Source code for "The competing benefits of noise and heterogeneity in neural coding" by Eric Hunsberger, Matthew Scott, and Chris Eliasmith. The manuscript included in this repository is a preprint of the final article to be published in Neural Computation (see http://www.mitpressjournals.org/doi/abs/10.1162/NECO_a_00621).
All the code in this repository has been run on a Linux machine. While ideally it should work without modification on other machines, if you are having trouble getting the code to run, or have any other questions relating to the paper or the code, please contact Eric Hunsberger at ehunsber@uwaterloo.ca.
The core requirement for this repository is Python 2.7. Python requirements
can be found in requirements.txt, and can be installed using pip with
pip install requirements.txt
However, the two first requirements are Numpy and Scipy, which must be installed specially on some machines. For instructions specific to your machine, please visit http://www.scipy.org/install.html.
Building the paper requires LaTeX, specifically the existence of
a pdflatex command at the command-line.
All the scripts in this project, namely those for running the simulations,
creating the figures, and putting together the paper, have been set up to
run easily with doit. doit is a make-like utility
for managing tasks, and running them only as necessary. The full list of tasks
can be seen by typing
doit list
into the main console. Each command can then be run by typing doit and the
command name. For example, to build the paper, type
doit paper
The simulation tasks have been set not to run if the target results file
exists at all. To rerun the simulations, manually remove all the results files
(ending in .npz) from the results directory. Then call
doit
to rerun all out-of-date tasks (which should be everything). To rerun specific tasks, such as only the simulations for the information plots, you can use a wildcard character like
doit sim_info*
to run sim_info_lif and sim_info_fhn. Note that some of the simulations may
take a while to run (on the order of one day). To run the simulations faster,
you can ask Theano to run them on your GPU by passing the argument gpu=true
to doit.
doit gpu=true
This requires Theano to work with your GPU (see the first paragraphs of http://deeplearning.net/software/theano/tutorial/using_gpu.html).
The scripts folder contains all the source code for conducting the numerical
experiments, as well as code for generating the plots.