Recurrent Network Architecture Research Articles

Multi-Directional Long-Term Recurrent Convolutional Network for Road Situation Recognition.

Understanding road conditions is essential for implementing effective road safety measures and driving solutions. Road situations encompass the day-to-day conditions of roads, including the presence of vehicles and pedestrians. Surveillance cameras strategically placed along streets have been instrumental in monitoring road situations and providing valuable information on pedestrians, moving vehicles, and objects within road environments. However, these video data and information are stored in large volumes, making analysis tedious and time-consuming. Deep learning models are increasingly utilized to monitor vehicles and identify and evaluate road and driving comfort situations. However, the current neural network model requires the recognition of situations using time-series video data. In this paper, we introduced a multi-directional detection model for road situations to uphold high accuracy. Deep learning methods often integrate long short-term memory (LSTM) into long-term recurrent network architectures. This approach effectively combines recurrent neural networks to capture temporal dependencies and convolutional neural networks (CNNs) to extract features from extensive video data. In our proposed method, we form a multi-directional long-term recurrent convolutional network approach with two groups equipped with CNN and two layers of LSTM. Additionally, we compare road situation recognition using convolutional neural networks, long short-term networks, and long-term recurrent convolutional networks. The paper presents a method for detecting and recognizing multi-directional road contexts using a modified LRCN. After balancing the dataset through data augmentation, the number of video files increased, resulting in our model achieving 91% accuracy, a significant improvement from the original dataset.

Heterogeneous Forgetting Rates and Greedy Allocation in Slot-Based Memory Networks Promotes Signal Retention.

A key question in the neuroscience of memory encoding pertains to the mechanisms by which afferent stimuli are allocated within memory networks. This issue is especially pronounced in the domain of working memory, where capacity is finite. Presumably the brain must embed some "policy" by which to allocate these mnemonic resources in an online manner in order to maximally represent and store afferent information for as long as possible and without interference from subsequent stimuli. Here, we engage this question through a top-down theoretical modeling framework. We formally optimize a gating mechanism that projects afferent stimuli onto a finite number of memory slots within a recurrent network architecture. In the absence of external input, the activity in each slot attenuates over time (i.e., a process of gradual forgetting). It turns out that the optimal gating policy consists of a direct projection from sensory activity to memory slots, alongside an activity-dependent lateral inhibition. Interestingly, allocating resources myopically (greedily with respect to the current stimulus) leads to efficient utilization of slots over time. In other words, later-arriving stimuli are distributed across slots in such a way that the network state is minimally shifted and so prior signals are minimally "overwritten." Further, networks with heterogeneity in the timescales of their forgetting rates retain stimuli better than those that are more homogeneous. Our results suggest how online, recurrent networks working on temporally localized objectives without high-level supervision can nonetheless implement efficient allocation of memory resources over time.

Inappropriate YouTube content detection and classification by using proposed novel auto-determined k-means clustering and PDBRNN architecture

To ensure a safe and pleasant user experience while watching content on YouTube, it is necessary to identify and classify inappropriate content, especially content that is inappropriate for children. In this work, we have concentrated on establishing an efficient system for detecting inappropriate content on YouTube. Most of the work focuses on manual pre-processing; however, it takes too much time, requires manpower support, and is not ideal for solving real-time problems. To address this challenge, we have proposed an automatic preprocessing scheme for selecting appropriate frames and removing unwanted frames such as noise and duplicate frames. For this purpose, we have utilized the proposed novel auto-determined k-means (PADK-means) algorithm. Our PADK-means algorithm automatically determines the optimal cluster count instead of manual specifications. By doing this, we have solved the manual cluster count specification problem in the traditional k-means clustering algorithm. On the other hand, to improve the system’s performance, we utilized the Proposed Feature Extraction (PFE) method, which includes two pre-trained models DenseNet121 and Inception V3 are utilized to extract local and global features from the frame. Finally, we employ a proposed double-branch recurrent network (PDBRNN) architecture, which includes bi-LSTM and GRU, to classify the video as appropriate or inappropriate. Our proposed automatic preprocessing mechanism, proposed feature extraction method, and proposed double-branch RNN classifier yielded an impressive accuracy of 97.9%.

