Gradients should stay on path: better estimators of the reverse- and forward KL divergence for normalizing flows

Lorenz Vaitl, Kim A. Nicoli, Shinichi Nakajima, Pan Kessel