Training Generative Adversarial Networks with Binary Neurons by End-to-end Backpropagation
This work addresses a specific challenge in modeling discrete distributions with GANs for researchers in generative models, though it is incremental as it builds on existing GAN frameworks with binary neurons.
The authors tackled the problem of training generative adversarial networks (GANs) with binary neurons by proposing BinaryGAN, which uses sigmoid-adjusted straight-through estimators for gradient estimation and end-to-end backpropagation, enabling direct generation of binary-valued outputs like binarized MNIST digits, with results showing it is feasible but preliminary.
We propose the BinaryGAN, a novel generative adversarial network (GAN) that uses binary neurons at the output layer of the generator. We employ the sigmoid-adjusted straight-through estimators to estimate the gradients for the binary neurons and train the whole network by end-to-end backpropogation. The proposed model is able to directly generate binary-valued predictions at test time. We implement such a model to generate binarized MNIST digits and experimentally compare the performance for different types of binary neurons, GAN objectives and network architectures. Although the results are still preliminary, we show that it is possible to train a GAN that has binary neurons and that the use of gradient estimators can be a promising direction for modeling discrete distributions with GANs. For reproducibility, the source code is available at https://github.com/salu133445/binarygan .