Software vulnerability detection has traditionally relied on classical machine learning (ML) models with handcrafted features, which are lightweight but limited in performance. Transformer-based Large Language Models (LLMs) trained on source code provide deeper semantic understanding and stronger generalization. This work presents a direct comparison between ML- and LLM-based approaches across binary and multi-label tasks. We benchmark Random Forest, SVM, and LSTM against LoRA-finetuned LLaMA 3.1 models on both synthetic and real-world datasets. Our LLaMA 3.18 B model achieves $87 \%$ macro $F 1$ in binary classification and $82 \%$ in multi-label classification on real-world data-substantially outperforming ML and DL baselines. Despite higher computational demands, parameter-efficient fine-tuning with Unsloth and LoRA makes LLMs viable for practical deployment. These findings highlight the superiority of LLMs in security-critical code analysis and establish strong baselines for future research.

Machine learning (ML) techniques are popular in many parameter estimation tasks; however, they face challenges in the real-world deployment due to the lack of robustness to errors. ML estimators are not able to ascertain performance in the presence of noise, variations in the data distribution, and anomalies in the test samples. This work proposes a novel framework, surrogate-based physical error correction (SPEC), which addresses the unmet need for measurement reliability estimation and self-correction under process data uncertainty, by bringing together physics- and network-based optimization. The workings of SPEC are demonstrated using the paradigm of gas parameter estimation in the laser absorption spectroscopy (LAS). It operates in two modes, estimation and correction. During estimation, SPEC provides an initial state estimate, with estimation reliability being assessed by the physics-driven anomaly detection (PAD) module, which uses a hybrid error, combining a nondifferentiable reconstruction error, calculated through an ensemble network, and a differentiable feasibility error. When an estimate is flagged as unreliable, the correction mode is enabled. This network-based optimization algorithm delivers efficient and robust state correction by using a greedy ensemble search. SPEC's performance is evaluated in a variety of experiments including outside-of-distribution and noisy data. Moreover, it offers reconfigurability through PAD configuration modification, eliminating the need for ML estimator retraining.

Electrical Impedance Tomography based sensors have emerged as a promising approach in tactile sensing, offering notable advantages such as affordability, portability, and low power consumption. However, the inherently ill-posed nature of the inverse problem often results in reconstruction errors, impacting on the accuracy of tactile information retrieval. In this work, an effective deep learning approach for tactile sensing is proposed, leveraging the concept of learnable masks, incorporated within a Convolutional Neural Network. The learnable masks support the selection of the most informative feature subsets from the associated voltage inputs, enabling the network to reconstruct conductivity distributions precisely. The proposed approach exhibited outstanding performance in image reconstruction, achieving a mean square error of 0.000041, a structural similarity index of 98.28, and a peak signal-to-noise ratio of 42.35 dB.

In autonomous driving, environment perception is an important step in understanding the driving scene. Objects in images captured through a vehicle camera can be detected and classified using semantic segmentation and depth estimation methods. Both these tasks are closely related to each other and this association helps in building a multi-task neural network where a single network is used to generate both views from a given monocular image. This approach gives the flexibility to include multiple related tasks in a single network. It helps reduce multiple independent networks and improve the performance of all related tasks. The main aim of our research presented in this paper is to build a multi-task deep learning network for simultaneous semantic segmentation and depth estimation from monocular images. Two decoder-focused U-N et-based multi-task networks that use a pre-trained Resnet-50 and DenseNet-121 which shared encoder and task-specific decoder networks with Attention Mechanisms are considered. We also employed multi-task optimization strategies such as equal weighting and dynamic weight averaging during the training of the models. The corresponding models' performance is evaluated using mean IoU for semantic segmentation and Root Mean Square Error for depth estimation. From our experiments, we found that the performance of these multi-task networks is on par with the corresponding single-task networks.

Early and accurate detection of brain tumors is very important to save the patient's life. Brain tumors are generally diagnosed manually by a radiologist by analyzing the patient”s brain MRI scans which is a time-consuming process. This led to our study of this research area for finding out a solution to automate the diagnosis to increase its speed and accuracy. In this study, we investigate the use of Residual Network deep learning architecture to diagnose and segment brain tumors. We proposed a two-step method involving a tumor detection stage, using ResNet50 architecture, and a tumor area segmentation stage using ResU-Net architecture. We adopt transfer learning on pre-trained models to help get the best performance out of the approach, as well as data augmentation to lessen the effect of data population imbalance and hyperparameter optimization to get the best set of training parameter values. Using a publicly available dataset as a testbed we show that our approach achieves 84.3 % performance outperforming the state-of-the-art using U-Net by 2% using the Dice Coefficient metric.