Deep learning-based human activity recognition using CNN, ConvLSTM, and LRCN

Human activity recognition (HAR) plays a crucial role in assisting the elderly and individuals with vascular dementia by providing support and monitoring for their daily activities. This paper presents a deep learning (DL)-based approach to HAR, leveraging convolutional neural network (CNN), convolutional long short-term memory (ConvLSTM), and long-term recurrent convolutional network (LRCN) architectures. These models are designed to extract spatial features and capture temporal dependencies in video data, enhancing the accuracy of activity classification. We conducted experiments on the UCF50 and HMDB51 video datasets, encompassing diverse human activities. Our evaluation demonstrates that the ConvLSTM model achieves an accuracy of 82% on UCF50 and 68% on HMDB51, while the LRCN model gives accuracies of 93.44% and 71.55%, respectively. Finally, the CNN model outperforms with an accuracy rate of 99.58% for the UCF50 and 92.70% for the HMDB51 datasets. These significant improvements showcase the effectiveness of integrating convolutional and recurrent neural networks for HAR tasks. Our research contributes to advancing HAR systems with potential applications in healthcare, assisted living, and surveillance. By accurately recognizing human activities, our models can assist in remote patient monitoring, fall detection, and public safety initiatives. These findings underscore the importance of DL in enhancing the quality of life and safety for individuals in various contexts.

Surrogate-Assisted Cooperative Co-evolutionary Reservoir Architecture Search for Liquid State Machines

Liquid state machines (LSMs) are biologically more plausible than feedforward spiking neural networks for brain-inspired computing and neuromorphic engineering. However, optimizing and training complex recurrent network architectures in the reservoir of LSMs remains challenging. Most existing algorithms aim to adjust the synaptic strength only, without fundamentally modifying the reservoir architecture of LSMs. Recently, it has become popular to simultaneously optimize the architecture and parameters of the reservoir in LSMs. However, most existing architecture representation schemes are too restricted to discover more powerful architectures of LSMs. To address the above issue, this paper proposes a generative liquid state machine, whose reservoir architecture is evolved using a cooperative co-evolutionary algorithm whose weights are tuned by synaptic plasticity rules. To reduce the computation time for evolving the reservoir, random forest is adopted to assist the cooperative co-evolutionary algorithm, together with a data parallelism strategy. The proposed algorithm is assessed on three sequence classification benchmarks and our experimental results show that the proposed algorithm outperforms the state-of-the-art on the benchmark problems. Meanwhile, our analysis shows that the data parallelism strategy is effective in speeding up the evaluation process.

Deep learning for tumor margin identification in electromagnetic imaging

In this work, a novel method for tumor margin identification in electromagnetic imaging is proposed to optimize the tumor removal surgery. This capability will enable the visualization of the border of the cancerous tissue for the surgeon prior or during the excision surgery. To this end, the border between the normal and tumor parts needs to be identified. Therefore, the images need to be segmented into tumor and normal areas. We propose a deep learning technique which divides the electromagnetic images into two regions: tumor and normal, with high accuracy. We formulate deep learning from a perspective relevant to electromagnetic image reconstruction. A recurrent auto-encoder network architecture (termed here DeepTMI) is presented. The effectiveness of the algorithm is demonstrated by segmenting the reconstructed images of an experimental tissue-mimicking phantom. The structure similarity measure (SSIM) and mean-square-error (MSE) average of normalized reconstructed results by the DeepTMI method are about 0.94 and 0.04 respectively, while that average obtained from the conventional backpropagation (BP) method can hardly overcome 0.35 and 0.41 respectively.

Multi-Sensors System and Deep Learning Models for Object Tracking.

Autonomous navigation relies on the crucial aspect of perceiving the environment to ensure the safe navigation of an autonomous platform, taking into consideration surrounding objects and their potential movements. Consequently, a fundamental requirement arises to accurately track and predict these objects' trajectories. Three deep recurrent network architectures were defined to achieve this, fine-tuning their weights to optimize the tracking process. The effectiveness of this proposed pipeline has been assessed, with diverse tracking scenarios demonstrated in both sub-urban and highway environments. The evaluations have yielded promising results, affirming the potential of this approach in enhancing autonomous navigation capabilities.

