Generative Adversarial Network Loss Research Articles

Non-coplanar CBCT image reconstruction using a generative adversarial network for non-coplanar radiotherapy.

To develop a non-coplanar cone-beam computed tomography (CBCT) image reconstruction method using projections within a limited angle range for non-coplanar radiotherapy. A generative adversarial network (GAN) was utilized to reconstruct non-coplanar CBCT images. Data from 40 patients with brain tumors and two head phantoms were used in this study. In the training stage, the generator of the GAN used coplanar CBCT and non-coplanar projections as the input, and an encoder with a dual-branch structure was utilized to extract features from the coplanar CBCT and non-coplanar projections separately. Non-coplanar CBCT images were then reconstructed using a decoder by combining the extracted features. To improve the reconstruction accuracy of the image details, the generator was adversarially trained using a patch-based convolutional neural network as the discriminator. A newly designed joint loss was used to improve the global structure consistency rather than the conventional GAN loss. The proposed model was evaluated using data from eight patients and two phantoms at four couch angles (±45°,±90°) that are most commonly used for brain non-coplanar radiotherapy in our department. The reconstructed accuracy was evaluated by calculating the root mean square error (RMSE) and an overall registration error ε, computed by integrating the rigid transformation parameters. In both patient data and phantom data studies, the qualitative and quantitative metrics results indicated that±45° couch angle models performed better than±90° couch angle models and had statistical differences. In the patient data study, the mean RMSE and ε values of couch angle at 45°, -45°, 90°, and -90° were 58.5 HU and 0.42mm, 56.8 HU and 0.41mm, 73.6 HU and 0.48mm, and 65.3 HU and 0.46mm, respectively. In the phantom data study, the mean RMSE and ε values of couch angle at 45°, -45°, 90°, and -90° were 91.2 HU and 0.46mm, 95.0 HU and 0.45mm, 114.6 HU and 0.58mm, and 102.9 HU and 0.52mm, respectively. The results show that the reconstructed non-coplanar CBCT images can potentially enable intra-treatment three-dimensional position verification for non-coplanar radiotherapy.

A deep generative model for selecting representative periods in renewable energy-integrated power systems

The extensive integration of renewable energies into power systems has led to a challenging and computationally demanding scenario for the system planning. This is due to the increased number of time series involved and the greater complexity of time series data integrated into power systems. In this paper, a new deep learning-based time aggregation method is proposed for selecting representative periods for renewable energy-integrated power system datasets where multiple variable energy resources are present. Relying on the capabilities of deep generative models, especially Generative Adversarial Network (GAN), the proposed method selects a representative period, including one or more representative days, obtained from multiple time series with spatio-temporal correlations among them. The proposed approach contains the Long Short-Term Memory (LSTM) Network in both generator and discriminator parts. Also, it includes a loss term, specifically designed for clustering tasks, in the minimax objective of the vanilla GAN loss function to enhance the clustering performance in the latent space. Furthermore, the learning rate of the proposed model, as the most important parameter of the learning algorithm, is adaptively fine-tuned during the training process, based on the training error, to enhance its learning performance. Two real-world test cases with various datasets and time series are used for data-based and model-based evaluations. Results obtained on these two real-world test cases confirm the superiority of the proposed model with respect to the state-of-the-art conventional and deep learning-based models, obtaining the performance improvement in the range of [38.55 %, 46.14 %] in terms of Mean Absolute Percentage Error (MAPE), in the range of [45.75 %, 70.16 %] in terms of Root Mean Squared Error (RMSE), and in the range of [65.81 %, 74.58 %] in terms of Mean Absolute Error (MAE). The proposed model can significantly enhance the tractability and scalability of the planning problem of renewable energy-integrated power systems, facilitating renewable energy integration into power systems which is a key issue for net zero transitioning of communities.

Image synthesis of interictal SPECT from MRI and PET using machine learning.

