Machine learning models have significantly progressed toward automation in many financial operations, including stock market prediction and investment portfolio development. Due to the volatility of the stock market, stock selection and asset allocation remain challenging for investors. The modern portfolio theory (MPT) has faced challenges in real-world cases due to its heavy reliance on variance, which fails to account for non-normal, heavy-tailed investment losses. In this research study, three advanced models are implemented and compared with the standard MPT framework, including conditional value-at-risk (CVaR), hierarchical risk parity (HPR), and a genetic algorithm (GA) for cardinality-constrained optimisation. The dataset includes historical prices for 18 US stocks from 2016 to 2025. Strict constraints are imposed during the implementation of these models, including long-only positioning and a per-asset cap. An evaluation of risk-adjusted performance using the sortino ratio (STR) demonstrates that clustering-driven HRP and tail-consistent CVaR models outperform MPT significantly. The CVaR-based portfolio has achieved a Sharpe ratio (SR) of 1.58 and a 2.42 STR, demonstrating its ability to mitigate downside risk. GA has achieved the highest annual return of 45.8% and 2.03 SR; simultaneously, drawdown and volatility are also high, proving the inherent trade-offs between maximising empirical SR and robust risk management. The research findings indicated that GA can efficiently build a risk-adjusted optimal investment portfolio while imposing real-world constraints.

We introduce a plug-and-play framework for reconstructing natural images from functional Magnetic Resonance Imaging (fMRI) using a frozen diffusion prior. The approach decouples appearance and semantics: early-visual activity is linearly decoded to a Variational Autoencoder (VAE) latent that preserves structure, while higher-level activity is decoded to a text embedding that guides content. A lightweight diagonal ridge calibration aligns predicted semantics with the model’s conditioning space. Reconstruction runs in image-to-image mode by stochastically encoding from the predicted appearance latent and denoising with U-Net conditioned on the calibrated semantics—without fine-tuning the generative model. On Natural Scenes Dataset (NSD), the method improves semantic alignment and perceptual fidelity over strong baselines, offering a simple and reproducible pathway to brain-guided image synthesis.

Imagine a classroom where every student gets a personal tutor trained in the Socratic art—not to give answers, but to ask just the right questions. That vision, however, has long run into a practical wall: how do you scale a method built on live, adaptive dialogue to hundreds or thousands of learners? In this article, we explore a generative-AI model designed to do exactly that. Instead of reducing Socratic teaching to a scripted Q&A, we treat the AI as a “midwife” for thinking—one that operates from deliberate ignorance, provides cognitive scaffolding, and dynamically shifts initiative back to the learner. Technically, the system works through a three-layer “Goal–Agent–Conversation” architecture, turning broad instructional aims into living, personalized dialogues. Learners move through a kind of cognitive spiral: they externalize their assumptions, run into contradictions, reflect critically, and gradually rebuild understanding. What we’re really trying to address here is an old tension in education—the tug‑of‑war between standardization and genuine personalization. Could this approach ease that friction? Looking ahead, we see the beginnings of a data‑informed learning model that doesn’t just deliver knowledge but cultivates the skill of thinking itself. We also sketch where the work goes from here—from technical fine‑tuning and new use cases to the ethical questions that will inevitably follow.

With the rapid development of UAV technology, their applications in urban traffic monitoring have become increasingly important. This study investigates the relationship between UAV flight altitude and overall operational performance. A performance metric, defined as the ratio of effective coverage area to operation-related energy consumption, is proposed to quantitatively evaluate UAV efficiency. Precise models are developed to capture the effects of altitude on both coverage and energy consumption, while other less relevant factors, such as UAV weight and image resolution, are held constant. The results reveal the existence of an optimal altitude that maximizes the proposed performance metric. Moreover, the findings are generalizable and can be adapted to different UAV-based traffic monitoring systems by adjusting model parameters.

This paper reviews automated palm tree identification methods based on traditional machine learning and deep learning, exploring various data sources (e.g., UAV, satellite, and aerial imagery) and imaging modalities (RGB, thermal, multispectral, and LiDAR). It systematically analyzes the strengths and limitations of segmentation, classification, and object detection algorithms, highlighting their practical applications in agriculture, ecological conservation, and urban planning. The study demonstrates that deep learning approaches (e.g., CNNs and Vision Transformers) excel in complex environments and large-scale datasets but face challenges such as species similarity, environmental variability, and computational constraints. Future research should focus on multimodal data fusion, lightweight model development, and dataset expansion.

