Abstract
We propose the BinaryGAN, a novel generative adversarial network (GAN) thatuses binary neurons at the output layer of the generator. We employ thesigmoid-adjusted straight-through estimators to estimate the gradients for thebinary neurons and train the whole network by end-to-end backpropogation. Theproposed model is able to directly generate binary-valued predictions at testtime. We implement such a model to generate binarized MNIST digits andexperimentally compare the performance for different types of binary neurons,GAN objectives and network architectures. Although the results are stillpreliminary, we show that it is possible to train a GAN that has binary neuronsand that the use of gradient estimators can be a promising direction formodeling discrete distributions with GANs. For reproducibility, the source codeis available at https://github.com/salu133445/binarygan .