Generative Adversarial Networks

Generative Adversarial Networks (GANs): a fun new framework for estimating generative models, introduced by Ian Goodfellow et al.
We are using a 2-layer network from scalar to scalar (with 30 hidden units and tanh nonlinearities) for modeling both generator and discriminator network.
We are trying to reproduce this figure from the paper:
Except as a live demo. As we'll see below things are not quite as nice in practice:

color code: true data distribution, generator network from uniform noise in U[0,1], discriminator network

Reset Toggle 1D/2D Increase learning rate Decrease learning rate