Automatic car damage detection and assessment are very useful in alleviating the burden of manual inspection associated with car insurance claims. This will help filter out any frivolous claims that can take up time and money to process. This problem falls into the image classification category and there has been significant progress in this field using deep learning. However, deep learning models require a large number of images for training and oftentimes this is hampered because of the lack of datasets of suitable images. This research investigates data augmentation techniques using Generative Adversarial Networks to increase the size and improve the class balance of a dataset used for training deep learning models for car damage detection and classification. We compare the performance of such an approach with one that uses a conventional data augmentation technique and with another that does not use any data augmentation. Our experiment shows that this approach has a significant improvement compared to another that does not use data augmentation and has a slight improvement compared to one that uses conventional data augmentation.

When deploying machine learning estimators in science and engineering (SAE) domains, it is critical to avoid failed estimations that can have disastrous consequences, e.g., in aero engine design. This work focuses on detecting and correcting failed state estimations before adopting them in SAE inverse problems, by utilizing simulations and performance metrics guided by physical laws. We suggest to flag a machine learning estimation when its physical model error exceeds a feasible threshold, and propose a novel approach, GEESE, to correct it through optimization, aiming at delivering both low error and high efficiency. The key designs of GEESE include (1) a hybrid surrogate error model to provide fast error estimations to reduce simulation cost and to enable gradient based backpropagation of error feedback, and (2) two generative models to approximate the probability distributions of the candidate states for simulating the exploitation and exploration behaviours. All three models are constructed as neural networks. GEESE is tested on three real-world SAE inverse problems and compared to a number of state-of-the-art optimization/search approaches. Results show that it fails the least number of times in terms of finding a feasible state correction, and requires physical evaluations less frequently in general.

Marine debris poses a significant ecological challenge, necessitating advanced methods for its accurate detection and segmentation. Deep learning enables advanced remote sensing capabilities for earth observation, however, its on-board deployment is hindered by limitations in resource availability. In this paper, octonion neural networks (ONNs) are proposed for developing a parameter-efficient solution to address the problem of marine debris segmentation. ONNs extend the capabilities of real-valued networks by incorporating octonions, an eight-dimensional hypercomplex number system. By harnessing the power of octonions, such as their ability to capture higher-dimensional relationships and extract robust feature representations, enhanced segmentation accuracy can be achieved. The proposed ONN model is evaluated on the MARIDA dataset, a comprehensive benchmark for marine debris segmentation. The results demonstrate that the proposed approach outperforms the state of the art, achieving remarkable improvements of 9.9% and 7.6% in terms of the Intersection over Union (IoU) and F1 metrics, respectively. Moreover, the ONN approach delivers performance similar to that of the real-valued architecture, while utilizing 1/13 of the network parameters.

Early detection of retinal diseases is one of the most important means of preventing partial or permanent blindness in patients. In this research, a novel multi-label classification system is proposed for the detection of multiple retinal diseases, using fundus images collected from a variety of sources. First, a new multi-label retinal disease dataset, the MuReD dataset, is constructed, using a number of publicly available datasets for fundus disease classification. Next, a sequence of post-processing steps is applied to ensure the quality of the image data and the range of diseases, present in the dataset. For the first time in fundus multi-label disease classification, a transformer-based model optimized through extensive experimentation is used for image analysis and decision making. Numerous experiments are performed to optimize the configuration of the proposed system. It is shown that the approach performs better than state-of-the-art works on the same task by 7.9% and 8.1% in terms of AUC score for disease detection and disease classification, respectively. The obtained results further support the potential applications of transformer-based architectures in the medical imaging field.