Cross-modality image estimation can be performed using generative adversarial networks (GANs). To date, SPECT image estimation from another medical imaging modality using this technique has not been considered. We evaluate the estimation of SPECT from MRI and PET, and additionally assess the necessity for cross-modality image registration for GAN training. We estimated interictal SPECT from PET and MRI as a single-channel input, and as a multi-channel input to the GAN. We collected data from 48 individuals with epilepsy and converted them to 3D isotropic images for consistence across the modalities. Training and testing data were prepared in native and template spaces. The Pix2pix framework within the GAN network was adopted. We evaluated the addition of the structural similarity index metric to the loss function in the GAN implementation. Root-mean-square error, structural similarity index, and peak signal-to-noise ratio were used to assess how well SPECT images were able to be synthesised. High quality SPECT images could be synthesised in each case. On average, the use of native space images resulted in a 5.4% percentage improvement in SSIM than the use of images registered to template space. The addition of structural similarity index metric to the GAN loss function did not result in improved synthetic SPECT images. Using PET in either the single channel or dual channel implementation led to the best results, however MRI could produce SPECT images close in quality. Synthesis of SPECT from MRI or PET can potentially reduce the number of scans needed for epilepsy patient evaluation and reduce patient exposure to radiation.

GAN‐MD: A myocarditis detection using multi‐channel convolutional neural networks and generative adversarial network‐based data augmentation

AbstractMyocarditis is a significant public health concern because of its potential to cause heart failure and sudden death. The standard invasive diagnostic method, endomyocardial biopsy, is typically reserved for cases with severe complications, limiting its widespread use. Conversely, non‐invasive cardiac magnetic resonance (CMR) imaging presents a promising alternative for detecting and monitoring myocarditis, because of its high signal contrast that reveals myocardial involvement. To assist medical professionals via artificial intelligence, the authors introduce generative adversarial networks ‐ multi discriminator (GAN‐MD), a deep learning model that uses binary classification to diagnose myocarditis from CMR images. Their approach employs a series of convolutional neural networks (CNNs) that extract and combine feature vectors for accurate diagnosis. The authors suggest a novel technique for improving the classification precision of CNNs. Using generative adversarial networks (GANs) to create synthetic images for data augmentation, the authors address challenges such as mode collapse and unstable training. Incorporating a reconstruction loss into the GAN loss function requires the generator to produce images reflecting the discriminator features, thus enhancing the generated images' quality to more accurately replicate authentic data patterns. Moreover, combining this loss function with other regularisation methods, such as gradient penalty, has proven to further improve the performance of diverse GAN models. A significant challenge in myocarditis diagnosis is the imbalance of classification, where one class dominates over the other. To mitigate this, the authors introduce a focal loss‐based training method that effectively trains the model on the minority class samples. The GAN‐MD approach, evaluated on the Z‐Alizadeh Sani myocarditis dataset, achieves superior results (F‐measure 86.2%; geometric mean 91.0%) compared with other deep learning models and traditional machine learning methods.

Deep Learning-Based Delayed PET Image Synthesis from Corresponding Early Scanned PET for Dosimetry Uptake Estimation.

The acquisition of in vivo radiopharmaceutical distribution through imaging is time-consuming due to dosimetry, which requires the subject to be scanned at several time points post-injection. This study aimed to generate delayed positron emission tomography images from early images using a deep-learning-based image generation model to mitigate the time cost and inconvenience. Eighteen healthy participants were recruited and injected with [18F]Fluorodeoxyglucose. A paired image-to-image translation model, based on a generative adversarial network (GAN), was used as the generation model. The standardized uptake value (SUV) mean of the generated image of each organ was compared with that of the ground-truth. The least square GAN and perceptual loss combinations displayed the best performance. As the uptake time of the early image became closer to that of the ground-truth image, the translation performance improved. The SUV mean values of the nominated organs were estimated reasonably accurately for the muscle, heart, liver, and spleen. The results demonstrate that the image-to-image translation deep learning model is applicable for the generation of a functional image from another functional image acquired from normal subjects, including predictions of organ-wise activity for specific normal organs.