Multi-Scale Cyclic Image Deblurring Based on PVC-Resnet

Aiming at the non-uniform blurring of image caused by optical system defects or external interference factors, such as camera shake, out-of-focus, and fast movement of object, a multi-scale cyclic image deblurring model based on a parallel void convolution-Resnet (PVC-Resnet) is proposed in this paper, in which a multi-scale recurrent network architecture and a coarse-to-fine strategy are used to restore blurred images. The backbone network is built based on Unet codec architecture, where a PVC-Resnet module designed by combinations of parallel dilated convolution and residual network is constructed in the encoder of the backbone network. The convolution receptive field is expanded with parallel dilated convolution to extract richer global features. Besides, a multi-scale feature extraction module is designed to extract the shallow features of different scale targets in blurred images, and then the extracted features are sent to the backbone network for feature refinement. The SSIM loss function and the L1 loss function are combined to construct the SSIM-L1 joint loss function for the optimization of the overall network to ensure that the image restoration at different stages can be optimized. The experimental results show that the average peak signal-to-noise ratio (PSNR) of the proposed model on different data sets is as high as 32.84 dB, and the structural similarity (SSIM) reaches 0.9235. and statistical structural similarity (Stat-SSIM) of 0.9249 on different datasets. Compared with other methods, the deblurred images generated by this method are superior to the methods proposed by Nah et al., Kupyn et al. and Cho S J et al., especially on the calibration board data set. The model proposed in this paper applies parallel dilated convolution and SSIM-L1 joint loss function to improve the performance of the network so that the edge and texture details of the restored image are clearer.

REMOVAL OF RAIN COMPONENTS FROM SINGLE IMAGES USING A RECURRENT NEURAL NETWORK

Context. Removing the undesirable consequences of rain effects from single images is an actual problem in many computer vision tasks, because rain streaks can significantly degrade the visual quality of images and seriously interfere with the operation of various intelligent systems, which are used for their processing and further analysis.
 Objective. The goal of the work is to develop a method for detecting and removing undesirable effects of the rain from single images, which is based on the using of a convolutional neural network with a recurrent structure.
 Method. The main component of the proposed method is a convolutional neural network, which has a recurrent multi-stage structure. A feature of this network architecture is the use of repeated blocks (layers), at the output of which you can get an intermediate result of «cleaning» the original image. Moreover at the output of each next layer of the network we get an image with less influence of rain components than on the previous one. Each network layer contains two independent sub-networks (branches) for parallel image processing. The main branch is designed to detect and remove the effect of rain from the image and the attention branch is used to improve and speed up the process of detecting undesirable rain components (for rain attention map formation).
 Results. An approach has been developed to automatically detect and remove the rain effect from single images. The process of “cleaning” the original image is based on the use of a convolutional neural network with a recurrent structure, which was trained on the Rain100H and Rain100L datasets. The results of computer experiments, which testifies to the effectiveness and expediency of using the proposed method for solving practical tasks of pre-processing “contaminated” images are presented.
 Conclusions. The advantage of the developed method for removing undesirable components of rain from images is that the recurrent multi-stage network architecture, on which it is based allows it to be potentially applied to solving tasks under conditions of limited computing resources. The proposed method can be successfully used in the development of intelligent systems for area monitoring with surveillance cameras, autonomous vehicles control, processing aerial photography results, etc. In the future, it should be considered the possibility of forming a separate sub-network to eliminate blurring in the image and train the network on datasets that contain image samples with different components of rain, which will make the method more «resistant» to different forms of the rain effect and increase the quality of image “cleaning”.

Application of coordinate systems for vessel trajectory prediction improvement using a recurrent neural networks

According to the Global Maritime Insurance annual report, among human and non-human risk factors, the number of accidents in maritime transport remains a significant issue. One of the factors is vessel collisions and anomalies at sea. Massive historical data from automatic identification systems are analyzed, and intelligent transportation systems are being developed to solve the problem of vessel trajectory prediction. The most ordinary attempt to improve accuracy is by evaluating the historical vessel behavior and learning the patterns and similarities of the predicted vessel movements. However, this paper shows that a better forecast also may be reached by choosing a different trajectory calculation strategy. The geographical or polar coordinate system values are used in a classical way, but several modifications, such as Universal Transverse Mercator (UTM), have been proposed in this study as an alternative to Mercator’s projection coordinates. Two main positioning of the vessels motion transformations were tested: by changing the coordinates to measurements of the angular distance (haversine) and displacement angle (azimuth) functions between different time steps; degree coordinates transformation into a Cartesian system using UTM with vector subtraction. The last case improves the accuracy of almost 30% in the available data sample by using the autoencoder architecture, compared to the longitude and latitude predictions even with computed delta features. The research generally compared three recurrent network architectures (with their hyperparameter – cell sizes): Autoencoder Long Short-Term Memory, Bi-directional Long Short-Term Memory, and Gated Recurrent Unit networks. The model calculations are performed in a real historical dataset, exclusively on cargo vessel type trajectories in the Netherlands (North Sea) coastal region. Also, the methods were validated in another dataset of the Baltic Sea Region.

