Online financial forums have become a popular channel through which investors express their opinions to the firm-level information. This raises a natural question: does user-generated sentiment predict short-term stock price movements? This study examines the relationship between firm-level online investor sentiment and next-day stock price changes using user comment data from a Chinese financial forum. We construct a daily sentiment measure using the proportion of negative comments and relate it to next-day stock price change. Besides, two regression models are employed for estimate if the stock price change can be predicted by daily sentiment measure. Our linear regression results show that, although the estimated sentiment effects are consistent in direction of price change, the model exhibits low explanatory power and the effects are not statistically significant. The logistic regression model achieves moderate classification accuracy in determining the direction of price change, but the ability is still limited. Overall, the results suggest that sentiment derived from online investor comments contains limited informative signals for next-day stock prices. The evidence is consistent with the interpretation that such comments primarily reflect immediate reactions or noise rather than forward-looking signals.

EEG is a non-invasive technique for recording brain activity, valued for its high temporal resolution and low cost. Independent Component Analysis (ICA), a blind source separation method, effectively decomposes EEG signals into independent components. Widely used in both medical research and brain-computer interfaces (BCIs), ICA has become essential for improving EEG data quality. As demand grows for real-time and robust neural signal processing, advanced methods like ICA remain crucial for advancing EEG applications. This paper investigated the application of ICA in EEG signal processing. The research first reviewed the core principles of ICA, including its mathematical model and algorithms such as Infomax and FastICA. The results clearly demonstrated that ICA can effectively decompose raw EEG signals into statistically independent components. This enables the identification and removal of artifacts while preserving neural information. It also showed that while ICA is a mature method for EEG preprocessing, challenges such as source number determination and real-time processing bottlenecks still exist. At the end of the research, future developing trends of ICA and its integration with other methods are discussed.

Underwater images serve as essential visual data for marine resource exploration, underwater robot operation, and other related fields. Due to light absorption by water, scattering by suspended particles, and refraction distortion caused by imaging devices, usually, there will be degradation problems such as color shift, blurring and low contrast. This study examines the negative effects of these issues on the efficiency of certain visual analysis tasks. We present a detailed survey on rehabilitation approaches for underwater images. We describe the disparate effects of underwater image degradation and summarize their physical causes. Existing image restoration approaches fall into three primary types: (1) physics-based methods, (2) non-physics-based methods, and (3) deep learning-based methods. We also analyze the pros and cons, as well as the integration perspective, of these methods. Additionally, we review the existing underwater image datasets and quality assessment frameworks to present the current state and issues of these methods. This study builds a knowledge base for subsequent works on underwater image restoration, advocating for a faster deployment of the restoration methods for underwater image rehabilitation to a wide range of applications in marine engineering and underwater exploration. This organized systematized literature review is prepared to aid researchers in the future by providing a structured reference framework regarding the restoration of degraded underwater images. We anticipate the implementation of the technologies detailed in this review to further improve the mitigation of underwater obstructions within a lateral distance of the submersible craft and the attire of underwater construction.

Credit scoring remains a standard tool in the banking industry because it provides an interpretable way to estimate default risk. In particular, scorecards based on Weight of Evidence (WOE), Information Value (IV), and Logistic Regression (LR) are widely adopted due to their transparency and practical usefulness. However, traditional scorecards are mainly used to assess customers' risk. They do not directly optimize subsequent actions, such as credit limit allocation. This paper proposes a two-stage framework for consumer credit analysis. In the first stage, it constructs a scorecard model using an anonymized bank dataset. The workflow includes data cleaning, decision-tree-based binning, WOE/IV transformation, correlation filtering, stepwise logistic regression, and scorecard construction. The resulting model achieves moderate predictive power, with a KS value of 0.386 and an AUC of 0.739. In the second stage, the scorecard is extended into a sequential decision-making framework. Predicted risk levels from the scorecard are used to define feasible credit-limit actions, and Thompson Sampling is applied to learn adaptive limit-allocation policies under uncertainty. In the simulation, three credit-limit levels are considered, and the reward is designed as a simplified risk-adjusted return. The final cumulative reward reaches 276,550, and the learned policy exhibits intuitive risk-consistent behavior: high-risk customers have their credit limits limited to low levels, medium-risk customers have their credit limits mainly at medium levels with a small portion at low levels, while low-risk customers are more likely to have higher credit limits. Overall, this study suggests how a traditional scorecard can be extended from pure risk prediction to scorecard-guided credit limit decision-making. The result also remains interpretable while introducing an adaptive bandit-based decision layer.

Fatigue is a critical human element affecting civil aviation flight safety. Accurate acquisition of crew fatigue data and experimental research are of great significance for establishing a scientific fatigue risk management system and ensuring flight safety. Based on existing relevant academic literature, this study systematically reviews the primary methods for fatigue data acquisition in aircraft engines, including physiological parameter collection such as Electroencephalogram (EEG), Electrocardiogram (ECG), and Electrooculogram (EOG), behavioral monitoring of eye movements and operational behaviors, subjective evaluation scales, and cockpit environmental parameter monitoring. Experimental studies have demonstrated that EEG is the most sensitive indicator of fatigue, while heart rate variability and eye movement features can effectively reflect variations in fatigue levels. Significant discrepancies exist between laboratory simulations and actual flight test results, as real flight environments are subject to greater interference from meteorological and psychological factors.The causes and characteristics of crew fatigue differ between long-haul and short-haul flight routes. Single-modal monitoring has limited reliability, while multimodal data fusion combined with machine learning can significantly improve fatigue identification accuracy. This study provides theoretical support and technical references for optimizing real-time fatigue monitoring systems in civil aircraft crews.