Multimodal-Boost: Multimodal Medical Image Super-Resolution Using Multi-Attention Network With Wavelet Transform.

Multimodal medical images are widely used by clinicians and physicians to analyze and retrieve complementary information from high-resolution images in a non-invasive manner. Loss of corresponding image resolution adversely affects the overall performance of medical image interpretation. Deep learning-based single image super resolution (SISR) algorithms have revolutionized the overall diagnosis framework by continually improving the architectural components and training strategies associated with convolutional neural networks (CNN) on low-resolution images. However, existing work lacks in two ways: i) the SR output produced exhibits poor texture details, and often produce blurred edges, ii) most of the models have been developed for a single modality, hence, require modification to adapt to a new one. This work addresses (i) by proposing generative adversarial network (GAN) with deep multi-attention modules to learn high-frequency information from low-frequency data. Existing approaches based on the GAN have yielded good SR results; however, the texture details of their SR output have been experimentally confirmed to be deficient for medical images particularly. The integration of wavelet transform (WT) and GANs in our proposed SR model addresses the aforementioned limitation concerning textons. While the WT divides the LR image into multiple frequency bands, the transferred GAN uses multi-attention and upsample blocks to predict high-frequency components. Additionally, we present a learning method for training domain-specific classifiers as perceptual loss functions. Using a combination of multi-attention GAN loss and a perceptual loss function results in an efficient and reliable performance. Applying the same model for medical images from diverse modalities is challenging, our work addresses (ii) by training and performing on several modalities via transfer learning. Using two medical datasets, we validate our proposed SR network against existing state-of-the-art approaches and achieve promising results in terms of structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR).

Unpaired deep adversarial learning for multi-class segmentation of instruments in robot-assisted surgical videos.

Image segmentation of instruments in the raw surgical videos is a critical component of intraoperative assistance softwares. Challenges include addressing rendered overlays occluding the instrument while providing pivotal input to instrument tracking frameworks and, train the segmentation process with limited labelled data available from surgical videos. The proposed adversarial network, InstruSegNet uses unpaired training (eliminating need for massive paired data) for automated multi-class surgical instrument segmentation in raw surgical videos with complex backgrounds. The proposed method is applied for single/multiple robotic and rigid instruments and optimised on least square Generative Adversarial Networks loss. Promising validation has been conducted on the publicly available dataset. Proposed approach for multi-class segmentation of robotic and rigid instruments meets outstanding performance in terms of accuracy and surpasses the existing methods. This work facilitates segmenting instrument information without manual interventions from raw videos providing means to code surgeon's actions for developing intelligent assistance software.

UPHDR-GAN: Generative Adversarial Network for High Dynamic Range Imaging With Unpaired Data

The paper proposes a method to effectively fuse multi-exposure inputs and generate high-quality high dynamic range (HDR) images with unpaired datasets. Deep learning-based HDR image generation methods rely heavily on paired datasets. The ground truth images play a leading role in generating reasonable HDR images. Datasets without ground truth are hard to be applied to train deep neural networks. Recently, Generative Adversarial Networks (GAN) have demonstrated their potentials of translating images from source domain X to target domain Y in the absence of paired examples. In this paper, we propose a GAN-based network for solving such problems while generating enjoyable HDR results, named UPHDR-GAN. The proposed method relaxes the constraint of the paired dataset and learns the mapping from the LDR domain to the HDR domain. Although the pair data are missing, UPHDR-GAN can properly handle the ghosting artifacts caused by moving objects or misalignments with the help of the modified GAN loss, the improved discriminator network and the useful initialization phase. The proposed method preserves the details of important regions and improves the total image perceptual quality. Qualitative and quantitative comparisons against the representative methods demonstrate the superiority of the proposed UPHDR-GAN.

