Learning Inverse Mapping by AutoEncoder Based Generative Adversarial Nets

Abstract

The inverse mapping of GANs'(Generative Adversarial Nets) generator has a great potential value. Hence, some works have been developed to construct the inverse function of generator by directly learning or adversarial learning. While the results are encouraging, the problem is highly challenging and the existing ways of training inverse models of GANs have many disadvantages, such as hard to train or poor performance. Due to these reasons, we propose a new approach based on using inverse generator (IG) model as encoder and pre-trained generator (G) as decoder of an AutoEncoder network to train the IG model. In the proposed model, the difference between the input and output, which are both the generated image of pre-trained GAN's generator, of Au-toEncoder is directly minimized. The optimizing method can overcome the difficulty in training and inverse model of an non one-to-one function. We also applied the inverse model of GANs' generators to image searching and translation. The experimental results prove that the proposed approach works better than the traditional approaches in image searching.

Inverting the Generator of a Generative Adversarial Network

“…In contrast to Luo et al [11], we demonstrate our inversion approach on data samples drawn from test sets of real data samples. To make inversion more challenging, in the case of the Omniglot dataset, we invert image samples that come from a different distribution to the training data.…”

“…On the other hand, Luo et al 3 [11], train an encoding network after the GAN has been trained, which means that their approach may be applied to pre-trained models. One concern about the approach of Luo et al [11], is that it may not be an accurate reflection of what the GAN has learned, since the learned decoder may over-fit to the examples it is trained on. For this reason, the approach of Luo et al [11] may not be suitable for inverting image samples that come from a different distribution to the training data.…”

“…One concern about the approach of Luo et al [11], is that it may not be an accurate reflection of what the GAN has learned, since the learned decoder may over-fit to the examples it is trained on. For this reason, the approach of Luo et al [11] may not be suitable for inverting image samples that come from a different distribution to the training data. Evidence of over-fitting may be found in Luo et al [11]- Figure 2, where "original" image samples being inverted are synthesised samples, rather than samples from a test set data; in other words Luo et al [11] show results ( Figure 2) for inverting synthesised samples, rather than real image samples from a test set.…”

Pioneer Networks: Progressively Growing Generative Autoencoder

“…We compare reconstructions against inference-capable models: AGE [31] and ALI [6]. We also train ProgGAN for reconstruction as follows (compare to e.g., [24,15,17,4]). We train the network normally until convergence, and then use the latent vector of the discriminator also as the latent input for the generator (properly normalized).…”

Adversarial Training in Affective Computing and Sentiment Analysis: Recent Advances and Perspectives [Review Article]

“…Network-type-based: in addition, several GAN variants have been named after the network topology used in the GAN configuration, such as the DCGAN based on deep convolutional neural networks [19], the AEGAN based on autoencoders [71], the C-RNN-GAN based on continuous recurrent neural networks [72], the AttnGAN based on attention mechanisms [73], and the CapsuleGAN based on capsule networks [74].…”