The repository contains code that was used to generate data and figures for Waniek, Transition scale-spaces: A computational theory for the discretized entorhinal cortex, 2019.
Please see below how to create raw data and figures that are presented in the
Experiments section of the paper. See the code in folder paper_figures for
code that was used to generate some of the other figures in the paper, such as
the entropy of binary search with overlapping receptive fields. If you want to
create particles on a hemisphere similar to the ones illustrated in the Appendix
of the paper, then have a look at README_hemisphere.md for some working
parameters. You can use those in combination with the hemisphere_...
scripts.
If you have any questions regarding the code, please don't hesitate from contacting me.
If you use this code, or base new experiments on it, please make sure to cite the corresponding article:
@article{waniek2019tss,
author = {Waniek, Nicolai},
title = {Transition scale-spaces: A computational theory for the discretized entorhinal cortex},
publisher = {MIT Press},
doi = {10.1162/neco\_a\_01255},
note ={PMID: 31835003},
journal = {Neural Computation},
volume = {32},
number = {2}
year = {2020},
pages = {330-394},
url = {https://doi.org/10.1162/neco_a_01255 },
eprint = {https://doi.org/10.1162/neco_a_01255}
}
All code in this repository is under MIT License. See LICENSE file for details.
Brief descriptions of the algorithms can be found in the paper. Of course you can also just have a look at the file. Most of the code is support-code to record activity of nodes in the graph. Have a look at the comments in the file to discern what is algorithm and what is support-code.
./demo_01_linear_track.py
./demo_02_wavepropagation.py --N 250 --startX 0.2 --startY 0.2 --targetX 0.8 --targetY 0.8 --W 1.0 --H 1.0 --period 0.2 --M 300 --pointgen rand --save-figures
./demo_03_scalespace_refinement_descending.py --N 250 --startX 0.4 --startY 0.1 --targetX 1.8 --targetY 4.8 --W 2.0 --H 5.0 --M 200 --max-i 2 --nscales 5 --save-figures
./demo_04_scalespace_refinement_ascending.py --N 250 --startX 0.4 --startY 0.1 --targetX 1.8 --targetY 4.8 --W 2.0 --H 5.0 --M 200 --max-i 2 --nscales 5 --save-figures
./demo_05_watermaze.py trajectory_data/005.hdf5 --M 100 --save-figures
Code to generate figures from paper can be found in subfolder
paper_figures
To recreate the example of geodesic computation given in the appendix of the paper, you first need to generate data for the field distribution on the hemisphere, then compute samples from start to target, and finally visualize. Or, after checking out this repository, directly jump to the third and final step of visualization. In detail, the steps are as follows.
First, run the file hemisphere_walker.py, which will spawn a virtual agent
on a hemisphere and keep it walking according to some statistics similar to a
rodent. While walking around, grid fields will be placed and self-organized
according to particle dynamics that were described in [1]. Specifically, the
closest particle to the agent will be moved closed to the agent (but slower than
the agent's movement). All other particles in the vicinity of the agent will be
repelled. After finishing the execution for some time (see line 58 of the file),
the particle locations will be stored to a .json file.
Second, run the file hemisphere_tss.py. Despite the name, this file will
not create the TSS but only perform trajectory sampling using Dijkstra and
backtracking. After finishing, this file will write its results to a file. To
change the file name, adapt line 61 of the file.
Third, and finally, visualize everything using hemisphere_visualizer.py.
If you changed the filenames in the other scripts, make sure to also adapt this
file at lines 47 -- 48.
[1] Waniek Hexagonal Grid Fields Optimally Encode Transitions in Spatiotemporal Sequences, 2018, Neural Comput. 2018 Oct;30(10):2691-2725. doi: 10.1162/neco_a_01122, https://www.mitpressjournals.org/doi/abs/10.1162/neco_a_01122