Multiple auxiliary classifiers GAN for controllable image generation: Application to license plate recognition

One of the main challenges in developing machine learning (ML) applications is the lack of labeled and balanced datasets. In the literature, different techniques tackle this problem via augmentation, rendering, and over-sampling. Still, these methods produce datasets that appear less natural, exhibit poor balance, and have less variation. One potential solution is to leverage the Generative Adversarial Network (GAN) which achieves remarkable results in the generation of high-fidelity natural images. However, expanding the ability of GANs' to control generated image attributes with supervisory information remains a challenge. This research aims to propose an efficient method to generate high-fidelity natural images with total control of its main attributes. Therefore, this paper proposes a novel Multiple Auxiliary Classifiers GAN (MAC-GAN) framework based on Auxiliary Classifier GAN (AC-GAN), multi-conditioning, Wasserstein distance, gradient penalty, and dynamic loss. It is therefore presented as an efficient solution for highly controllable image synthesis red that allows to enrich and re-balance datasets beyond data augmentation. Furthermore, the effectiveness of MAC-GAN images on a target ML application called Automatic License Plate Recognition (ALPR) under limited resource constraints is probed. The improvement achieved is over 5% accuracy, which is mainly due to the ability of the MAC-GAN to create a balanced dataset with controllable synthesis and produce multiple (different) images with the same attributes, thus increasing the variation of the dataset in a more elaborate way than data augmentation techniques.

AAGAN: Accuracy-Aware Generative Adversarial Network for Supervised Tasks

Owing to their great success in unsupervised tasks, generative adversarial networks (GANs) are widely adopted for supervised (conditional) image-generation tasks such as in-painting. Unfortunately, designing an objective function using GANs is not trivial. For supervised tasks, the generator trained only with GAN loss does not yield the corresponding target for the given input because GAN is trained to match the data distribution, not to find the exact answer. Therefore, the loss function of a generator is often formulated by linearly combining supervised loss and GAN loss as similar to multi-task learning, expecting that each loss’s weakness is complemented. Contrary to expectation, both losses cause a conflict in practice due to different optimum of each loss, yielding low objective and subjective image qualities. To address this problem, we empirically investigated the conflict caused by using a conventional GAN with pixel-wise losses; we then propose a novel (relativistic) accuracy-aware discriminator. Based on the proposed discriminator, we developed an accuracy-aware GAN (AAGAN) and proved its optimality under an ideal assumption. We then propose a relativistic accuracy-aware GAN (RAAGAN) by considering practical assumptions. Experimental results on supervised tasks demonstrated that the proposed schemes alleviated the competition between losses and outperformed conventional GANs in terms of both objective and subjective qualities.

Attentive Generative Adversarial Network with Dual Encoder-Decoder for Shadow Removal

Shadow removal is a fundamental task that aims at restoring dark areas in an image where the light source is blocked by an opaque object, to improve the visibility of shadowed areas. Existing shadow removal methods have developed for decades and yielded many promising results, but most of them are poor at maintaining consistency between shadowed regions and shadow-free regions, resulting in obvious artifacts in restored areas. In this paper, we propose a two-stage (i.e., shadow detection and shadow removal) method based on the Generative Adversarial Network (GAN) to remove shadows. In the shadow detection stage, a Recurrent Neural Network (RNN) is trained to obtain the attention map of shadowed areas. Then the attention map is injected into both generator and discriminator to guide the shadow removal stage. The generator is a dual encoder-decoder that processes the shadowed regions and shadow-free regions separately to reduce inconsistency. The whole network is trained with a spatial variant reconstruction loss along with the GAN loss to make the recovered images more natural. In addition, a novel feature-level perceptual loss is proposed to ensure enhanced images more similar to ground truths. Quantitative metrics like PSNR and SSIM on the ISTD dataset demonstrate that our method outperforms other compared methods. In the meantime, the qualitative comparison shows our approach can effectively avoid artifacts in the restored shadowed areas while keeping structural consistency between shadowed regions and shadow-free regions.

