Björn Kindler

Statistical Analysis of Network Dynamics

In spiking neural networks, both information processing and (short term) memory are encoded in the network dynamics, i.e. in the spike times of the different neurons. But due to noise, a neural network will never respond in exactly the same way to identical input. Furthermore, in biological systems even similar input information results in varying spike patterns sent to the cortex: E.g. the signal generated by the retina changes while viewing a single object.

Obviously, the occurrence or absence of single spikes does not affect the function of a large network in principle. In many cases neural information processing rather reveals a statistical nature. Many significant facets of a network's dynamics can be subsumed as firing activities of different populations of neurons and their causal and temporal dependencies on each other. Therefore, we aim for a statistical description of networks' behaviour. Commonly used measures are mean firing rates (for overall activity), pairwise correlations (whether neurons fire in synchrony) or the so-called Fano-Factor (whether the spikes occur clocked or at random). Depending on the particular task and connectivity of the network further measures need to be applied.

Similar to PyNN, which allows a platform-independent description of networks and experiments, the members of the FACETS research project developed a package of classes and functions - named NeuroTools - to process, analyse and plot the experiments' results. Since PyNN provides uniform data formats on any simulator back-end, NeuroTools is compatible with any platform's results.

Example of a small network on the "Spikey" chip firing synchronous (top) and asynchronous (bottom). Plots generated with NeuroTools.