Reconstructing the human connectome at the neuron level presents a formidable challenge, requiring approximately 400 trillion hours of manual work by trained neuroscientists. However, with advancements in machine learning and deep learning, it has become feasible to automate this process. This paper explores various automated methods for reconstructing the 3D geometry of neural tissue through electron microscopy (EM) image segmentation. We focus on enhancing existing segmentation pipelines by improving both their efficiency and accuracy. Our contributions include the development of a modular segmentation pipeline, DeepSeg, written in Python, allowing for flexible experimentation at different stages of the computational pipeline. Furthermore, we investigate multiple neural network architectures to improve segmentation accuracy and robustness, mitigating error propagation across stages. By leveraging learned alignment strategies and evaluating the performance of different approaches, this work offers a promising path towards the ultimate goal of generating neuron-level connectivity graphs, which can significantly advance our understanding of neural structures and brain function.

In today's rapidly evolving cyber landscape, the growing sophistication of attacks, including the rise of zero-day exploits, poses critical challenges for network intrusion detection. Traditional Intrusion Detection Systems (IDSs) often struggle with the complexity and high dimensionality of modern cyber threats. Quantum Machine Learning (QML) seamlessly integrates the computational power of quantum computing with the adaptability of machine learning, offering an innovative approach to solving intricate and highdimensional challenges. A key factor in QML's performance is the method used to encode classical data into quantum states, as it defines how data is represented and processed in quantum circuits. QML offers promising advances for IDS, particularly through hybrid quantum-classical models. This study presents an in-depth comparative analysis of quantum-classical data encoding techniques for QML-based IDS. To the best of our knowledge, this is the first study to comprehensively evaluate the performance impact of different quantum encoding methods and provide a thorough evaluation of their impacts on the overall model performances. To achieve this, we first present a comprehensive evaluation of quantum and classical data encoding techniques, focusing on four key encoding techniques namely, Amplitude Embedding, Angle Embedding, Instantaneous Quantum Polynomial (IQP) Encoding, and Quantum Approximate Optimization Algorithm (QAOA) Embedding. Then, we develop a hybrid quantum-classical QML model to analyze how each encoding affects classification performance for malicious traffic. Finally, we conduct extensive experiments using two well-known, real-world network attack datasets to assess the accuracy and efficiency of each encoding approach. Our obtained results show notable differences in classification accuracy, underscoring the importance of encoding choice in optimizing QML-based IDS. This study aims to advance the application of quantum methodologies in network security by identifying effective encoding strategies for intrusion detection.

State-of-the-art deep neural networks (DNNs) for inverse scattering are typically trained for fixed arrangements of transmitters and receivers, restricting their applicability to setups with fixed viewing angles. Consequently, any change in these angles necessitates retraining the network. One way to avoid retraining is to train the DNN on all possible combinations of viewing angles for varying numbers of views, but this can be tedious. Instead, we employ a more efficient approach that uses two U-Nets. First, the scattered field from each view is independently processed through a shared U-Net to produce view-wise contrast reconstructions. These view-wise reconstructions are aggregated and then refined using the second U-Net to obtain the final reconstruction. Using two different sets of contrast profiles-one with circular cylinders and the other from the MNIST dataset-we demonstrate that this approach achieves good contrast reconstruction for arbitrary viewing angles and an arbitrary number of views. However, performance decreases on rotated MNIST images, as this dataset is not rotation-augmented. While data augmentation through image rotations is one solution, we propose training the U-Nets on data transformed into polar coordinates. We show that this method accommodates arbitrary views and is robust to rotated versions of the contrast profiles without relying on data augmentation.