A variability aware GAN for improving spatial representativeness of discrete geobodies

Generative Adversarial Networks (GAN) have shown great potential in not only producing acceptable realizations of geologically complex models but also successfully reparametrizing them. Training GANs is quite challenging. One such challenge is mode collapse. When generating realizations of spatial property, mode collapse causes reduction in variability, compared to the input training dataset, and thus, the realizations become spatially biased at specific locations. To address this issue, we developed a new GAN architecture where a regularization term is introduced to maintain the variability and reduce mode collapse. This is achieved by using a probability map to evaluate variability and spatial bias of generated realizations and modifying the GAN loss function to minimize this bias. We applied the new architecture to a binary channelized permeability distribution and compared the results with those generated by Deep Convolutional GAN (DCGAN) and Wasserstein GAN with gradient penalty (WGAN-GP). Our results show that the proposed architecture significantly enhances variability and reduces the spatial bias induced by mode collapse, outperforming both DCGAN and WGAN-GP in the application of generating subsurface property distributions.

ECG Biosignal Deidentification Using Conditional Generative Adversarial Networks.

Electrocardiogram (ECG) signals provide rich information on individuals' potential cardiovascular conditions and disease, ranging from coronary artery disease to the risk of a heart attack. While health providers store and share these information for medical and research purposes, such data is highly vulnerable to privacy concerns, similar to many other types of healthcare data. Recent works have shown the feasibility of identifying and authenticating individuals by using ECG as a biometric due to the highly individualized nature of ECG signals. However, to the best of our knowledge, there does not exist a method in the literature attempting to de-identify ECG signals. In this paper, to address this privacy protection gap, we propose a Generative Adversarial Network (GAN)-based framework for de-identification of ECG signals. We leverage a combination of a standard GAN loss, an Ordinary Differential Equations (ODE)-based, and identity-based loss values to train a generator that de-identifies a ECG signal while preserving structure the ECG signal and information regarding the target cardio vascular condition. We evaluate our framework in terms of both qualitative and quantitative metrics considering different weightings over the above-mentioned losses. Our experiments demonstrate the efficiency of our framework in terms of privacy protection and ECG signal structural preservation.

Multi-Conditional Constraint Generative Adversarial Network-Based MR Imaging from CT Scan Data.

Magnetic resonance (MR) imaging is an important computer-aided diagnosis technique with rich pathological information. The factor of physical and physiological constraint seriously affects the applicability of that technique. Thus, computed tomography (CT)-based radiotherapy is more popular on account of its imaging rapidity and environmental simplicity. Therefore, it is of great theoretical and practical significance to design a method that can construct an MR image from the corresponding CT image. In this paper, we treat MR imaging as a machine vision problem and propose a multi-conditional constraint generative adversarial network (GAN) for MR imaging from CT scan data. Considering reversibility of GAN, both generator and reverse generator are designed for MR and CT imaging, respectively, which can constrain each other and improve consistency between features of CT and MR images. In addition, we innovatively treat the real and generated MR image discrimination as object re-identification; cosine error fusing with original GAN loss is designed to enhance verisimilitude and textural features of the MR image. The experimental results with the challenging public CT-MR image dataset show distinct performance improvement over other GANs utilized in medical imaging and demonstrate the effect of our method for medical image modal transformation.

