In the quest of efficient hardware implementations of dynamic neural fields : an experimental study on the influence of the kernel shape

On this page, you will find some additional figures not included in the submitted paper as well as the simulation scripts and optimized parameters

Simulation scripts : HowTo

The simulations are written in Python; The optimisation is performed with the Popot library and its Python wrapper. Your also need numpy and matplotlib. The scripts used for the simulations are ijcnn-2016.tar.gz

Then, the optimization is performed with a single script “optimize.py” which takes as input the scenario name (selection or wm), the trial number and the kernel name (dog, doe, dol, step).

In order to reproduce the illustrations of the paper, you simply need to call the script plot_hist_results.py

Reading of the result files

In the next section, we provide the result files. For each kernel and scenario, we provide a fitness.data and params.data file. The fitness.data file is a matrix of nruns x nsteps . Each row contains the values of the cost function for one trial. The params.data file is a matrix of nruns x (2 + nparams). Each row is for one trial. The first column is the index of the last step of the optimization, the second column is the minimal fitness obtained at the end of the optimization and then come the 6 parameters [dt/tau, h, Ae, ks, ka, si].

Competition scenario

Below are the parameter and fitness files for the different kernels :

Below are the histograms from which some elements are given in the paper.

Selection dog Selection dog all the weights Selection doe Selection doe all the weights Selection dol Selection dol all the weights Selection step Selection dol all the weights

Working memory scenario

Below are the parameter and fitness files for the different kernels :

Below are the histograms from which some elements are given in the paper.

Wm dog Wm dog all the weights Wm doe Wm doe all the weights Wm dol Wm dol all the weights Wm step Wm step all the weights