The concept of Semantic Vector Collapse proposes a novel approach to recalibrating vector representations within large-scale transformer models, aiming to mitigate the effects of semantic drift and improve memory retention over extended input sequences. Through the dynamic computation of relevance-weighted projection matrices, the method refines highdimensional embeddings, enhancing their ability to prioritize critical contextual information while suppressing redundancy. Experiments demonstrate significant improvements in sequence coherence and contextual alignment, with marked reductions in entropy observed across varying input lengths. The framework adapts seamlessly to diverse linguistic tasks, achieving notable robustness to noise and variability in input data, as evidenced through qualitative and quantitative analyses. Results from embedding space evaluations indicate optimized dimensionality utilization, enabling efficient memory representation and facilitating scalable applications in real-world scenarios. The method's integration into pre-trained architectures requires minimal modifications, ensuring practical applicability without introducing prohibitive computational overhead. Observations of improved dependency parsing and thematic continuity suggest a profound impact on the model's ability to handle complex semantic relationships across long sequences. Memory retention rates and token-level similarity metrics reflect the transformative potential of the recalibration mechanism in addressing challenges inherent to extended text processing. Additional insights into embedding compression ratios and robustness metrics emphasize the scalability and efficiency of the approach. The mathematical rigor underpinning Semantic Vector Collapse highlights its innovative contribution to advancing semantic representation dynamics within transformer-based language models. Its ability to achieve significant contextual improvements with resourceefficient techniques marks a step forward in the design and functionality of large-scale natural language processing systems.

Hardware accelerators for artificial intelligence (AI) applications must often meet stringent constraints for accuracy and throughput. In addition to architecture/algorithm improvements, high performance computational techniques such as mixed precision are also required. In this paper, a floating-point fused multiply-add (FMA) unit supporting mixed/multiple precision is proposed. A wide range of conventional precision formats (such as half and single, formats) as well as emerging precision formats (including E4M3, E5M2, DLFloat, BFLoat16 and TF32) are supported in the proposed design. In addition to all these formats, the proposed design is flexible in manipulating the exponent and mantissa lengths for 8, 16 and 32-bit floating-point numbers based on the needs of an application. The proposed FMA can be configured to perform normal FMA operation supporting multiple precision, or alternatively perform mixed precision in ASIC. The proposed FMA is fully pipelined and in each cycle, the input bit streams are processed based on the provided configuration, so independent of the previous cycles. For normal FMA, the proposed design utilizes sharing of resources to parallelize several operations based on the available hardware and required precision. Mixed precision tends to accumulate the lower precision dot products into higher precision to avoid overflow/underflow. The proposed design improves computational accuracy by adding all possible dot products at the same time while decreasing the number of rounding operations to prevent rounding errors. An innovative method to accumulate the dot products and the aligned addend is also proposed for normal/mixed precision FMA operation. By considering tradeoffs between reusing the available hardware and removing unnecessary complex units, a more efficient and flexible design is attained in terms of hardware metrics and supported mixed precision computation compared to other designs found in the technical literature. Extensive simulation results for a comparative analysis are provided.

Although deep learning (DL) has led to several breakthroughs in many disciplines of science, technology, engineering, and mathematics (STEM), a fundamental understanding of why and how it is empirically successful remains elusive. In attacking this fundamental problem and unraveling the mysteries behind DL's empirical successes, significant innovations toward a (unified) theory of DL have been made. Even though these innovations encompass nearly fundamental advances in optimization, generalization, and approximation, no work to date has quantified the testing performance of a DL-based algorithm employed to solve a pattern classification problem. In overcoming this fundamental challenge in part, this paper exposes the fundamental testing performance limits of feedforward neural network (FNN)-based binary classifiers trained with hinge loss. For binary classifiers based on hinge-loss-trained deep FNNs with ReLU and Tanh activation, we derive their fundamental asymptotic testing performance limits that we validate using extensive computer experiments.

Although deep learning (DL) has led to several breakthroughs in many disciplines as diverse as chemistry, computer science, electrical engineering, mathematics, medicine, neuroscience, and physics, a comprehensive understanding of why and how it is empirically successful remains fundamentally elusive. To attack this fundamental problem and unravel the mysteries behind DL's empirical successes, the DL research community has made significant innovations in optimization, generalization, and approximation. However, no work to date has quantified the testing performance of a DL-based algorithm employed to solve a pattern classification problem. To address this fundamental challenge in part, we expose the testing performance limits of binary classifiers based on deep rectified linear unit (ReLU) feedforward neural networks (FNNs) that are trained with hinge loss by driving their fundamental asymptotic testing performance limits. These limits are validated by our extensive computer experiments.

