• We propose WEViT, the first framework that integrates Neural Architecture Search (NAS) with vision transformers for Weakly Supervised Semantic Segmentation (WSSS). • A one-shot weight-sharing strategy is employed to efficiently train an overparameterized supernet, enabling fast and effective transformer architecture search. • The framework utilizes multi-class tokens to extract class-specific attention from transformers, enhancing the precision of localization maps. • We introduce a novel Refinement Patch Affinity strategy that suppresses background noise and improves class focus in multi-class images. • A regularization loss function is designed to promote class-discriminative attention, with experiments highlighting the importance of transformer layer selection for maximizing segmentation accuracy. Weakly Supervised Semantic Segmentation (WSSS) is a challenging task in computer vision, as it relies on limited supervision to generate precise object localization maps, often using Class Activation Maps (CAMs). Traditional methods struggle with balancing localization accuracy and scalability due to their reliance on fixed network architectures and handcrafted strategies. Neural Architecture Search (NAS), despite its proven success in optimizing network designs across tasks, has not yet been explored in WSSS due to the need for efficient weight sharing. To address these limitations, we propose WEViT, a novel framework that integrates NAS with transformers to optimize network architectures and generate accurate and class-specific object localization maps for WSSS. Our approach leverages the weight entanglement strategy, enabling the supernet to train multiple subnets simultaneously while ensuring high-quality weight inheritance. This eliminates the need for retraining subnets from scratch, significantly reducing computational cost. The best-performing architecture, obtained through the evolutionary algorithm, is then utilized to extract attention weights from transformer heads. These weights are further refined using a Refinement Patch Affinity strategy, effectively removing background noise and enhancing focus on relevant classes in multi-class images. We also incorporate a regularization loss function during training to enhance the generation of class-discriminative localization maps, with experiments highlighting the critical role of transformer layer selection in this process. WEViT achieves state-of-the-art performance on PASCAL VOC 2012 and MS COCO, demonstrating the efficacy of applying NAS to WSSS for the first time and paving the way for scalable, efficient, and accurate segmentation solutions.

DUO-Net is proposed, a unified multi-task learning framework for joint 2D object detection and depth estimation. The architecture employs a shared ResNet-based backbone with attention modules and task-specific heads to simultaneously perform bounding box localisation and dense depth prediction. A two-stage training method is adopted to sequentially pretrain each task and subsequently refine them through joint learning, enhancing both features and convergence reliability. To address task imbalance and noisy supervision, we incorporate uncertainty-aware loss weighting, enabling the model to dynamically adjust task contributions during training. Evaluated on the KITTI and JRDB datasets, DUO-Net demonstrates robust performance across both tasks while maintaining efficiency and scalability.

ABSTRACT Object detection is a critical aspect of computer vision (CV) applications, especially within autonomous driving systems (AVs), where it is fundamental to ensuring safety and reducing traffic accidents. Recent advancements in computational resources have enabled the widespread adoption of Deep Learning (DL) techniques, significantly enhancing the efficiency and accuracy of object detection tasks. However, the technology for autonomous driving has yet to reach a level of maturity that guarantees consistent performance, reliability, and safety, with several challenges remaining unresolved. This study specifically focuses on 2D image‐based object detection methods, which offer several advantages over other modalities, such as cost‐effectiveness and the ability to capture visual features like colour and texture that are not detectable by LiDAR. We provide a comprehensive survey of DL‐based strategies for detecting vehicles and pedestrians using 2D images, analysing both one‐stage and two‐stage detection frameworks. Additionally, we review the most commonly used publicly available datasets in autonomous driving research and highlight their relevance to 2D detection tasks. The paper concludes by discussing the current challenges in this domain and proposing potential future directions, aiming to bridge the gap between the capabilities of 2D image‐based models and the requirements of real‐world autonomous driving applications. Comparative tables are included to facilitate a clear understanding of the different approaches and datasets.

Abstract Food recommendation systems have become increasingly popular due to the proliferation of online food service websites. Accordingly, the ratings assigned by users are one of the most important resources in these systems. However, users generally express their opinions about a few foods, which results in data sparsity. Furthermore, food recommendation is a health-critical task, as recommending unhealthy foods to users may threaten their health. In this paper, we developed a novel rating profile expansion approach for food recommenders that considers both health and reliability measures. This approach enhances the efficiency of the user’s rating profile by including healthy and reliable virtual ratings. Specifically, we introduce a probabilistic rating profile evaluation technique to determine whether a profile needs to be expanded. Then, those profiles with an insufficient number of ratings are automatically expanded by adding virtual ratings obtained using the opinions of users who belong to the target user’s community. For this purpose, the users are grouped using a novel time-aware community detection algorithm based on their preferences. Moreover, a health-aware reliability measure is proposed so that only the most reliable virtual ratings are accounted for in the target user’s rating profile expansion. Therefore, the developed approach not only mitigates issues stemming from sparse data in food recommendation systems but also makes them more effective in recommending healthy foods to users. Experiments conducted on two publicly available real-world datasets demonstrated that the developed system is superior to other baseline models.

