Be more theoretically grounded in parameter settings.

[bcipolli]

• Should we choose a different distribution than the "donut normal" (normal distribution, but suppressing the center pixel)?
• Should we add noise? If so, how much?
• Should we train strongly (to use as much neighbor information as possible) or weakly (because when you train strongly, asymmetries seem to disappear)?
• Should we use logpolar? If so, when do things start to break?
• What's the distribution differences for LH/RH at the different image sizes?