SORN

This codes simulates the self-organizing recurrent neural network (SORN).

Install

Just git clone https://github.com/chrhartm/SORN.

Code organization

The code is set up in a somewhat modular manner:

• the common folder contains code common to all experiments such as the sorn, default parameters, ...
• the utils folder contains code for backing up files, submitting jobs to clusters, ...
• the other folders (examples, chartmann) contain the experiments implemented so far.

In general, this is scientific code so it is neither pretty nor well documented. The files in the common folder are somewhat documented and especially the default parameter file (defaults.py) might be useful to get started.

Getting started

To run experiments, navigate to the common folder and run python test_single.py your_parameter_file.

For example, run python test_single.py examples.zheng2013.param_noinput to replicate some figures of Zheng2013.

Designing your own experiments

If you are planning to also work with SORN and use this code, feel free to fork and set up your experiments in a folder like chartmann. Once you are ready to publish, issue a pull request to publish your code in this repository. If you need to modify files in the common folder, try to keep backwards compatibility to the other experiments (e.g. by introducing default parameters).

Exploring parameters and averaging independent simulations on a cluster

Apart from running single experiments, you can also explore parameters with MPI (using mpi4py) and average over multiple instantiations of the same network. Both is done with the file test_cluster.py. For this you can:

• adapt the submit-sleuth file in the utils folder to your needs and run from the common folder ../utils/submit-sleuth -n N_JOBS test_cluster.py your_parameter_file
• do this locally by running mpirun -n N_JOBS python test_cluster.py your_parameter_file

Please note that the code is written in such a way that N_JOBS has to match the number of different parameters to be explored (c.cluster.params in the parameter files).

Parameter files for Hartmann2015

The simulations will always produce more plots than there are in the paper.

c indicates cluster simulations. Most also return plots for single experiments.

• Fig.1: chartmann.spont.param_spont_randletters
• Fig.2: c chartmann.spont.param_sequence (switch the test_words in the param file to get both a and b)
• Fig.3: chartmann.spont.param_spont, c chartmann.spont.param_fiser
• Fig.4: c chartmann.spont.param_spont
• Fig.5: c chartmann.spont.param_hesselmann
• Fig.7: c chartmann.spont.param_hesselmann (set eta_stdp to 0 for b and d)
• Fig.8: chartmann.spont.param_spont

Parameter files for Hartmann2016

• Fig.1: see notebook chartmann/alignment/kesten_notebbok/kesten_process_alignment.ipynb
• Fig.2: see Fig.1
• Fig.3: see code at http://fias.uni-frankfurt.de/~miner/
• Fig.4: chartmann.alignment.param_alignment_potbias
• Fig.5: chartmann.alignment.param_alignment_depbias
• Fig.6: see Fig.4 and Fig.5

UML

The following simplified UML diagram sketches the simulation setup:

When running test_single.py with a parameter file, the script will instantiate the Experiment specified in the parameters. The Experiment then creates Sources that specify the stimulation paradigm and Stats that record data at every time step. These Stats are then collected in a StatsCollection that will be an attribute of Sorn. Finally, the Experiment instantiates a Sorn with the parameters from the parameter file. The Sorn then instantiates a number of SynapticMatrices that can be either sparse or dense and will perform stdp or other plasticity mechanisms as specified in the parameters. After the initialization, the Experiment is run by the test_single.py script and the results are analyzed and plottet according to the plot_single method of the Experiment. If one runs multiple simulations with test_cluster.py, the Experiment is reset by the script before run is called again.