Abstract Integrating machine learning techniques into medical diagnostic systems holds great promise for enhancing disease identification and treatment. Among the various options for training such systems, the extreme learning machine (ELM) stands out due to its rapid learning capability and computational efficiency. However, the random selection of input weights and hidden neuron biases in the ELM can lead to suboptimal performance. To address this issue, our study introduces a novel approach called modified Harris hawks optimizer (MHHO) to optimize these parameters in ELM for medical classification tasks. By applying the MHHO‐based method to seven medical datasets, our experimental results demonstrate its superiority over seven other evolutionary‐based ELM trainer models. The findings strongly suggest that the MHHO approach can serve as a valuable tool for enhancing the performance of ELM in medical diagnosis.

Neural Architecture Search (NAS) has significantly improved the accuracy of image classification and segmentation. However, these methods concentrate on finding segmentation structures for natural or medical applications. In this study, we introduce a NAS approach based on gradient optimization to identify ideal cell designs for road segmentation. To the best of our knowledge, this work represents the first application of gradient-based NAS to road extraction. Taking insight from the U-Net model and its successful variations in different image segmentation tasks, we propose NAS-enhanced U-Net, illustrated by an equal number of cells in both encoder and decoder levels. While cross-entropy combined with dice loss is commonly used in many segmentation tasks, road extraction brings up a unique challenge due to class imbalance. To address this, we introduce a combination of loss function. This function merges cross-entropy with weighted Dice loss, focusing on elevating the importance of the road class by assigning it a weight (⍵), while background Dice values are disregarded. The results indicate that the optimal weight for the proposed model equals 2. Additionally, our work challenges the assumption that increased model parameters or depth inherently leads to improved performance. Therefore, we establish search spaces 2,3,4,5,6,7 and 8 to automatically choose the optimal depth for model. We present promising segmentation results for our proposed method, achieved without any pretraining on the Massachusetts road dataset. Furthermore, these results are compared with those of 14 models categorized into four groups: U-Net, Segnet, FCN8, and Nas-U-Net.

Precise image segmentation is one of the dominant factors in disease diagnosis. A typical application is the segmentation of breast ultrasound images, allowing radiologists to suggest what to do next. After emerging deep learning technology especially convolutional neural networks (CNNs), the image segmentation model achieved state-of-the-art performance in various medical applications such as cancer detection and classification, lung node segmentation, cell segmentation and so on. However, despite these successes, a big question arises: to what extent is the model certain about the predicted result? Generally, most deep learning models focus on high accuracy but not on uncertainty of predicted results, which is not enough to make a critical real-life decision such as a disease diagnosis, where a wrong decision can be life-threatening. Hence for making a crucial decision, it is essential that the predicted result will provide not only accuracy but also estimate model uncertainty. Our contribution to this research is to build a system that predicts pixel-wise semantic segmentation and provides uncertainty estimation of the predicted results. It is achieved by adding a dropout layer during training and using Monte Carlo dropout in inference. We evaluate our model with the breast ultrasound image dataset (BUSI) and compare the results with a few other state-of-the-art methods where our method outperforms others in terms of IoU.

Food recommendation systems aim to provide recommendations according to a user’s diet, recipes, and preferences. These systems are deemed useful for assisting users in changing their eating habits towards a healthy diet that aligns with their preferences. Most previous food recommendation systems do not consider the health and nutrition of foods, which restricts their ability to generate healthy recommendations. This paper develops a novel health-aware food recommendation system that explicitly accounts for food ingredients, food categories, and the factor of time, predicting the user’s preference through time-aware collaborative filtering and a food ingredient content-based model. Based on the user's predicted preferences and the health factor of each food, our model provides final recommendations to the target user. The performance of our model was compared to several state-of-the-art recommender systems in terms of five distinct metrics: Precision, Recall, F1, AUC, and NDCG. Experimental analysis of datasets extracted from the websites Allrecipes.com and Food.com demonstrated that our proposed food recommender system performs well compared to previous food recommendation models.

