3 ms·
As the origin is special, instead of training in a linear space, what would training in an affine space do?
by 082349872349872 2y ago
As the origin is special, instead of training in a linear space, what would training in an affine space do?
- whatshisface 2y agoI think they are training in an affine space, but I see what you're saying. The initialization of the bias must be breaking the symmetry in a way that makes the origin special. Of course to some degree that's unavoidable since we have to initialize on distributions with compact support.