Video Capsule Endoscopy (VCE) presents a significant challenge, particularly due to the vast and unstructured data generated through the GI tract which may lead to difficulties in realtime analysis and classification. To address these issues, this study presents a semi-supervised methodology for image classification that leverages texture-based features. This method initially addresses the imbalance in the provided dataset and utilizes advanced texture-based feature extraction techniques such as Gray Level Co-Occurrence Matrix, Local Binary Pattern, and deep features derived from VGG16. Then, a semi-supervised learning approach, Transductive Learning Algorithm has been carried out that strengthens the model's robustness and ability to classify the normal and abnormal classes, yielding improved accuracy. Additional classification models like Random Forest and K-Means Clustering are also carried out for comparison study. The Gray Level Co-Occurrence Matrix with transductive learning outperformed the other approaches in accurately classifying the images by achieving an effective accuracy of 95.14% with 0.99 mean AUC. In conclusion, our approach which has been carried out achieved 23 rd rank in the Capsule Vision 2024 Challenge, paves the way for enhancing diagnostic capabilities, ultimately contributing to more accurate and timely identification of gastrointestinal conditions.

Recent advances in stable diffusion models have revolutionized text-to-image generation. However, these models struggle with spatial relationship comprehension in remote sensing scenarios, limiting their ability to generate spatially accurate imagery. We present a novel framework for generating remote sensing images from spatial relationship descriptions with precise spatial control. Our approach introduces a two-stage pipeline: first, a spatial relationship semantic structuring model converts formalized spatial relationship descriptions into controlled layouts; second, an enhanced diffusion model incorporates positional prompts and a layout attention mechanism to generate the final image. The positional prompts explicitly encode spatial information, while the layout attention mechanism enables focused region learning. Comprehensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches in both spatial accuracy and image quality.

This is a proposal of a novel framework leveraging graph gauge theory to develop the Quantum Virtual Mind (QVM), a computational platform integrating classical and quantum algorithms to mimic human brain neural activity. The core hypothesis posits that an artificially trained quantum connectome can emulate neural patterns by mapping electromagnetic (EM) field distortions onto symmetry-breaking phenomena within a quantum computational framework. EM distortions, captured via Brain-Computer Interface (BCI) devices, characterize the quantum connectome and enable quantum simulations in the QVM, allowing real-time computation of an individual's emotional signatures, referred to as the Mental Stress Tensor (MST). This work bridges gauge theory, connectome studies, and neural networks to facilitate computation of agentic components in Multi-agent AI systems.

The Gene Regulatory Network (GRN) in biological cells orchestrates essential functions for adaptation and survival in diverse environments, drawing on structural similarities with the Artificial Neural Network (ANN) to transform into the Gene Regulatory Neural Network (GRNN). This transformation enables exploration of their natural computing capabilities regarding network reconfigurability and controllability, facilitating dynamic adjustments of gene-gene interaction weights to regulate biological processes. In this paper, we present a control-theoretic model for the GRNN that determines optimal chemical input concentrations, steering the GRNN towards desired weight configurations using the Linear Quadratic Regulator (LQR) approach. This method enhances network robustness by optimizing stability and minimizing plasticity, ensuring adaptive reconfigurability in dynamic environments. We develop mathematical models to identify critical genes using a Continuous-Time Markov Chain (CTMC) and derive temporal weight configurations, providing insights into the system's reconfiguration dynamics, while also quantifying stability and plasticity. Our findings demonstrate the effectiveness of the control model in mitigating Clostridioides difficile biofilm formation, outperforming sub-optimal and stochastic inputs, and highlighting the importance of determining optimal inputs for robust network behavior across diverse complexities.