Wind power forecasting is very crucial for power system planning and scheduling. Deep neural networks (DNNs) are widely used in forecasting applications due to their exceptional performance. However, the DNNs’ architectural configuration has a significant impact on their performance, and the selection of proper hyper-parameters determines the success or failure of these models. Therefore, one of the challenging issues in DNNs is how to assess their hyper-parameter values effectively. Most of the previous researches in the literature have tuned the DNNs’ hyper-parameters manually, which is a weak and time-consuming task. Using optimization/evolutionary algorithms is an effective way to obtain the optimal values of DNNs’ hyper-parameters automatically. In this article, we propose a novel evolutionary algorithm that is based on the grasshopper optimization algorithm (GOA) improved by adding two evolutionary operators, opposition-based learning and chaos theory, to the optimization process. Overall, a novel probabilistic wind power forecasting model named neural GOA deep auto-regressive (NGOA-DeepAr) is proposed based on an auto-regressive recurrent neural network in which the proposed evolutionary algorithm has optimized its hyper-parameters. The performance of the proposed NGOA-DeepAr model is tested on two different datasets: One is the publicly available GEFCom-2014 dataset and the other is the Australian Energy Market Operator dataset. The prediction interval coverage probability and pinball loss for the two datasets are$[0.902, 0.320]$and$[0.933, 1.4885]$, respectively. According to the experimental findings, our proposed NGOA-DeepAr is much faster in learning and outperforms the benchmark DNNs and the other neuroevolutionary models.

Driving simulators are utilized for many applications including basic driver training, human factor studies, human-machine interaction, and vehicle prototyping in automobile industries. The main purpose of using driving simulator is to provide realistic driving experience. Since simulator platforms have physical limitations, Motion Cueing Algorithms (MCAs) are used to generate driving sensation for the simulator user while considering the simulator's physical and dynamical constraints. When using a model predictive control (MPC)-based MCA, the principle of MPC is leveraged to predict an optimized future behavior of the simulator where a series of control actions is developed across a defined future horizon using the explicitly specified process model. Corresponding to the pre-positioning or time-varying reference MPC, it is crucial to predict the future vehicle motion signals for the simulator accurately. The existing methods for predicting vehicle motion signals do not excel in predicting time-series of a long sequence due to the missing feedback loop or limited memory size. To address this issue, the Long Short-Term Memory (LSTM) model is developed to predict motion signals using Python. The performance of LSTM is compared with those from different traditional methods using several measurements criteria, which include the root mean squared error (RMSE), mean absolute error (MAE), and Pearson’s correlation coefficient (r). The results indicate that LSTM outperforms RNN by producing more accurate motion allowing the MCA to deliver realistic motion sensations, the LSTM model can be employed in a wide range of applications including autonomous vehicles trajectory prediction, and other prediction problems.

Abstract Cardiovascular diseases are the leading cause of death in recent decades, which are increasing due to changes in people's lifestyles. Their treatment has high costs and a long treatment process. Therefore, predicting such diseases can provide care, and prevention services and treatment programs can be very useful to increase the quality of life and reduce the cost of treatment and the risk of death for patients. Various artificial neural network (ANN) techniques and machine learning (ML) algorithms can be used as efficient and reliable methods to automatically analyze and detect the hidden patterns of patient medical records data collected through medical examinations related to cardiovascular diseases. In this paper, the multilayer perceptron (MLP) neural network is employed as a supervised learning approach to detect cardiovascular diseases. Moreover, we propose a modified version of moth‐flame optimization algorithm named as MMFO which is used to achieve the optimal values of weights and biases in the MLP to speed‐up the training process and provide more accurate predictions. The effectiveness of the proposed method is assessed according to performing extensive experiments on three cardiovascular disease datasets from the UCI repository, and its performance is compared with different state‐of‐the‐art classification approaches. The results reveal that the proposed method performs better than other models in terms of all medical datasets.

Accurate prediction of solar energy is an important issue for photovoltaic power plants to enable early participation in energy auction industries and cost-effective resource planning. This article introduces a new deep learning-based multistep ahead approach to improve the forecasting performance of global horizontal irradiance (GHI). A deep convolutional long short-term memory is used to extract optimal features for accurate prediction of the GHI. The performance of such deep neural networks directly depends on their architectures. To deal with this problem, a swarm evolutionary optimization method, called the sine-cosine algorithm, is applied and advanced to automatically optimize the network architecture. A three-phase modification model is proposed to increase the diversity of population and avoid premature convergence in the optimization mechanism. The performance of the proposed method is investigated using three datasets collected from three solar stations in the east of the United States. The experimental results demonstrate the superiority of the proposed method in comparison to other forecasting models.

