This repository contains the code to reproduce the results presented in the paper "Optimizing semantic pointer representations for symbol-like processing in spiking neural networks".

• Python 2.7 (Python 3 might work with minor changes, but has not been tested.)

The following Python packages that should be pip installable:

• NumPy
• SciPy
• pandas
• nengo_spinnaker
• Jupyter
• matplotlib
• Seaborn

The following Python packages that are not yet pip installable in the required version:

• Nengo at commit 17f5e81bd432aaa272b3d69ec2ba179f6f0bf6ef (or newer)
• psyrun at commit 852da71434873a060d88fbc00a7bc5df6bb9e6b6

To install these to python packages run the following commands:

git clone https://github.com/nengo/nengo.git
git clone https://github.com/jgosmann/psyrun.git
cd nengo
git checkout 17f5e81bd432aaa272b3d69ec2ba179f6f0bf6ef
python setup.py install --user
cd ../psyrun
git checkout 852da71434873a060d88fbc00a7bc5df6bb9e6b6
python setup.py install --user

The code in this repository will likely work with the upcoming Nengo 2.1 release as well. Thus, following the release installing this specific Nengo version should not be required anymore and a pip install nengo should suffice. The psyrun API might still change at this point and I recommend using the version given here.

To generate all the data change to the repositories root directory and run psy-doit twice. This will require a configured SpiNNaker board connected to the computer. To only generate specific data sets run psy-doit <task_name>. To get a list of all task names run psy-doit list.

The tasks ending in _less_neurons require data from the task with the same name without the suffix. As long as that data does not exist, the _less_neurons tasks will do nothing. (That is why psy-doit has to be run twice to generate all the data.)

To generate the plots from the data open Data analysis.ipynb with Jupyter notebook and run the cells.