With the in-depth integration of communication technology and neurology, the application of communication technology in neural signal acquisition, transmission, analysis, diagnosis and treatment has become increasingly widespread. This paper mainly adopts the literature analysis method to sort out the research status of communication technology in the field of neurology. First, it outlines the basic characteristics of neural signals and related communication support technologies, focuses on analyzing three major application scenarios: brain-computer interface (BCI), neurological disease monitoring and implantable neural devices, then summarizes the main challenges faced by current technologies, and looks forward to the future development trends. The research finds that core communication technologies such as wireless transmission, signal processing and decoding have been initially applied in scenarios such as neural rehabilitation, disease monitoring and implantable devices, effectively solving key problems such as high-fidelity transmission and remote regulation of neural signals. However, there are still prominent bottlenecks in biocompatibility, technical standardization and ethical norms. In the future, it is necessary to rely on technologies such as 6G and artificial intelligence to break through existing limitations and promote the in-depth integration and clinical transformation of communication technology and neurology.

The fast development of artificial intelligence toward big models, many parallel calculations, and edge-cloud cooperation has further shown the limits in performance and energy use of traditional von Neumann structures. In these structures, memory and processing units are physically separated, which makes data transfer happen very often. This leads to too much delay and extra energy waste linked with the memory wall. Computing-in-memory is a good new choice because it does arithmetic tasks directly inside memory arrays, so it cuts down data moving as much as possible. Among new nonvolatile memory devices, ferroelectric field-effect transistors are very good for CIM. They have many advantages, such as nonvolatility, high ability to read data correctly, low energy use for writing data, and good compatibility with CMOS integration. Especially, FeFETs made of Hf₁₋ₓZrₓO₂ show steady ferroelectric properties even in very small nanoscale sizes. This makes it possible to make CIM hardware that can be scaled up and uses energy efficiently. In this work, we use a crossbar array based on HZO FeFETs to do matrix multiplication in a CIM system. We put forward a calculation way where weight values are stored in FeFET polarization states, input activation signals are sent through word lines, and multiply–accumulate operations are done by adding currents in bit lines. We design a 3×3 HZO FeFET array to test the matrix-multiplication method in real experiments. Test results show that the array does in-memory MAC operations correctly, and the output results are the same as theoretical values. We also study its scalability, and it shows that this design can be extended from a 3×3 array to any common n×n array without changing the basic calculation principle. These findings prove that HZO FeFET-based CIM can work well for efficient matrix operations. It is a practical hardware way for future AI computing systems that use low power and have high parallel working ability.

With the rapid development of deep learning techniques within medical image analysis, convolutional neural networks (CNNs) that play a crucial role in disease detection and auxiliary diagnosis are emerging. However, the traditional deep learning methods are extremely dependent on human expertise when it comes to model construction and hyperparameter adjustment which makes the development cycle rather long the technical threshold quite high and the resource consumption also quite large. In order to enhance the modeling efficiency and reduce the dependence on expertise automated machine learning (AutoML) has indeed become a crucial direction in medical AI research. It is able to automatically handle the key processes, like the things such as model selection, neural architecture search and hyperparameter optimization. Taking the classification of brain tumor magnetic resonance imaging images as an example this article by means of the Baidu EasyDL platform constructs an AutoML model based on the public datasets from Kaggle. A systematic comparative analysis is carried out on traditional self-defined CNN models. It turns out that AutoML has struck a fairly good balance between efficiency and performance, which has greatly reduced human intervention as well as model design costs and at the same time has maintained a relatively high classification accuracy. However, this research also points out the existing limitations in the interpretability of the model and the problems in overfitting control.

With the acceleration of global urbanization, ground space resources are becoming increasingly saturated. The low-altitude economy, a new growth pole of the spatial economy, is strategically important for the sustainable development of smart cities. This study aims to illustrate the thematic importance of the integration between the low-altitude economy and smart cities, and to analyze how to resolve the contradiction between the fragmentation of urban airspace resources and the complexity of security governance through technological means. This paper systematically reviews the current development status of the low-altitude economy at home and abroad, outlines the dynamic allocation logic of airspace under big data collaborative governance, and the application paths of artificial intelligence in specific scenarios. The results indicate that the large-scale development of the low-altitude economy highly relies on cross-departmental data sharing mechanisms and the precise support of AI algorithms. The conclusion points out that the deep integration of the low-altitude economy and smart cities can significantly improve urban operational efficiency, and proposes optimization directions for future construction in response to current challenges such as inconsistent standards and weak infrastructure.

In traditional racing games, rival characters are authored by designers during development, and players race against whoever shipped with the game. Large language models make per-player rival generation feasible, but turning that possibility into a usable tool depends on a question of shape: what kind of input should humans and the model share, so the user can specify a rival in terms they can relate to and the model can act on? RivalCraft is our attempt at an answer. A rival is specified through six named personality sliders (aggression, patience, risk-taking, sportsmanship, psychological play, consistency) layered on an archetype; these are the shared vocabulary between the user and the LLM. The model returns a structured profile that the user can edit in place. In a formative study with eight participants using SUS, a six-item perception questionnaire, and short open-ended questions, the mean SUS was 78.1 (SD = 11.9); six of eight scored above the 68-point above-average threshold. Participants rated the tool highest on alignment with the sliders (M = 4.4/5) and satisfaction (M = 4.3), and lowest on believability (M = 3.25). In open-ended answers, participants consistently preferred the rival they had themselves shaped through the sliders and described AI-only drafts as "formulaic" or "lacking humanity." We argue that the sliders work less as parameters and more as a shared language that carries user taste into the draft.