22:15? Reverse KL 35:40 Instead of optimizing over all possible functions T, optimize over an arbitrary set of neural network architectures. Similar flavor to importance sampling in VAEs. 36:45? Essentially the same as minimax training objective of GANs, p and q are expectations, T is the discriminator and optimization does not depend on likelihoods (p and q) 40:00? Supporting hyperplanes, convex hull, tangent? 41:35 KL divergences vs f divergences, doesn't depend on likelihoods, doesn't measure compression of data, flexible loss functions 48:40 Summary slide on divergences and training objectives 51:00? Example 1:04:00? Lipschitz constant 1:08:00? Earth mover distance, how is this different to the minimax training objective of GAN 1:09:30 Unlike f divergence, Earth mover distance doesn't give bounds 1:22:20 BiGAN, encoder is similar to VAE except deterministic, not trained by minimizing KL divergences like in the ELBO, trained by minimizing two sample test optimized by the discriminator