tactile virtual media.

Using these methods, we demonstrated that gravity sensors are organized for parallel distributed processing of information. They have non-modular receptive fields that are organized into overlapping, dynamic cell assemblies. These provide a basis for functional degeneracy and graceful degradation. The sensors have two intrinsic microcircuits that are prototypic of more advanced systems. These microcircuits are highly channeled (type I cell to a nerve terminal called a calyx) and distributed modifying (type II cells and feedforward/feedback neural lOops). A circuit of extrinsic origin likely biases the intrinsic circuits. We use simulation methods to study the effects of intrinsic feedback-feedforward lOops and of extrinsically driven biases on discharge patterns. These and similar investigations into the functioning of huge assemblies of neurons require supercomputer capabilities and pave the way for studies of human brain functioning as a grand challenge in supercomputer applications.


(accessed February 3, 1998)

« Previous
Continue »