www-ai.cs.tu-dortmund.de/LEHRE/SEMINARE/WS2122/TrustworthyAIMachineLearning/yurochkin2020a.pdf
considers differences
3
Published as a conference paper at ICLR 2020
3 2 1 0 1 2
1.5
1.0
0.5
0.0
0.5
1.0
1.5
(a) unfair classifier 3 2 1 0 1 2
1.5
1.0
0.5
0.0
0.5
1.0
1.5
(b) unfair map 3 2 1 0 1 2
1.5
1.0 [...] of the loss class. Proposition 3.1 (uniform convergence). Under Assumptions A1–A3, equation 3.1 satisfies
δn ≤ 48C(L)√
n +
48LD2
√ nε
+ LδcD
2
√ ε
+M( log 2
t
2n )
1 2 (3.2)
with probability at least 1− [...] convergence is equation 3.3 is close to its empirical counterpart
supP :W (P,Pn)≤ε EP [ `(Z, θ)
] − EPn
[ `(Z, θ)
] . (3.5)
In other words, the gap generalizes. This implies equation 3.5 is a certificate of …