V2X has extremely high requirements for real-time information transmission, demanding as low a latency as possible to obtain timely information and maintaining a sufficiently high reliability to ensure complete and secure information. RSMA avoids complex successive interference cancellation (SIC) operations, such as those performed in NOMA, through a layered signal processing method. Although RSMA can enhance robustness, it has a high dependence on instantaneous channel knowledge. The segmentation, coding, and resource allocation for different parts of the signal all require more complex algorithm support, gradually increasing the complexity of RSMA and making it unable to meet the information transmission requirements in the V2X environment. This paper introduces a reconfigurable intelligent surface (RIS) into RSMA, using only second-order channel statistics to construct an algorithm with a quasi-closed-form solution, eliminating the dependence on instantaneous channel knowledge. Simulation results show that this method can effectively reduce computational overhead and is expected to become a solution for the development of next-generation 6G vehicle communication systems.

Existing quantization and sparsity algorithms are basically intended for tradition neural network, which makes them unsuitable to adapt to the Transformer architecture. This paper proposes a dynamically adjustable EBSP+GQSA fusion compression method. We improved the EBSP algorithm by introducing a dynamic sparsity strategy and block-level dynamic shift factors to enhance its adaptability to attention weights. Simultaneously, we optimized the GQSA algorithm by proposing a dynamic grouping strategy and layer-wise weight analysis to improve compression flexibility and efficiency. Experimental results on typical Transformer models, including BERT-base, ViT-base, and GPT-2-small, demonstrate that the EBSP+GQSA fusion method achieves minimal accuracy loss (<1%), the highest compression ratio (>70%), the greatest inference speedup (>2×), and the most significant energy reduction (>55%), outperforming the individual use of EBSP or GQSA. This method significantly enhances the deployment capability of models on resource-constrained edge devices, providing a feasible technical path for efficient Transformer inference.

To enhance the accuracy and naturalness of human-machine interaction (HMI) for upper limb exoskeletons, the key lies in overcoming the adaptive bottleneck of their force feedback systems. Traditional methods have limitations in multimodal information fusion and dynamic control. Therefore, this study aims to develop an adaptive force feedback system based on multimodal sensors. By deeply fusing surface electromyography (sEMG) and mechanical sensor data, and introducing deep learning algorithms, a control strategy capable of dynamically adapting to user intentions was constructed. Notably, we independently designed and fabricated core temperature and pressure sensors. Experimental data shows that their characteristic equations are y = 0.041x - 0.90 and y = 0.1x, respectively, and their performance is highly consistent with that of standard sensors. This study confirms that the proposed scheme can effectively improve the sensing accuracy and adaptive capability of the force feedback system, providing a reliable hardware foundation and a novel technical approach for achieving more intelligent and collaborative human-machine interaction.

Artificial intelligence (AI) shows great potential for improving surgical efficiency, precision, and autonomy in surgical robotic systems. However, the robustness of deep learning-based algorithms remains a critical challenge as the surgical environments shows much variance in real application. Most deep learning-based segmentation models, though highly effective on benchmarking datasets, often fail during unforeseen nonadversarial corruptions such as occlusions, bleeding, or low brightness. In this study, we introduce a domain-specific augmentation strategy to enhance model robustness against possible surgical corruptions that is not seen in the training data. Our method simulates key corruptions, including blood simulation, brightness adjustment, and contrast adjustment. Based on the SegSTRONG-C benchmark, we evaluate a baseline U-Net model on a binary surgical tool segmentation task. While the baseline shows strong performance on clean images, its accuracy drops substantially on the corrupted test data. Incorporating our proposed augmentations significantly improves performance on corrupted inputs while preserving accuracy on the clean domain. These findings underscore the importance of specific augmentation for model’s robustness and demonstrate a practical pathway toward more reliable and generalizable segmentation models for real-world surgical robotics applications

The advancement of colorectal cancer emphasizes how important it is for colonoscopic imaging to accurately segment polyps. Learning-based techniques have made significant progress in the field of polyp medical image segmentation; however, recurring issues such as the identification of small object segments, poorly defined lesion boundaries, and complex backgrounds still exist. In order to overcome these constraints, we introduce CDBT-Unet, a brand-new framework that enhances segmentation performance by integrating two significant innovations. Initially, the transformer layer's convolutional prior speeds up convergence and extracts the fine-grained local texture that is essential for tiny flat polyps. By prioritizing horizontal-vertical background relationships through cross-shaped attention, it improves boundary delineation in complex backgrounds by reducing computation and accelerating convergence. The intricate background and edge blurring issue of polyp segmentation is well-considered in this point. Second, in order to improve accuracy, our dual-path encoder uses the MaxViT block to strategically balance global dependency modeling and local feature preservation. Combining multilevel feature fusion with coordinate space focus mechanisms and channel refinement improves edge response in multiscale fusion. The issue of boundary blurring is the main focus. Under the same experimental setup, our model outperforms the state-of-the-art ConDseg model by 3.72% and the baseline (TransUnet) by 7.32% in terms of Dice scores when tested on the Kvasir-SEG and CVC-ClinicDB datasets. Even in the presence of motion artifacts or low contrast, the framework demonstrates exceptional robustness in segmenting polyps of various sizes. Furthermore, the attention maps that were produced enhanced interpretability and gave physicians practical knowledge about how to make decisions when modeling.