Today, the challenges and opportunities that are now inherent in the development of generative artificial intelligence (GAI) for cybersecurity are evolving rapidly. As such, this article examines the proactive capabilities of GAI, with a focus on how it has revolutionized the concept of cyber threats and the approaches to threat intelligence. It investigates the predictive and adaptive potential of GAI to combat sophisticated attack vectors, zero-day vulnerabilities, and automated threat modeling. According to studies, this research highlights that GAI can predict attack patterns with 87% accuracy and detect zero-day vulnerabilities with 80% precision. In addition, GAI-based intrusion detection systems (IDS) exhibit high detection rates (98% for known threats and 92% for unknown threats) at low false positive rates. This data shows that integrating GAI into cybersecurity enables organizations to discover vulnerabilities prior to exploitation, simulate realistic attack scenarios, and automate threat reactions. However, as with GAI in cybersecurity, there are significant. ethical and operational risks of GAI in cybersecurity, including the capacity to enable adversaries and magnify existing vulnerabilities that need close regulatory oversight. Organizations leverage their predictive power to anticipate and proactively secure their digital infrastructure from emerging threats, moving from an earlier reactive to a more anticipatory security posture. In the end, the findings from this research point to the need to strategically leverage GAI to strengthen cybersecurity strategies by creating fortified defenses in an ever more complicated digital environment.

This paper investigates the physical layer performance of 5G New Radio (NR) in a cell-free (CF) massive MIMO system, focusing on the transmission of the Physical Uplink Shared Channel (PUSCH) over frequency-selective Rayleigh fading channels. A comprehensive link-level simulator, compliant with 3GPP standards, is developed to evaluate the system's performance across various scenarios. Key parameters such as subcarrier spacing (SCS), modulation and coding schemes (MCS) (QPSK, 16-QAM, 64-QAM, 256-QAM), and varying numbers of distributed radio units (RUs) are systematically analyzed. The results demonstrate the dual benefits of increasing the number of RUs: enhanced spatial diversity and improved proximity between user equipment (UE) and RUs, leading to significant improvements in reliability and error rate performance. The study also highlights the impact of higher SCS values in leveraging frequency diversity, particularly when signal bandwidth exceeds the channel's coherence bandwidth. Practical channel estimation is incorporated to validate performance under realistic conditions. Block Error Ratio (BLER) is used as the primary performance metric, providing valuable insights into the interplay of spatial and frequency diversity in CF 5G NR systems, thereby confirming their potential to achieve robust and efficient communication in challenging wireless environments.

Ultrasonic attenuation can be used to characterize tissue properties of the human breast. Both quantitative ultrasound (QUS) and ultrasound tomography (USCT) can provide attenuation estimation. However, limitations have been identified for both approaches. In QUS, the generation of attenuation maps involves separating the whole image into different data blocks. The optimal size of the data block is around 15 to 30 pulse lengths, which dramatically decreases the spatial resolution for attenuation imaging. In USCT, the attenuation is often estimated with a full wave inversion (FWI) method, which is affected by background noise. In order to achieve a high resolution attenuation image with low variance, a deep learning (DL) based method is proposed. In the approach, RF data from 60 angle views from the QTI Breast Acoustic CT TM Scanner are acquired as the input and attenuation images as the output. To improve image quality for the DL method, the spatial correlation between speed of sound (SOS) and attenuation were used as a constraint in the model. The results indicate that by including the SOS structural information, the performance of the model is better. With a higher spatial resolution attenuation image, further segmentation of the breast can be achieved. Segmentation of breast using the DL-based attenuation image had a high spatial correlation with the segmentation from the FWI-based SOS image, i.e., with correlation coefficients greater than 0.64, and Intersection over Union of the ROI greater than 0.73, which confirms that the DL-generated attenuation images can be used to do image segmentation.

The release of Segment Anything Models (SAM1 and SAM2) by Meta has significantly influenced various domains, including medical imaging. However, existing implementations of SAM primarily focus on standalone tools that require local installation and configuration, limiting accessibility and ease of use. In this work, we present the web-based extension of SAM2 integrated into the Open Health Imaging Foundation (OHIF) viewer. Our solution supports all SAM2 prompt types, including points and bounding boxes, and enables multi-label predictions. This web-based integration eliminates installation requirements, offering a more user-friendly interface. The implementation is open-source and available at https://github.com/CCI-Bonn/OHIF-SAM2.

