Template based black box optimization of Dynamic Neural Fields¶

On this page, you will find some scripts and illustrations used in the paper “A Template based black box optimization of Dynamic Neural Fields”.

In order to use the code for optimizing dynamic neural fields, you need two additional libraries :

• the neuralfield library which provides C++ templates for setting up a parametrized neural field with various types of transfer functions, weight functions ,etc.. The version 2.08 was used for the paper(neuralfield-2.08),
• the popot library which implements particle swarm optimization. The version 1.20 was used in the paper (popot-1.20),
• the matlab script for CMA-ES is included in the archive

At the bottom of the page, you will find links to the source packages of the version of the libraries used for the simulations.

Once both neuralfield and popot are installed, you can now compile the code for optimizing dynamic neural fields. The archive DnfPSO-cmake.tar.gz contains all the scripts for simulating the examples mentioned in the paper. Big files containing the matrix of the optimal parameters are given in the archive as well as below.

Some examples of optimized behaviors are provided below with a video for the 2D tracking scenario.

For the competition scenario, the responses of one field to the three conditions are shown below:

For the Wm scenario, you also see below the input and firing rates of the field as well as the firing rates and templates at t=80 and t=190:

For the tracking scenario, we show below a run with some optimal parameters.

Parameter files:

• Competition scenarion , PSO-selection, CMAES-selection
• Working memory scenario, PSO-wm, CMAES-wm
• Tracking scenario, PSO-tracking, CMAES-tracking

All the parameter files read:

TrialNumber Fitness a b u0 Ap Sp ka ks

with a,b,u0 the parameters of the transfer function, Ap and Sp the amplitude and variance of the excitatory gaussian of the lateral weights, the amplitude and variance of the inhibitory gaussian begin ka*Ap and ks * Sp.