Combating Mode Collapse in GAN Training: An Empirical Analysis using Hessian Eigenvalues

Durall, Ricard; Chatzimichailidis, Avraam; Labus, Peter; Keuper, Janis

doi:10.5220/0010167902110218

Cited by 22 publications

(13 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The training of a GAN model is not a simple task to perform, due to a complicated nature of finding an optimal point upon which both G and D can converge on, which many studies describe as an equivalent to the search for Nash equilibrium in a high-dimensional, highly non-convex optimization space [ 25 ]. Both G and D need to be closely monitored and trained in relation to each other.…”

Section: Methodsmentioning

confidence: 99%

Multiclass Image Classification Using GANs and CNN Based on Holes Drilled in Laminated Chipboard

Wieczorek

Chlebuś

Gajda

et al. 2021

Sensors

View full text Add to dashboard Cite

The multiclass prediction approach to the problem of recognizing the state of the drill by classifying images of drilled holes into three classes is presented. Expert judgement was made on the basis of the quality of the hole, by dividing the collected photographs into the classes: “very fine,” “acceptable,” and “unacceptable.” The aim of the research was to create a model capable of identifying different levels of quality of the holes, where the reduced quality would serve as a warning that the drill is about to wear down. This could reduce the damage caused by a blunt tool. To perform this task, real-world data were gathered, normalized, and scaled down, and additional instances were created with the use of data-augmentation techniques, a self-developed transformation, and with general adversarial networks. This approach also allowed us to achieve a slight rebalance of the dataset, by creating higher numbers of images belonging to the less-represented classes. The datasets generated were then fed into a series of convolutional neural networks, with different numbers of convolution layers used, modelled to carry out the multiclass prediction. The performance of the so-designed model was compared to predictions generated by Microsoft’s Custom Vision service, trained on the same data, which was treated as the benchmark. Several trained models obtained by adjusting the structure and hyperparameters of the model were able to provide better recognition of less-represented classes than the benchmark.

show abstract

Section: Methodsmentioning

confidence: 99%

Multiclass Image Classification Using GANs and CNN Based on Holes Drilled in Laminated Chipboard

Wieczorek

Chlebuś

Gajda

et al. 2021

Sensors

View full text Add to dashboard Cite

show abstract

“…The authors in [ 24 ] evaluated and compared the performance of WGAN, AdaGAN, VEEGAN, and Unrolled GAN concerning mode collapse limitations. The experiments were performed over both synthetic and real datasets using the same architectures and training procedures for the GAN algorithms.…”

Section: Gan Framework and Problem Statementmentioning

confidence: 99%

“…For the coverage metrics (also measures the distance between real and generated sample distribution), VEEGAN recorded the highest coverage on the 2D ring dataset, WGAN and AdaGAN recorded the highest coverage on the 2D ring dataset. The WGAN generated low quality samples [ 24 ].…”

Section: Gan Framework and Problem Statementmentioning

confidence: 99%

“…For the high dimensional synthetic dataset with nine modes, AdaGAN was able to capture all the modes while VEEGAN was able to generate samples that were closest to the real data, had the highest quality samples, and captured six modes. As for Unrolled GAN, it failed to capture any modes [ 24 ]. The experiment also covered MNIST real dataset.…”

Section: Gan Framework and Problem Statementmentioning

confidence: 99%

“…The experiment also covered MNIST real dataset. The experiment showed that AdaGAn had the lowest Wassertein distance while all methods except VEEGAN were able to achieve a score of 0.9 for coverage metrics and captured each of the 10 modes [ 24 ]. The authors of [ 24 ] also applied second-order gradient information to tackle the issue of mode collapse.…”

Section: Gan Framework and Problem Statementmentioning

confidence: 99%

See 2 more Smart Citations

On the Performance of Generative Adversarial Network by Limiting Mode Collapse for Malware Detection Systems

Murray

Rawat

2021

Sensors

View full text Add to dashboard Cite

Generative adversarial network (GAN) has been regarded as a promising solution to many machine learning problems, and it comprises of a generator and discriminator, determining patterns and anomalies in the input data. However, GANs have several common failure modes. Typically, a mode collapse occurs when a GAN fails to fit the set optimizations and leads to several instabilities in the generative model, diminishing the capability to generate new content regardless of the dataset. In this paper, we study conditional limiter solutions for mode collapse for the Intrusion Detection System (IDS) Control Flow GAN (ICF-GAN) model. Specifically, the ICF-GAN’s mode collapse instances are limited by a mini-batch method that significantly improves the model accuracy. Performance evaluation is conducted using numerical results obtained from experiments.

show abstract

Generative Adversarial Networks: A Survey on Training, Variants, and Applications

Farajzadeh-Zanjani

Razavi–Far

Saif

et al. 2022

Intelligent Systems Reference Library

View full text Add to dashboard Cite

Combating Mode Collapse in GAN Training: An Empirical Analysis using Hessian Eigenvalues

Cited by 22 publications

References 0 publications

Multiclass Image Classification Using GANs and CNN Based on Holes Drilled in Laminated Chipboard

Multiclass Image Classification Using GANs and CNN Based on Holes Drilled in Laminated Chipboard

On the Performance of Generative Adversarial Network by Limiting Mode Collapse for Malware Detection Systems

Generative Adversarial Networks: A Survey on Training, Variants, and Applications

Contact Info

Product

Resources

About