Least-squares GAN (LSGAN)

As discussed in the previous section, the original GAN is difficult to train. The problem arises when the GAN optimizes its loss function; it's actually optimizing the Jensen-Shannon divergence, D_JS. It is difficult to optimize D_JS when there is little to no overlap between two distribution functions.

WGAN proposed to address the problem by using the EMD or Wasserstein 1 loss function which has a smooth differentiable function even when there is little or no overlap between the two distributions. However, WGAN is not concerned with the generated image quality. Apart from stability issues, there are still areas of improvement in terms of perceptive quality in the generated images of the original GAN. LSGAN theorizes that the twin problems can be solved simultaneously.

LSGAN proposes the least squares loss. Figure 5.2.1 demonstrates why the use of a sigmoid cross entropy loss in the GAN results in poorly generated data quality. Ideally, the fake samples distribution...

Auxiliary classifier GAN (ACGAN)

ACGAN is similar in principle to the Conditional GAN (CGAN) that we discussed in the previous chapter. We're going to compare both CGANand ACGAN. For both CGAN and ACGAN, the generator inputs are noise and its label. The output is a fake image belonging to the input class label. For CGAN, the inputs to the discriminator are an image (fake or real) and its label. The output is the probability that the image is real. For ACGAN, the input to the discriminator is an image, whilst the output is the probability that the image is real and its class label. Following figure highlights the difference between CGAN and ACGAN during generator training:

Figure 5.3.1: CGAN vs. ACGAN generator training. The main difference is the input and output of the discriminator.

Essentially, in CGAN we feed the network with side information (label). In ACGAN, we try to reconstruct the side information using an auxiliary class decoder network. ACGAN argued that forcing the network to...

Conclusion

In this chapter, we've presented various improvements in the original algorithm of GAN, first introduced in the previous chapter. WGAN proposed an algorithm to improve the stability of training by using the EMD or Wassertein 1 loss. LSGAN argued that the original cross-entropy function of GAN is prone to vanishing gradients, unlike least squares loss. LSGAN proposed an algorithm to achieve stable training and quality outputs. ACGAN convincingly improved the quality of the conditional generation of MNIST digits by requiring the discriminator to perform classification task on top of determining whether the input image is fake or real.

In the next chapter, we'll study how to control the attributes of generator outputs. Whilst CGAN and ACGAN are able to indicate the desired digits to produce; we have not analyzed GANs that can specify the attributes of outputs. For example, we may want to control the writing style of the MNIST digits such as roundness, tilt angle, and thickness. Therefore...

References

1. Ian Goodfellow and others. Generative Adversarial Nets. Advances in neural information processing systems, 2014(http://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf).

2. Martin Arjovsky, Soumith Chintala, and Léon Bottou, Wasserstein GAN. arXiv preprint, 2017(https://arxiv.org/pdf/1701.07875.pdf).

3. Xudong Mao and others. Least Squares Generative Adversarial Networks. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE 2017(http://openaccess.thecvf.com/content_ICCV_2017/papers/Mao_Least_Squares_Generative_ICCV_2017_paper.pdf).

4. Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional Image Synthesis with Auxiliary Classifier GANs. ICML, 2017(http://proceedings.mlr.press/v70/odena17a/odena17a.pdf).

Conclusion

In this chapter, we've presented various improvements in the original algorithm of GAN, first introduced in the previous chapter. WGAN proposed an algorithm to improve the stability of training by using the EMD or Wassertein 1 loss. LSGAN argued that the original cross-entropy function of GAN is prone to vanishing gradients, unlike least squares loss. LSGAN proposed an algorithm to achieve stable training and quality outputs. ACGAN convincingly improved the quality of the conditional generation of MNIST digits by requiring the discriminator to perform classification task on top of determining whether the input image is fake or real.

In the next chapter, we'll study how to control the attributes of generator outputs. Whilst CGAN and ACGAN are able to indicate the desired digits to produce; we have not analyzed GANs that can specify the attributes of outputs. For example, we may want to control the writing style of the MNIST digits such as roundness, tilt angle, and thickness...

References

1. Ian Goodfellow and others. Generative Adversarial Nets. Advances in neural information processing systems, 2014(http://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf).
2. Martin Arjovsky, Soumith Chintala, and Léon Bottou, Wasserstein GAN. arXiv preprint, 2017(https://arxiv.org/pdf/1701.07875.pdf).
3. Xudong Mao and others. Least Squares Generative Adversarial Networks. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE 2017(http://openaccess.thecvf.com/content_ICCV_2017/papers/Mao_Least_Squares_Generative_ICCV_2017_paper.pdf).
4. Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional Image Synthesis with Auxiliary Classifier GANs. ICML, 2017(http://proceedings.mlr.press/v70/odena17a/odena17a.pdf).