Formula omitted]-GAN: Robust generative adversarial networks

Generative adversarial networks (GAN) training is subject to problems including mode collapse, gradient vanishing, and instability. Although many different losses have been proposed to alleviate these shortcomings, they heavily rely on a fixed-value function with limited expressive power in terms of robustness, whereby failing to perform consistently over multiple data sets. To solve this problem, we propose a parametric and robust αβ-loss function that can improve the performances of GAN on different data sets. Specifically, unlike standard GAN loss function it exploits the αβ-divergence (AB-divergence) to weigh the likelihood ratio associated with each data point. This weighing mechanism makes the model robust to noises and yields better models in terms of FID score. To reduce the cost of searching for the optimal α and β, we further propose an adaptive version to systematically update these parameters according to statistics of the discriminator’s output. Moreover, αβ-loss can be reduced to Least Square GAN (LS-GAN) and standard GAN (SGAN) loss function as special cases. We conduct extensive experiments on both synthetic and real-world data sets. Experimental results over the synthetic data sets (2D Gaussian ring and grid) demonstrate that our approach can significantly alleviate the issue of mode collapse. Additionally, by constraining the gradient of the discriminator that is fed back to the generator via finely adjusting the hyper-parameters α and β, our approach can improve the quality of synthetic images, as can be seen from the decrease of FID from 40 to 23.71 on the data set CIFAR10 using the SN-DCGAN architecture.

Throwaway Shadows Using Parallel Encoders Generative Adversarial Network

Face photographs taken on a bright sunny day or in floodlight contain unnecessary shadows of objects on the face. Most previous works deal with removing shadow from scene images and struggle with doing so for facial images. Faces have a complex semantic structure, due to which shadow removal is challenging. The aim of this research is to remove the shadow of an object in facial images. We propose a novel generative adversarial network (GAN) based image-to-image translation approach for shadow removal in face images. The first stage of our model automatically produces a binary segmentation mask for the shadow region. Then, the second stage, which is a GAN-based network, removes the object shadow and synthesizes the effected region. The generator network of our GAN has two parallel encoders—one is standard convolution path and the other is a partial convolution. We find that this combination in the generator results not only in learning an incorporated semantic structure but also in disentangling visual discrepancies problems under the shadow area. In addition to GAN loss, we exploit low level L1, structural level SSIM and perceptual loss from a pre-trained loss network for better texture and perceptual quality, respectively. Since there is no paired dataset for the shadow removal problem, we created a synthetic shadow dataset for training our network in a supervised manner. The proposed approach effectively removes shadows from real and synthetic test samples, while retaining complex facial semantics. Experimental evaluations consistently show the advantages of the proposed method over several representative state-of-the-art approaches.

Equipment Maintenance Support Effectiveness Evaluation Based on Improved Generative Adversarial Network and Radial Basis Function Network

Due to the lack of maintenance support samples, maintenance support effectiveness evaluation based on the deep neural network often faces the problem of small sample overfitting and low generalization ability. In this paper, a neural network evaluation model based on an improved generative adversarial network (GAN) and radial basis function (RBF) network is proposed to amplify maintenance support samples. It adds category constraint based on category probability vector reordering function to GAN loss function, avoids the simplification of generated sample categories, and enhances the quality of generated samples. It also designs a parameter initialization method based on parameter components equidistant variation for RBF network, which enhances the response of correct feature information and reduces the risk of training overfitting. The comparison results show that the mean square error (MSE) of the improved GAN-RBF model is5.921×10−4, which is approximately 1/2 of the RBF model, 1/3 of the Elman model, and 1/5 of the BP model, while its complexity remains at a reasonable level. Compared with traditional neural network evaluation methods, the improved GAN-RBF model has higher evaluation accuracy, better solves the problem of poor generalization ability caused by insufficient training samples, and can be more effectively applied to maintenance support effectiveness evaluation. At the same time, it also provides a good reference for evaluation research in other fields.