Twitter sentiment analysis using hybrid gated attention recurrent network

Sentiment analysis is the most trending and ongoing research in the field of data mining. Nowadays, several social media platforms are developed, among that twitter is a significant tool for sharing and acquiring peoples’ opinions, emotions, views, and attitudes towards particular entities. This made sentiment analysis a fascinating process in the natural language processing (NLP) domain. Different techniques are developed for sentiment analysis, whereas there still exists a space for further enhancement in accuracy and system efficacy. An efficient and effective optimization based feature selection and deep learning based sentiment analysis is developed in the proposed architecture to fulfil it. In this work, the sentiment 140 dataset is used for analysing the performance of proposed gated attention recurrent network (GARN) architecture. Initially, the available dataset is pre-processed to clean and filter out the dataset. Then, a term weight-based feature extraction termed Log Term Frequency-based Modified Inverse Class Frequency (LTF-MICF) model is used to extract the sentiment-based features from the pre-processed data. In the third phase, a hybrid mutation-based white shark optimizer (HMWSO) is introduced for feature selection. Using the selected features, the sentiment classes, such as positive, negative, and neutral, are classified using the GARN architecture, which combines recurrent neural networks (RNN) and attention mechanisms. Finally, the performance analysis between the proposed and existing classifiers is performed. The evaluated performance metrics and the gained value for such metrics using the proposed GARN are accuracy 97.86%, precision 96.65%, recall 96.76% and f-measure 96.70%, respectively.

Combined space–time reduced-order model with three-dimensional deep convolution for extrapolating fluid dynamics

There is a critical need for efficient and reliable active flow control strategies to reduce drag and noise in aerospace and marine engineering applications. While traditional full-order models based on the Navier–Stokes equations are not feasible, advanced model reduction techniques can be inefficient for active control tasks, especially with strong non-linearity and convection-dominated phenomena. Using convolutional recurrent autoencoder network architectures, deep-learning-based reduced-order models have been recently shown to be effective while performing several orders of magnitude faster than full-order simulations. However, these models encounter significant challenges outside the training data, limiting their effectiveness for active control and optimization tasks. In this study, we aim to improve the extrapolation capability by modifying the network architecture and integrating coupled space–time physics as an implicit bias. Reduced-order models via deep learning generally employ decoupling in spatial and temporal dimensions, which can introduce modeling and approximation errors. To alleviate these errors, we propose a novel technique for learning coupled spatial–temporal correlation using a three-dimensional convolution network. We assess the proposed technique against a standard encoder–propagator–decoder model and demonstrate a superior extrapolation performance. To demonstrate the effectiveness of the three-dimensional convolution network, we consider a benchmark problem of the flow past a circular cylinder at laminar flow conditions and use the spatiotemporal snapshots from the full-order simulations. Our proposed three-dimensional convolution architecture accurately captures the velocity and pressure fields for varying Reynolds numbers. Compared to the standard encoder–propagator–decoder network, the spatiotemporal-based three-dimensional convolution network improves the prediction range of Reynolds numbers outside of the training data.

Deep dual recurrence optical flow learning for time-resolved particle image velocimetry

Motion fields estimated from image data have been widely used in physics and engineering. Time-resolved particle image velocimetry (TR-PIV) is considered as an advanced flow visualization technique that measures multi-frame velocity fields from successive images. Contrary to conventional PIV, TR-PIV essentially estimates a velocity field video that provides both temporal and spatial information. However, performing TR-PIV with high computational efficiency and high computational accuracy is still a challenge for current algorithms. To solve these problems, we put forward a novel deep learning network named Deep-TRPIV in this study, to effectively estimate fluid motions from multi-frame particle images in an end-to-end manner. First, based on particle image data, we modify the optical flow model known as recurrent all-pairs field transforms that iteratively updates flow fields through a convolutional gated recurrent unit. Second, we specifically design a temporal recurrent network architecture based on this optical flow model by conveying features and flow information from previous frame. When N successive images are fed, the network can efficiently estimate N – 1 motion fields. Moreover, we generate a dataset containing multi-frame particle images and true fluid motions to train the network supervised. Eventually, we conduct extensive experiments on synthetic and experimental data to evaluate the performance of the proposed model. Experimental evaluation results demonstrate that our proposed approach achieves high accuracy and computational efficiency, compared with classical approaches and related deep learning models.

