On integral probability metrics, φ-divergences and binary classification (2009, background): “= µ(X|Y = +1), Q(X) := µ(X|Y = −1), ε:= µ(M,Y = +1). Here, P and Q represent the class-conditional distributions and εis the prior distribution of class +1. By appropriately choosing L, Nguyen et al. [33] have shown an equivalence between φ-divergences (betweenPand Q) and RL F⋆ . In particular, they”
Improving MMD-GAN Training with Repulsive Loss Function (2018, background): “• The Hinge loss (Tran et al. (2017)): LD = EPX [ReLU(1−D(x))] +EPZ [ReLU(1 +D(G(z)))], LG = EPZ [−D(G(z))], which is notably known for usage in support vector machines and is related to the total variation (TV) distance (Nguyen et al. (2009)).”
Composite Multiclass Losses (2011, background): “(Nguyen et al., 2009) Explicit 1:1 correspondence for binary and multiclass case (Reid and Williamson, 2011; Garcı́aGarcı́a and Williamson, 2012).”
Generative Adversarial Imitation Learning (2016, background): “Abbeel and Ng [1] and Syed et al. [29] use, respectively, C linear leads to feature expectation matching [1], which minimizes (cid:96) 2 distance between expected feature vectors: Meanwhile, C convex leads to MWAL [28] and LPAL [29], which minimize worst-case excess cost among the individual basis…”
On the Estimation of α-Divergences (2011, methodology): “Recently, Nguyen et al. (2009, 2010) developed methods for estimating f -divergences using their variational characterization properties.”