Vehicular Knowledge Networking (VKN) is a paradigm where vehicles exchange knowledge instead of data. Named Data Networking (NDN) is a resilient architecture for sharing data between vehicles without information about the hosting vehicle. NDN might not be adapted for sharing knowledge, as the peculiar complexity of knowledge requires knowledge-driven NDN functions as well as a specific simulation environment enabling joint knowledge perception, inference and reasoning. In this paper, we present an open-source cosimulation framework connecting NDN, traffic, dynamic control and perception simulators for knowledge perception & inference, and with an AI-as-a-Service (AIaaS) platform for knowledge reasoning. This paper notably describes new interfaces and functions between the NDN daemon and the AIaaS micro-services for handling interest naming, caching and forwarding between knowledge producers and consumers. We demonstrate the benefit of the proposed framework and knowledge-specific extensions through an AI-driven vehicular intersection management.

The intricate challenge of maintaining semantic coherence and contextual alignment in text generation has driven the development of innovative embedding recalibration methodologies. Contextual Embedding Recalibration introduces a dynamic mechanism that iteratively adjusts token representations to reflect evolving semantic requirements within extended sequences. The recalibration framework leverages hierarchical token encoding and adaptive weight computation to balance localized fidelity and global coherence, addressing limitations of traditional static embedding strategies. Empirical evaluations across diverse tasks demonstrate substantial improvements in perplexity reduction, contextual consistency, and robustness under varied conditions, including noisy inputs and cross-domain generalization. A key feature of the recalibration layer is its seamless integration with existing transformer architectures, achieved through non-linear transformations and entropy-based regularization mechanisms. Quantitative analyses reveal that the recalibrated model excels in maintaining long-range dependencies, reducing semantic drift, and aligning tokens effectively in multi-turn dialogues and narrative completions. The ability to process out-of-vocabulary tokens with improved accuracy further highlights its adaptability to real-world linguistic variability. By incorporating computationally efficient techniques, the recalibration process achieves scalability while preserving performance gains across large datasets. The presented results demonstrate the efficacy of embedding recalibration in enhancing the representational depth and contextual reasoning capabilities of large-scale transformer models. Cross-domain evaluations confirm the robustness of the recalibrated model, indicating its versatility for applications requiring context-sensitive natural language generation. The findings not only address fundamental challenges in token representation but also offer a foundational framework for advancing machine learning methodologies in broader linguistic contexts. The proposed recalibration approach establishes a benchmark for future explorations into adaptive embedding mechanisms within artificial intelligence frameworks.

Parametric Semantic Lattices (PSL) offer a new approach to enhancing semantic representation and contextual processing in advanced language modeling architectures. By embedding hierarchical structures that capture both local and global linguistic dependencies, PSL addresses key limitations observed in traditional transformer-based frameworks. The adaptive nature of the lattice configurations enables the dynamic interpretation of syntactic and semantic relationships across diverse linguistic inputs. Quantitative analysis reveals significant reductions in perplexity and marked improvements in semantic coherence when compared to baseline models. Through leveraging multi-dimensional encoding, the PSL framework enhances interpretability while maintaining computational efficiency across long text sequences. Experiments demonstrate its scalability across extensive datasets and various hardware configurations, validating its application to resource-intensive natural language processing tasks. Evaluation under conditions of noisy input highlights its robustness in maintaining contextual fidelity and accuracy. Practical applications span areas such as automated translation, sentiment analysis, and advanced query resolution, showcasing its versatility. The integration of PSL into transformers introduces minimal computational overhead while yielding substantial performance gains. Results further highlight its ability to retain long-range dependencies, a critical factor for tasks requiring deeper contextual understanding. By redefining structural modeling within transformers, PSL opens new avenues for the development of contextually aware artificial intelligence systems. The findings collectively underscore its transformative potential in advancing the state of language modeling and representation.