search

ioam / topographica

A general-purpose neural simulator focusing on topographic maps.
topographica.org
BSD 3-Clause "New" or "Revised" License
53 stars 32 forks source link

what is the value of parameter" area" in the gcal script to produce the paper? #626

Closed windform2015 closed 9 years ago

windform2015 commented 9 years ago

I am confused by the value of "area" in the GCAL model. In the GCAL model script, it is 1.0 , but in the notebook tutorial "stevens_jn13.ipynb" which is in the directory topographica/models/stevens.jn13/, is 1.5. I would like to know, which is the value used to reproduce the stevens. et al(2013) paper ? Thanks.

jbednar commented 9 years ago

In most of our work, the value is 1.0, and I believe that all of the plots show a 1.0x1.0 area in the JN13 paper. However, GCAL (like any laterally connected model with non-periodic boundary conditions) will have border effects near the edges of the cortical sheet, which can affect the density of pinwheels in those areas. To avoid such issues, for the JN13 paper we increased the area to 1.5x1.5 for simulation, then cut off all but the central 1.0x1.0 region.

So, if you want to reproduce the paper exactly, you need to do 1.5x1.5, but we usually just use 1.0x1.0 ourselves and accept that there will be some effects of the border on the local details of the map shape.

To summarize: The values for exact reproduction of the paper are in models/stevens.jn13, but the value in examples/gcal.ty (or examples/gcal_oo_or.ty, a pared-down version that only works for orientation) is what we would normally recommend as a starting point for a new project. Another way to think of it is that examples/*.ty changes over time as we gradually improve and change things, while models/stevens.jn13 will always stay exactly as it is.

windform2015 commented 9 years ago

Thanks for your quick and nice answer. I am also wandering ,changing area=1.5 to area=1.0, pinwheel densities generated will be changed a little or not ?

jbednar commented 9 years ago

I'm not sure how significant the changes will be; any individual run is variable and we'd have to run the same number of runs as in jn13 if we wanted compare rigorously, and we haven't done that recently (though maybe Jean-Luc still remembers). In any case the network should work the same; the number of pinweels is important just for making detailed comparisons with experimental data.

jlstevens commented 9 years ago

Using area 1.0 definitely distorts the pinwheel density due to easily visible edge effects. To get a good estimate of the pinwheel density for the models, we had to lots of runs (and then average the pinwheel density across them) as well as simulate area 1.5 before cropping to unit area. So in short, I can confirm Jim's response! :-)

As the the magnitude of the effect, I'm afraid I can't remember exactly so whatever I say here may be wrong. I am fairly sure the pinwheel density went up due to edge effects so at a very rough guess, I would say you might expect a pinwheel density of around 3.5 or so if you simulate GCAL with area 1.0 (assuming you average across enough simulations to get a reliable estimate).

windform2015 commented 9 years ago

Interesting! I would like to see how the pinwheel density changes and test the expectation. Thanks !