Figure 1. Previous works used class information p (y ) to align
s
the learned source distribution p (y , u|x) with the balanced target
s
distribution p (y , u|x), without considering the space u and class
t
y relationship i.e. p (y , u). FRACAL captures p (y , u), using
s s
the fractal dimension, and embeds fractal margins during infer-
ence, aligning the learned distribution p (y , u|x) with the target
s
p (y , u|x) better than previous works.
t