Traffic forecasting with graph spatial–temporal position recurrent network

With the development of social economy and smart technology, the explosive growth of vehicles has caused traffic forecasting to become a daunting challenge, especially for smart cities. Recent methods exploit graph spatial–temporal characteristics, including constructing the shared patterns of traffic data, and modeling the topological space of traffic data. However, existing methods fail to consider the spatial position information and only utilize little spatial neighborhood information. To tackle above limitation, we design a Graph Spatial–Temporal Position Recurrent Network (GSTPRN) architecture for traffic forecasting. We first construct a position graph convolution module based on self-attention and calculate the dependence strengths among the nodes to capture the spatial dependence relationship. Next, we develop approximate personalized propagation that extends the propagation range of spatial dimension information to obtain more spatial neighborhood information. Finally, we systematically integrate the position graph convolution, approximate personalized propagation and adaptive graph learning into a recurrent network (i.e. Gated Recurrent Units). Experimental evaluation on two benchmark traffic datasets demonstrates that GSTPRN is superior to the state-of-art methods.

Physics-integrated neural differentiable (PiNDiff) model for composites manufacturing

Various manufacturing technologies are being developed to improve the manufacturing of composites owing to their low weight and high performance. The mechanical properties of the composites depend on various variables and parameters of the manufacturing process, which are challenging, if not impossible, to determine and optimize experimentally. Traditional first-principle modeling approaches are not accessible due to the complex physics involved. A hybrid model that combines incomplete physics knowledge with available measurement data within a differentiable programming framework opens up new avenues to tackle the challenges. In this work, a physics-integrated neural differentiable (PiNDiff) model is developed, where the partially known physics is integrated into the recurrent network architecture to enable effective learning and generalization. The merit and potential of the proposed method have been demonstrated in modeling the curing process of thick thermoset composite laminates, whose governing physics is partially given. The proposed PiNDiff model shows the capability to learn unknown physics from the limited, indirect data and, meanwhile, can be used to infer unobserved variables and parameters. The performance of the PiNDiff model has been compared with two state-of-the-art (SOTA) black-box deep learning models, and its advantages over the purely data-driven models and first-principles physics-based models have been discussed in detail. The demonstrated PiNDiff strategy may provide a general strategy to model phenomena where physics is only partially known and sparse, indirect data are available.

Multi-Branch and Progressive Network for Low-Light Image Enhancement

Low-light images incur several complicated degradation factors such as poor brightness, low contrast, color degradation, and noise. Most previous deep learning-based approaches, however, only learn the mapping relationship of single channel between the input low-light images and the expected normal-light images, which is insufficient enough to deal with low-light images captured under uncertain imaging environment. Moreover, too deeper network architecture is not conducive to recover low-light images due to extremely low values in pixels. To surmount aforementioned issues, in this paper we propose a novel multi-branch and progressive network (MBPNet) for low-light image enhancement. To be more specific, the proposed MBPNet is comprised of four different branches which build the mapping relationship at different scales. The followed fusion is performed on the outputs obtained from four different branches for the final enhanced image. Furthermore, to better handle the difficulty of delivering structural information of low-light images with low values in pixels, a progressive enhancement strategy is applied in the proposed method, where four convolutional long short-term memory networks (LSTM) are embedded in four branches and an recurrent network architecture is developed to iteratively perform the enhancement process. In addition, a joint loss function consisting of the pixel loss, the multi-scale perceptual loss, the adversarial loss, the gradient loss, and the color loss is framed to optimize the model parameters. To evaluate the effectiveness of proposed MBPNet, three popularly used benchmark databases are used for both quantitative and qualitative assessments. The experimental results confirm that the proposed MBPNet obviously outperforms other state-of-the-art approaches in terms of quantitative and qualitative results. The code will be available at https://github.com/kbzhang0505/MBPNet.

Predicting transmission loss in underwater acoustics using convolutional recurrent autoencoder network.

Underwater noise transmission in the ocean environment is a complex physical phenomenon involving not only widely varying physical parameters and dynamical scales but also uncertainties in the ocean parameters. It is challenging to construct generalized physical models that can predict transmission loss in a broad range of situations. In this regard, we propose a convolutional recurrent autoencoder network (CRAN) architecture, which is a data-driven deep learning model for learning far-field acoustic propagation. Being data-driven, the CRAN model relies only on the quality of the data and is agnostic to how the data are obtained. The CRAN model can learn a reduced-dimensional representation of physical data and can predict the far-field acoustic signal transmission loss distribution in the ocean environment. We demonstrate the ability of the CRAN model to learn far-field transmission loss distribution in a two-dimensional ocean domain with depth-dependent sources. Results show that the CRAN can learn the essential physical elements of acoustic signal transmission loss generated due to geometric spreading, refraction, and reflection from the ocean surface and bottom. Such ability of the CRAN to learn complex ocean acoustics transmission has the potential for real-time far-field underwater noise prediction for marine vessel decision-making and online control.