Recommender systems are known as intelligent systems which have many applications in enormous domains such as social networks, e-commerce services, and online shopping. Deep neural networks have shown significant improvement in the performance of recommender systems by learning the latent features of users/items based on input data. However, it is a challenging issue to how to apply deep neural networks on different resources and how to integrate their results. In this regard, we propose a recommender system in this paper based on deep sparse autoencoder and particle swarm optimization. In particular, a deep sparse autoencoder is utilized to learn latent features based on the ratings matrix, trust relationships, and tag information. Then, particle swarm optimization is used to find the optimal weights of these latent features in calculating unknown ratings. Experiments on two datasets show the superiority of the proposed method in comparison with state of the art recommender algorithms.

Deep learning-based collaborative filtering methods are studied in recommendation systems as efficient feature mapping techniques. The aim of these methods is to project the users and items to a common representation space and obtain their latent features. Although these methods have been widely used in the literature, they suffer from the limited expressiveness of Dot product function. In other words, Dot product cannot describe different impacts of various latent factors. To solve this issue, we propose a novel recommender system named Deep-MSR which exploits the multilayer perceptron (MLP) neural network and stacked auto-encoder network (SAN) to extract item latent factors and user latent factors from user-item interaction matrix. The obtained latent factors are used in the proposed rating prediction module which integrates user preferences and item features in the recommendation process. Our experiments on two well-known datasets show that our method can outperform the competitive baseline recommendation methods.

The problem of electricity load forecasting has emerged as an essential topic for power systems and electricity markets seeking to minimize costs. However, this topic has a high level of complexity. Over the past few years, convolutional neural networks (CNNs) have been used to solve several complex deep learning challenges, making substantial progress in some fields and contributing to state of the art performances. Nevertheless, CNN architecture design remains a challenging problem. Moreover, designing an optimal architecture for CNNs leads to improve their performance in the prediction process. This article proposes an effective approach for the electricity load forecasting problem using a deep neuroevolution algorithm to automatically design the CNN structures using a novel modified evolutionary algorithm called enhanced grey wolf optimizer (EGWO). The architecture of CNNs and its hyperparameters are optimized by the novel discrete EGWO algorithm for enhancing its load forecasting accuracy. The proposed method is evaluated on real time data obtained from datasets of Australian Energy Market Operator in the year 2018. The simulation results demonstrated that the proposed method outperforms other compared forecasting algorithms based on different evaluation metrics.

The Motion Cueing Algorithm (MCA) oversees regenerating the motion feeling of the real vehicle for the simulation-based motion platform (SBMP) within the physical limitations. Model Predictive Control (MPC) is recently employed as an MCA, which is called MPC-based MCA due to the consideration of the plant's boundaries in finding the optimal input signal. The computational load of the MPC directly relates to the control horizon and prediction horizon of the MPC. In this paper, a new optimisation method using butterfly optimisation algorithm is developed to find the optimal control horizon and prediction horizon of MPC-based MCA. The proposed method reduces the time of the tuning process of the MPC-based MCA, which is usually carried out via trial-and-error and genetic algorithm methods. Also, the trial-and-error method increases the motion sensation error and insufficient usage of the SBMP. The model is validated using MATLAB simulation environment, and the outcomes show that the developed butterfly optimisation algorithm will lead better motion sensation with less wrong motion signals and low computational burden compared with the trial-and-error and genetic algorithm method.

End to end learning for autonomous navigation and driving has become a growing research trend in both industry and academia in recent years. Its promise is in treating the whole driving pipeline as the development of a deep neural network (DNN). Its Achilles' heel is access to thousands of images required for training of the DNN. This paper comprehensively investigates the applicability of the deep transfer learning for the specific task of end to end learning of autonomous navigation. Five state of the art DNNs including ResNet, AlexNet, and Densenet are applied here for extracting features from images taken by the front-facing camera of a mobile robot. Extracted features have different information values as DNNs have different architectures and learning capabilities. These features are then processed by a multilayer fully connected neural network to estimate the robot angular velocity. Obtained results for different DNNs indicate that the transfer learning-based models show a promising performance for accurately estimating the angular velocity purely using visual information. According to obtained results, AlexNet-base model outperforms others in terms of the estimation accuracy and the performance consistency.