Patch-Based DCNN Method for CBCT Image Enhancement

Cone-beam computed tomography (CBCT) often has suboptimal image quality compared with standard fan-beam CT due to increased sensitivity to scattering and motion artifacts. In radiation treatment of cancers, due to the convenience in image acquisition as it is fully incorporated into the treatment station, CBCT plays a significant role in treatment plan re-evaluation and adaptation in adaptive radiation therapy but is limited by the image quality. In this paper, we propose a deep convolutional neural network (DCNN) based method for head and neck (HN) CBCT enhancement, where the network learns to transform the low-quality CBCT images to high-quality images.CT and CBCT image pairs from 50 HN cancer patients who received radiotherapy treatment were used for model training, validation and testing. An additional 22 raw CBCT images from another vendor were also used for testing. Pre-processing procedures include rigid registration, center-cropping, and Otsu's thresholding to segment the anatomical region. Data quality control was applied before training a 2D U-Net to reduce the adverse impact from the unmatched pairs. A patch-based strategy was employed during training, since the randomly extracted patches could reduce the receptive field and thus reduce the probability of overfitting based on the global anatomy. We explored the impact of patch size by varying it to be 64 × 64, 128 × 128, 192 × 192 and 256 × 256. After fixing to the optimal patch size, a Pix2Pix network combining L1 loss and conditional generative adversarial network (GAN) loss was also trained.It was found that the U-Net trained with a patch size of 128 × 128 reached the optimal performance and improved the mean absolute error (MAE) from 57.91 to 34.78 Hounsfield units (HU), peak signal-to-noise ratio (PSNR) from 30. 56 to 32.84 dB, and structural similarity (SSIM) from 0.7546 to 0.8455 compared with the original CBCT. With the addition of conditional GAN loss, the trained Pix2Pix network better kept the image details, especially on the external testing set.In this work, we developed a DCNN-based method to improve the CBCT image quality. This method has the advantages of high adaptability, vendor neutrality and low cost, since it operates purely in the image domain. Our data quality control pipeline automatically removed the unmatched CT and CBCT image pairs from the training set. The proposed DCNN-based method with the patch-based training strategy showed promising results for fast CBCT image enhancement.

Low-light Image Enhancement Based on Conditional Generative Adversarial Network

We present an end-to-end low-light image enhancement learning method. This learning is based on the conditional generative adversarial networks(GAN) and realizes low-light images enhancement. Specifically, our method uses a convolutional neural network containing residual structures as a generator and WGAN-GP as a discriminator to generate an effective low-light enhancement model under the constraints of GAN loss, Perceptual loss and Structural similarity loss. The model can retain the detailed information of the original image, improve the brightness of the image without generating noise interference, while the generated images are more natural and have higher quality. Extensive experimental results show that our method has reached the state-of-art in multiple objective evaluation indicators of image quality, and the visual appearance is superior.

Least kth-Order and Rényi Generative Adversarial Networks.

We investigate the use of parameterized families of information-theoretic measures to generalize the loss functions of generative adversarial networks (GANs) with the objective of improving performance. A new generator loss function, least kth-order GAN (LkGAN), is introduced, generalizing the least squares GANs (LSGANs) by using a kth-order absolute error distortion measure with k≥1 (which recovers the LSGAN loss function when k=2). It is shown that minimizing this generalized loss function under an (unconstrained) optimal discriminator is equivalent to minimizing the kth-order Pearson-Vajda divergence. Another novel GAN generator loss function is next proposed in terms of Rényi cross-entropy functionals with order α>0, α≠1. It is demonstrated that this Rényi-centric generalized loss function, which provably reduces to the original GAN loss function as α→1, preserves the equilibrium point satisfied by the original GAN based on the Jensen-Rényi divergence, a natural extension of the Jensen-Shannon divergence. Experimental results indicate that the proposed loss functions, applied to the MNIST and CelebA data sets, under both DCGAN and StyleGAN architectures, confer performance benefits by virtue of the extra degrees of freedom provided by the parameters k and α, respectively. More specifically, experiments show improvements with regard to the quality of the generated images as measured by the Fréchet inception distance score and training stability. While it was applied to GANs in this study, the proposed approach is generic and can be used in other applications of information theory to deep learning, for example, the issues of fairness or privacy in artificial intelligence.