Weather forecasting is a critical research area that may lead to serious consequences, if not done accurately. In order to manage the impact of adverse weather effects short time forecasting is employed, since it can be highly accurate to predict a short time, e.g., hours, into the future. In this work, we tackle the challenge of short time weather forecasting using a novel approach based on a modified UNet based model. Specifically, all convolution-based building blocks were replaced by 3D shifted window transformers in both encoder and decoder branches. Shifted window transformers greatly reduce computational complexity requirements, a major constraint of self-attention, without sacrificing performance. To support the pre-training of the transformer-based model a carefully crafted augmentation scheme was proposed. The model was tested on the IEEE Big Data Weather4cast Competition data, which requires the prediction of 8 hours ahead frames (4 per hour) from an hourly weather product sequence. We show the importance of including other weather products in encouraging spatial generalization, while this may not be optimal for temporal generalization. The model demonstrates highly competitive performance on both the validation and test datasets. The code is available online at https://github.com/bojesomo/Weather4Cast2021-SwinUNet3D

Over the last decade, robotics has experienced a rapid increase in research related to human-robot interaction. Developments in artificial skin research can equip robots with tactile sensing in a similar manner to the human sense of touch. This capability will make human-robot communication more natural and safer since an important part of perception indeed relies on tactile sensing. Electrical impedance tomography (EIT)-based sensors are considered as a potentially promising alternative for tactile sensing. These sensors can reconstruct images of the conductivity variation, which appear as a response to the applied pressure. However, due to the ill-posedness of the EIT inverse problem, reconstructed images have low spatial resolution and object boundaries are not preserved. In this research, we explore the hypothesis that performing image segmentation in conjunction with preserving the object boundaries may increase the accuracy of a subsequent classification of the reconstructed images. We compare the quality of EIT images segmented by a splitting and merging segmentation algorithm, Morphological Active Contours without Edges, Random Walker and transfer learning. While the explored classical techniques appear to have predilection towards over-segmentation, the deep learning approach results to a remarkable improvement of approximately 118% in terms of the similarity index.

Freezing of Gait (FoG) is a motor symptom of Parkinson's disease (PD) that frequently occurs in the long-term sufferers of the disease. FoG may result to nursing home admission as it can lead to falls, and therefore, it impacts negatively on the quality of life. The focus of this study is the systematic evaluation of machine learning techniques in conjunction with varying size time windows and time/frequency domain feature sets in predicting a FoG event before its onset. In the experiments, the Daphnet FoG dataset is used to benchmark performance. This consists of accelerometer signals obtained from sensors mounted on the ankle, thigh and trunk of the PD patients. The dataset is annotated with instances of normal activity events, and FoG events. To predict the onset of FoG, the dataset is augmented with an additional class, termed ‘transition’, which relates to a manually defined period prior to the occurrence of a FoG episode. In this research, five machine learning models are used, namely, Random Forest, Extreme Gradient Boosting, Gradient Boosting, Support Vector Machines using Radial Basis Functions, and Neural Networks. Support Vector Machines with Radial Basis kernels provided the best performance achieving sensitivity values of 72.34%, 91.49%, 75.00%, and specificity values of 87.36%, 88.51% and 93.62%, for the FoG, transition and normal activity classes, respectively.

Vascular structures in the retina contain important information for the detection and analysis of ocular diseases, including age-related macular degeneration, diabetic retinopathy and glaucoma. Commonly used modalities in diagnosis of these diseases are fundus photography, scanning laser ophthalmoscope (SLO) and fluorescein angiography (FA). Typically, retinal vessel segmentation is carried out either manually or interactively, which makes it time consuming and prone to human errors. In this research, we propose a new multi-modal framework for vessel segmentation called ELEMENT (vEsseL sEgmentation using Machine lEarning and coNnecTivity). This framework consists of feature extraction and pixel-based classification using region growing and machine learning. The proposed features capture complementary evidence based on grey level and vessel connectivity properties. The latter information is seamlessly propagated through the pixels at the classification phase. ELEMENT reduces inconsistencies and speeds up the segmentation throughput. We analyze and compare the performance of the proposed approach against state-of-the-art vessel segmentation algorithms in three major groups of experiments, for each of the ocular modalities. Our method produced higher overall performance, with an overall accuracy of 97.40%, compared to 25 of the 26 state-of-the-art approaches, including six works based on deep learning, evaluated on the widely known DRIVE fundus image dataset. In the case of the STARE, CHASE-DB, VAMPIRE FA, IOSTAR SLO and RC-SLO datasets, the proposed framework outperformed all of the state-of-the-art methods with accuracies of 98.27%, 97.78%, 98.34%, 98.04% and 98.35%, respectively.

