Lissom color fixes for Blue-RedGreen LGN sheets

[Tobias-Fischer]

This pull request consists of two small bugfixes related to the Blue-RedGreen LGN sheets in lissom.ty. 1: So far a fixed value of 4.7 was used as strength from the retina sheets to the Blue-RedGreen LGN sheets. However, this value is only suitable for natural image datasets, and only if speed=0. Instead, strength_scale should be used which depends on

• Whether it is a natural image dataset or Gaussian input
• How fast patterns are moving

2: The red and green retina sheets were connected to the Blue-RedGreen LGN sheets using center weights. Instead, they should be connected using surround weights. Only the blue retina sheets should use center weights (as they do correctly). This is a bug in lissom_oo_or_cr.ty as well as lissom.ty

[jbednar]

The strength_scale change sounds useful, thanks!

However, using the center weights for the B/Y surround is not a bug; it's implementing spatially coextensive B/Y weights, which is what some studies argue are the most common in animals. De Paula (2007) used coextensive weights, and the lissom files with color are meant to replicate his results, so they should do the same. That said, other authors argue for spatially opponent B/Y cells, and so it would be useful to easily switch between those two options. If you want to support that, you can add a parameter to allow switching between center-sized and surround-sized Y weights, which should be straightforward (though it may require other parameter changes elsewhere to compensate).

[Tobias-Fischer]

Right, I see. The code would just look much more regular if there were spatially opponent cells ;).

For now, I deleted the commits related to coextensive/spatially opponent cells. I will add a parameter for that in the model I will be working on soon.

Therefore this pull request only consists of the very small change related to the strength_scale.

[ceball]

On Thu, Jun 19, 2014, at 02:15 PM, Tobias Fischer wrote: [...]

2: The red and green retina sheets were connected to the Blue-RedGreen LGN sheets using center weights. Instead, they should be connected using surround weights. Only the blue retina sheets should use center weights (as they do correctly).

This code's supposed to be modeling the B/RG pathway as spatially co-extensive (e.g. http://www.annualreviews.org/doi/abs/10.1146/annurev.neuro.23.1.743). You could argue it's not this way in reality, but probably we should do that outside of a pull request ;)

Chris