A large-scale whole visual system model integrated by PLATO and its implementation on high performance computer

 To understand the details of visual function, a large-scale system model that reflects anatomical and neurophysiological characteristics of visual system needs to be implemented. Though numerous computational models of various brain areas related with the visual function have been proposed, it is important to integrate those models into a large-scale visual system model in order to understand the details of visual function. Currently, we have been developing eye movement, optics, retina, and visual cortex models, and integrating into a large-scale whole visual system model on the PLATO (Platform for a coLlaborative brAin sysTem mOdeling) environment to understand visual processing underlying perception, illusion, learning and memory. In the eye movement model, the saccade related pathway, burst neurons, omnipause neurons, integrator neurons, and eye plant, were explicitly described to generate both saccade and microsaccade with drift eye movement. In the eye optics model, realistic ±10deg spectral images (1000 x 1000 pixels) on the retina were calculated from an external image covering the spectral ranges from 380 to 780 nm in 4 nm steps. The present retinal model consisted of approximately two million L, M, and S cone photoreceptors. In the retinal model, an alignment of those cones so-called “cone mosaic” was replicated. The visual cortex model was described as a filter model, and carried out motion detection.

 In the implementation on high performance computer (HPC), the simulation program of each model was transformed into efficient parallel version; each model was then connected with an interface system based on a common data format. Currently, we were using the netCDF as the common data format; this is expected to induce communication overhead due to file transaction in parallel simulation of HPC. The simulation of the models, and data exchange between the models was managed by the agent system of PLATO. We confirmed that the implemented models were worked as a large-scale visual system by visualizing the signal processing in each sub-system. We also made a performance and scalability analyses of the present model on the HPC in RIKEN consisting of 1024 CPU nodes (maximum 8192 processors). We obtained linear parallel performance, and transaction overhead of the common data format was small enough compared to the execution time of each model.

 

Acknowledgments

 

 This research was supported by the next-generation integrated simulation of living matter', part of the development and use of the next-generation supercomputer project of the ministry of education, culture, sports, science and technology. Part of the simulation was conducted on the RIKEN Integrated Combined Cluster.