- Abstract: Implicit generative models are difficult to train as no explicit probability density functions are defined. Generative adversarial nets (GANs) propose a minimax framework to train such models, which suffer from mode collapse in practice due to the nature of the JS-divergence. In contrast, we propose a learning by teaching (LBT) framework to learn implicit models, which intrinsically avoid the mode collapse problem because of using the KL-divergence. In LBT, an auxiliary explicit model is introduced to learn the distribution defined by the implicit model while the later one's goal is to teach the explicit model to match the data distribution. LBT is formulated as a bilevel optimization problem, whose optimum implies that we obtain the maximum likelihood estimation of the implicit model. We adopt an unrolling approach to solve the challenging learning problem. Experimental results demonstrate the effectiveness of our method.