Analysing and bounding numerical error in spiking neural network simulations

This study explores how numerical error occurs in simulations of spiking neural network models, and also how this error propagates through the simulation, changing its observed behaviour. The issue of non-reproducibility in parallel spiking neural network simulations is illustrated, and a method to bound all possible trajectories is discussed. The base method used in this study is known as mixed interval and affine arithmetic (mixed IA/AA), but some extra modifications are made to improve the tightness of the error bounds. I introduce Arpra, my new software, which is an arbitrary precision range analysis library, based on the GNU MPFR library. It improves on other implementations by enabling computations in custom floating-point precisions, and reduces the overhead rounding error of mixed IA/AA by computing in extended precision internally. It also implements a new error trimming technique, which reduces the error term whilst preserving correct boundaries. Arpra also implements deviation term condensing functions, which can reduce the number of floating-point operations per function significantly. Arpra is tested by simulating the Hénon map dynamical system, and found to produce tighter ranges than those of INTLAB, an alternative mixed IA/AA implementation. Arpra is used to bound the trajectories of fan-in spiking neural network simulations. Despite performing better than interval arithmetic, the mixed IA/AA method used by Arpra is shown to be inadequate for bounding the simulation trajectories, due to the highly nonlinear nature of spiking neural networks. A stability analysis of the neural network model is performed, and it is found that error boundaries are moderately tight in non-spiking regions of state space, where linear dynamics dominate, but error boundaries explode in spiking regions of state space, where nonlinear dynamics dominate