Communication delays are known to create stability and performance issues in multilateral teleoperation systems. Multilateral teleoperation configurations usually include more than two communication channels, which can become problematic for robot control when limitations in network bandwidth results in delays and uncertainties in data transmission routes. This study develops a sliding surface based on the synchronization errors characterized between each sides of the considered multilateral teleoperation system. Here, two slave robots receive commands from the master system to cooperatively execute the desired teleoperation task in the remote, shared workspace. Lyapunov stability analysis approach guarantees the performance of the proposed controller. Moreover, the effectiveness of the controller is experimentally evaluated through a real-world Internet-based double-slave teleoperation system.

The field of neuroevolution has received great attention in recent years due to its promising capability for developing well-performing models. It has been applied to many real-world problems ranging from medical diagnosis to autonomous robots. The choice of the evolutionary algorithm (EA) has a huge impact on the neuroevolution overall performance. Despite recent progress in the field, it is not clear what the best choice of EA is. The problem becomes more severe considering a dozen of EAs available for neuroevolution applications. In this paper, six state of the art EAs are applied for the task of autonomous robot navigation. These EAs are MultiVerse optimizer (MVO), moth-flame optimization (MFO), particle swarm optimization (PSO), cuckoo search (CS), Grey wolf optimizer (GWO) and bat algorithm. MLP networks are trained using these six evolutionary algorithms to solve the classification task related to the autonomous robot navigation. Comprehensive experiments are conducted using three datasets and obtained results are visually and statistically compared. To the best knowledge of the authors, comparison among the aforementioned algorithms has not been considered in the literature. It is found that neuroevolution methods perform well for the task of autonomous robot navigation. Amongst investigated EAs, MVOtrained achieves the highest and most consistent performance metrics.

Cardiovascular diseases are one of the main causes of death among individuals over the last decade. Early diagnosis and recognizing of warning signs of this disease facilitate medical treatment for patients. Angiography is considered a reliable tool to diagnose coronary artery disease (CAD), however, it has some demerits such as complications and costs. Data mining techniques are considered as reliable and powerful tools for early diagnosis of diseases and are widely used in the medicine filed for recent years. In this paper, we use these techniques for early detection of CAD by applying them on a well-known CAD dataset named Z-Alizadeh sani. Thus, an effective nature-inspired optimization algorithm named Multi-verse optimizer (MVO) based on Multilayer perceptron (MLP) training as well as nine states of the art supervised learning techniques are employed for CAD prediction. As this dataset has 54 features, before applying the supervised learning algorithms, we used a feature selection method to identify the most effective features. This procedure enhances the prediction capability of the utilized algorithms. The classification rates of all algorithms are compared with each other using the most usable evaluation metrics including accuracy and area under the curve. Eventually, the experimental results show that the most appropriate model to classify CAD patients is the MLP model trained by MVO among all other nine supervised learning methods.

Imitation learning is an efficient paradigm for teaching and controlling intelligent autonomous cars. Obtaining a set of suitable demonstrations to learn an end-to-end policy from raw pixels is a challenging task in imitation learning problems. Deep neural networks have recently shown outstanding results in learning from raw high dimensional data for solving a wide range of real-world applications. The success of deep neural networks depends on finding suitable hyperparameters for constructing network architecture. Besides, designing hand-crafted deep architectures is not an efficient way for achieving the best performance. To address this issue, this paper performs a neuro-evolution method based on genetic algorithm for finding the optimal deep neural networks architecture in terms of hyperparameters. The experimental results show the effectiveness of the proposed approach for training an autonomous vehicle.

Deep neural networks have proven themselves as a strong approach in image classification and object detection with high accuracy. However, they are computationally demanding and the trained networks contain millions of active parameters and connections. Two recent trends of having deeper and dense architectures and the deployment of trained networks on resource-constrained devices such as smart phones and portable tablets bring new challenges. Instead of deploying an ensemble of smaller networks, we propose a pruning methodology on a trained network so that a smaller version of a fully trained network has the same and even better accuracy in comparison to the original one. We achieve two objectives with the pruning scheme. First, we have a smaller network with a better accuracy level, and we make the trained model avoids overfitting. Accordingly, an evolutionary based framework including three steps is defined to perform further tuning on trained deep network using dropping nodes and connections. This study shows that implementing genetic algorithm, after preprocessing and training stages, not only results in partially connected networks, but also increases performance and reduces overfitting specially when the depth and width of fully connected networks are investigated in small datasets.