This work proposes a variable neighbourhood search (FTS) that uses a fractal‐based local search primarily designed for images. Searching for specific content in images is posed as an optimisation problem, where evidence elements are expected to be present. Evidence elements improve the odds of finding the desired content and are closely associated to it in terms of spatial location. The proposed local search algorithm follows the fashion of a chain of triangles that engulf each other and grow indefinitely in a fractal fashion, while their orientation varies in each iteration. The authors carried out an extensive set of experiments, which confirmed that FTS outperforms state‐of‐the‐art metaheuristics. On average, FTS was able to locate content faster, visiting less incorrect image locations. In the first group of experiments, FTS was faster in seven out of nine cases, being >8% faster on average, when compared to the second best search method. In the second group, FTS was faster in six out of seven cases, and it was >22% faster on average when compared to the approach ranked second best. FTS tends to outperform other metaheuristics substantially as the size of the image increases.

This research is concerned with the development of a simulation framework for patient-specific image-based blood flow studies in the context of computing physiologically realistic flow patterns, which could provide accurate information for individual patients taking into account the specific characteristics of their Cardio Vascular System (CVS). The underlying principle is based on the application of the universal methods of Computational Fluid Dynamics (CFD) for the solution of the flow-governing Navier-Stokes equations, defined over the blood volume within the 3D vessel lumen of coronary arteries reconstructed from medical Computed Tomography (CT) images. In this work, we introduce the fundamental stages of our 0D modeling approach, which allows for the representation of a complex 3D artery lumen geometry through a simple lumped parameter volume. The framework is exemplified through the paradigm of a single arterial branch. We present preliminary results of the computed pressure along the length of the branch.

In this paper, we present a novel two-step algorithm for segmentation of coronary arteries in computed tomography images based on the framework of active contours. In the proposed method, both global and local intensity information is utilized in the energy calculation. The global term is defined as a normalized cumulative distribution function, which contributes to the overall active contour energy in an adaptive fashion based on image histograms, to deform the active contour away from local stationary points. Possible outliers, such as kissing vessel artifacts, are removed in the postprocessing stage by a slice-by-slice correction scheme based on multiregion competition, where both arteries and kissing vessels are identified and tracked through the slices. The efficiency and the accuracy of the proposed technique are demonstrated on both synthetic and real datasets. The results on clinical datasets show that the method is able to extract the major branches of arteries with an average distance of 0.73 voxels to the manually delineated ground truth data. In the presence of kissing vessel artifacts, the outer surface of the entire coronary tree, extracted by the proposed algorithm, is smooth and contains fewer erroneous regions, originating in kissing vessel artifacts, as compared to the initial segmentation.

This paper investigates the capabilities of hierarchical fuzzy systems to approximate functions on discrete input spaces. First, it is shown that any function on a discrete space has an arbitrary separable hierarchical structure and can be naturally approximated by hierarchical fuzzy systems. As a by-product of this result, a discrete version of Kolmogorov's theorem is obtained; second, it is proven that any function on a discrete space can be approximated to any degree of accuracy by hierarchical fuzzy systems with any desired separable hierarchical structure. That is, functions on discrete spaces can be approximated more simply and flexibly than those on continuous spaces; third, a hierarchical fuzzy system identification method is proposed in which human knowledge and numerical data are combined for system construction and identification. Finally, the proposed method is applied to the market condition performance modeling problem in site selection decision support and shows the better performance in both accuracy and interpretability than the regression and neural network approaches. In additions, the reason and mechanism why hierarchical fuzzy systems outperform regression and neural networks in this type of application are analyzed.

