Slide 6
Slide 6 text
Wasserstein natural proximal
The update scheme follows:
✓k+1 = arg min
✓2⇥
F(⇢✓) +
1
2h
dW (✓, ✓k)2.
where ✓ is the parameters of the generator, F(⇢✓) is the loss function,
and dW
is the Wasserstein metric.
In practice, we approximate the Wasserstein metric to obtain the
following update:
✓k+1 = arg min
✓2⇥
F(⇢✓) +
1
B
B
X
i
1
2h
kg✓(zi) g✓k
(zi)k2,
where g✓
is the generator, B is the batch size, and zi
⇠ p(z) are inputs
to the generator.
25