|
Hi, Most existing theoretical results in active learning assume points distributed uniformly over the unit sphere and homogeneous (passing through the origin) linear separators in R^d. While these assumptions makes analysis simpler by using the geometry of the unit sphere, they are also severely limiting in that they rarely resemble real-life active learning scenarios. The point that I was mulling over is: do people know what other reasons these assumptions might have other than simplifying the theoretical analysis. Does anyone know of existing worst case (i.e., distribution free) analysis in these theoretical work being done in this area? |
I don't think the distinction between homogeneous and non-homogeneous (i.e., with an explicit bias term) matters all that much in high dimensions (specially as you can add bias features if you feel like it, although it rarely improves performance significatively).