[Lisa_teatalk] Talk: David Duvenaud on Gradient-based hyperparameter optimization through reversible learning