Oscillatory Fourier Neural Network: A Compact and Efficient Architecture for Sequential Processing

Tremendous progress has been made in sequential processing with the recent advances in recurrent neural networks. However, recurrent architectures face the challenge of exploding/vanishing gradients during training, and require significant computational resources to execute back-propagation through time. Moreover, large models are typically needed for executing complex sequential tasks. To address these challenges, we propose a novel neuron model that has cosine activation with a time varying component for sequential processing. The proposed neuron provides an efficient building block for projecting sequential inputs into spectral domain, which helps to retain long-term dependencies with minimal extra model parameters and computation. A new type of recurrent network architecture, named Oscillatory Fourier Neural Network, based on the proposed neuron is presented and applied to various types of sequential tasks. We demonstrate that recurrent neural network with the proposed neuron model is mathematically equivalent to a simplified form of discrete Fourier transform applied onto periodical activation. In particular, the computationally intensive back-propagation through time in training is eliminated, leading to faster training while achieving the state of the art inference accuracy in a diverse group of sequential tasks. For instance, applying the proposed model to sentiment analysis on IMDB review dataset reaches 89.4% test accuracy within 5 epochs, accompanied by over 35x reduction in the model size compared to LSTM. The proposed novel RNN architecture is well poised for intelligent sequential processing in resource constrained hardware.

Recurrent neural networks for water quality assessment in complex coastal lagoon environments: A case study on the Venice Lagoon

Eutrophication represents an important ecological and environmental issue in coastal lagoons. This paper presents an extensive study of recurrent cell and network architectures to model eutrophication processes in the Venice lagoon, a very complex and fragile ecosystem that has been strongly altered by anthropic activities over years. Experimental results showed that recurrent models outperformed Random Forests (RF) significantly on two datasets, performing similarly to CNNs on one of the datasets, while outperforming CNNs on the other one. Additionally, the transferability potential of the trained models was investigated. The empirical analysis has shown that recurrent neural network models with lower computational complexity provide the highest eutrophication prediction accuracy when their trained models were tested on a new dataset. Designed models represent effective tools for early-warning eutrophication prediction that can support the implementation of relevant EU acquis (EU Marine Strategy and Water Framework Directives) and achievement of their environmental targets.

SoK] The Great GAN Bake Off, An Extensive Systematic Evaluation of Generative Adversarial Network Architectures for Time Series Synthesis

There is no standard approach to compare the success ofdifferent neural network architectures utilized for time seriessynthesis. This hinders the evaluation and decision process,as to which architecture should be leveraged for an unknowndata set. We propose a combination of metrics, which empiri-cally evaluate the performance of neural network architecturestrained for time series synthesis. With these measurementswe are able to account for temporal correlations, spatial cor-relations and mode collapse issues within the generated timeseries.We further investigate the interaction of different genera-tor and discriminator architectures between each other. Theconsidered architectures include recurrent neural networks,temporal convolutional networks and transformer-based net-works. So far, the application of transformer-based models islimited for time series synthesis. Hence, we propose a newtransformer-based architecture, which is able to synthesisetime series. We evaluate the proposed architectures and theircombinations in over 500 experiments, amounting to over2500 computing hours. We provide results for four data sets,one univariate and three multivariate. The data sets vary withregard to length, as well as patterns in temporal and spatialcorrelations.We use our metrics to compare the performance of genera-tive adversarial network architectures for time series synthesis.To verify our findings we utilize quantitative and qualitativeevaluations. Our results indicate that temporal convolutionalnetworks currently outperform recurrent neural network andtransformer based approaches with regard to fidelity and flex-ibility of the generated time series data. Temporal convolu-tional network architecture are the most stable architecture fora mode collapse prone data set. The performance of the trans-former models strongly depends on the data set characteristics,it struggled to synthesise data sets with high temporal andspatial correlations. Discriminators with recurrent networkarchitectures suffer from vanishing gradients. We also show,that the performance of the generative adversarial networksdepends more on the discriminator rather than the generator.