Wearable human movement measurement systems are increasingly popular as a means of capturing human movement data in real-world situations. Previous work has attempted to estimate segment kinematics during walking from foot acceleration and angular velocity data. In this paper, we propose a novel neural network [GRNN with Auxiliary Similarity Information (GASI)] that estimates joint kinematics by taking account of proximity and gait trajectory slope information through adaptive weighting. Furthermore, multiple kernel bandwidth parameters are used that can adapt to the local data density. To demonstrate the value of the GASI algorithm, hip, knee, and ankle joint motions are estimated from acceleration and angular velocity data for the foot and shank, collected using commercially available wearable sensors. Reference hip, knee, and ankle kinematic data were obtained using externally mounted reflective markers and infrared cameras for subjects while they walked at different speeds. The results provide further evidence that a neural net approach to the estimation of joint kinematics is feasible and shows promise, but other practical issues must be addressed before this approach is mature enough for clinical implementation. Furthermore, they demonstrate the utility of the new GASI algorithm for making estimates from continuous periodic data that include noise and a significant level of variability.

This paper proposes a new nonparametric regression method, based on the combination of generalized regression neural networks (GRNNs), density-dependent multiple kernel bandwidths, and regularization. The presented model is generic and substitutes the very large number of bandwidths with a much smaller number of trainable weights that control the regression model. It depends on sets of extracted data density features which reflect the density properties and distribution irregularities of the training data sets. We provide an efficient initialization scheme and a second-order algorithm to train the model, as well as an overfitting control mechanism based on Bayesian regularization. Numerical results show that the proposed network manages to reduce significantly the computational demands of having individual bandwidths, while at the same time, provides competitive function approximation accuracy in relation to existing methods.

This paper proposes a novel algorithm for function approximation that extends the standard generalized regression neural network. Instead of a single bandwidth for all the kernels, we employ a multiple-bandwidth configuration. However, unlike previous works that use clustering of the training data for the reduction of the number of bandwidths, we propose a distinct scheme that manages a dramatic bandwidth reduction while preserving the required model complexity. In this scheme, the algorithm partitions the training patterns to groups, where all patterns within each group share the same bandwidth. Grouping relies on the analysis of the local nearest neighbor distance information around the patterns and the principal component analysis with fuzzy clustering. Furthermore, we use a hybrid optimization procedure combining a very efficient variant of the particle swarm optimizer and a quasi-Newton method for global optimization and locally optimal fine-tuning of the network bandwidths. Training is based on the minimization of a flexible adaptation of the leave-one-out validation error that enhances the network generalization. We test the proposed algorithm with real and synthetic datasets, and results show that it exhibits competitive regression performance compared to other techniques.

In this paper, we propose a novel approach to stereo correspondence based on the optimization of a continuous disparity surface defined parametrically using radial basis functions. Principal advantages over other methods include the use of constrained nonlinear programming to perform regularization as a hierarchical multiobjective optimization which differs from the standard weighted sum approach, so that regularization becomes more consistent with the notion of Pareto optimality. Furthermore, the optimization algorithm is capable of handling arbitrary constraints on the sought parameters, so that a variety of types of a priori scene information can be incorporated explicitly to the problem definition. To exemplify this we derive a new continuous unary formulation of the disparity gradient limit constraint and propose other types of potential constraints for a priori knowledge. Furthermore, the optimization employs a smoothness oriented regularization operator to preserve surface discontinuities, a flexible block decomposition approach of the disparity surface to allow parallelization and a correlation-based fitting with heuristics to initialize the parameters and avoid local optima effectively. Experiments with standard stereo imagery show that the method handles adequately the imposed constraints and produces surfaces with accurate level of elevation detail.

This paper proposes a novel optimization algorithm for image-space matching and three-dimensional space analysis, using an adapted scheme of evolutionary computation that employs the concept of symbiosis in a collective of homogeneous populations. It is applied to the automatic generation of disparity surfaces used for depth estimation in stereo vision. The global task of approximating the complete disparity surface is decomposed to a large number of smaller local problems, each solvable by a smaller processing unit. Coevolution is sustained in such a way as to counteract the arbitrary decomposition of the original super-problem, so that the local evolutions of all the subproblems become interlocked. This, in the long run, provides a consistent global solution, and it does so via an asynchronous and massively parallel architecture. The entire surface is partitioned to a set of adjoining patches represented by distinct species or populations, with phenotypes corresponding to different polynomial functionals. The credit assignment functions take into account both self and symbiotic terms in an adaptive and dynamic manner, in order to produce disparity patches that are fit within their own domain and at the same time fit in association with their symbionts. This persistent propagation of local interactions to a global scale throughout evolution generates a unified disparity surface composed of the many smaller patch surfaces.