Financial technology is playing an increasingly vital role in loan decision-making, and financial institutions are increasingly relying on machine learning techniques to support credit decisions. The purpose of this review is to provide a critical overview of analysis comparing the practical applicability of deep neural network (DNN) and logistic regression (LR) models within the credit scoring domain. This paper systematically collects existing studies on the application of DNN and LR models in credit scoring. The research objects include DNN and LR models, as well as their improved variants developed on the original model frameworks. On this basis, it integrates theoretical research findings with comprehensive analyses to investigate and evaluate the practicality of DNN models. The research results indicate that DNN models still exhibit significant limitations in credit scoring applications. Further model improvements or hybrid integration with other models are therefore required to enhance their practical applicability in real-world scenarios.

Spike sorting is a fundamental step in extracellular neural signal analysis, but educational implementations are often difficult to inspect and reproduce because processing assumptions, parameter choices, and evaluation procedures are distributed across multiple stages. This paper constructs a transparent and reproducible baseline workflow for simulated neural signals rather than proposing a new sorting algorithm. The workflow integrates band-pass filtering, threshold-based spike detection, fixed-window waveform extraction, principal component analysis, KMeans clustering, one-to-one temporal matching, and automated metric export in a configurable Python implementation. Experiments are conducted on repository-generated synthetic datasets with easy, medium, and hard noise settings and on the Wave_Clus Easy1 noise series. Detection F1-score decreases from 0.812 to 0.457 as the synthetic setting becomes harder, while Wave_Clus Easy1 results range from 0.340 to 0.907 across noise conditions. The low clustering scores observed in several settings further show that accurate event detection does not necessarily imply reliable unit separation. The study contributes an auditable baseline, a reproducible evaluation process, and a teaching-oriented example that makes the distinction between detection and clustering performance explicit.

Dust storms develop rapidly and can affect transport, ecosystems, and public health over large areas. The integration of satellite remote sensing and artificial intelligence (AI) has significantly improved capabilities in dust storm identification, parameter retrieval, and short-term forecasting. This paper reviews how AI algorithms have been applied to remote sensing monitoring of dust storms. The review first summarizes the remote sensing response characteristics of dust aerosols and the primary data sources. It then focuses on machine learning-based pixel classification and parameter retrieval, deep learning-based dust mask segmentation and short-term forecasting, and the application of physics-guided hybrid methods in dust storm monitoring. The review also discusses the applications of these dust monitoring products and it points out several remaining problems, including identification instability in complex environments, lack of labeled samples, the underestimation of extreme dust events, and the lack of physical consistency in short-term forecasts. Future studies should focus on improving joint observations of multi-source data, building cross-regional datasets, expanding the pool of extreme event samples, and developing interpretable short-term forecasting models. Such progress will help improve the quantification, continuous observation and practical application of AI-based dust storm remote sensing monitoring.

The rapid development of brain-computer interfaces has made memory uploading, brain controlling of some basic functions of computers, and the treatment of diseases or disabilities like paralysis and blindness possible. However, human consciousness cannot be fully replicated. This paper analyzes why brain-computer interface (BCI) cannot completely replicate human consciousness via a literature review method from three aspects: the limitations of BCI technology, the non-replicability of consciousness, and the challenges of interdisciplinary integration. Research shows that technical issues such as low signal precision and data scarcity restrict its reliability, subjective experiences and nonlinear neural activities of consciousness are difficult to simulate by programs, and that decoding neural signals is not equivalent to reproducing consciousness. Existing theories of consciousness lack engineerable equations, so human consciousness cannot be replicated completely under current situation.

With the advancement of artificial intelligence and robotics, indoor service robots are gradually being deployed in complex environments including residences and hospitals to perform diverse tasks. However, relying solely on traditional localization and obstacle avoidance capabilities can no longer meet the complex demands of real-world scenarios. Robots must possess semantic-level understanding of environmental objects, spatial relationships, and service objectives. Currently, environmental representation for indoor robots is undergoing a transition from being geometry-dominated to semantic-enhanced. This study centers on scene perception and task semantic understanding for indoor service robots. It aims to explore how robots build a systematic cognition of scenes, objects, user commands and task workflows based on low-level visual and spatial data. Using literature review and case analysis methods, this paper synthesizes representative achievements in fields such as semantic mapping, semantic navigation, semantic SLAM, explicit knowledge representation, and task planning. The study argues that scene perception and task semantic understanding constitute a continuous intelligent chain from "environment recognition" to "task execution". In the future, indoor service robots require further improvements in multimodal information fusion, knowledge-driven modeling and task reasoning for open scenarios, so as to enhance their practical performance and operational reliability.

Foldable active-matrix organic light-emitting diode (AMOLED) displays are now widely used in flexible electronic products, but repeated bending and heat generated during operation can induce thermal-mechanical coupling stress (TMCS). This stress causes threshold-voltage drift in thin-film transistors (TFTs), leading to uneven luminance, flicker, afterimages, and reduced device reliability. Existing reviews mainly focus on rigid displays or individual circuit topologies and rarely compare compensation schemes under the specific stress conditions of foldable panels. This paper reviews three mainstream pixel compensation routes: in-pixel compensation circuits, external compensation in display driver integrated circuits (DDICs), and integrated self-compensation based on ferroelectric TFTs. A multi-dimensional comparison is developed from the perspectives of compensation accuracy, pixel aperture ratio, power consumption, manufacturing cost, high-refresh-rate adaptability, and bending durability. The paper also discusses the gap between current industrial requirements and available technologies, with attention to AI-adaptive DDICs and intelligent compensation strategies for next-generation flexible display systems. The review provides a concise technical reference for the design and optimization of reliable foldable AMOLED driver circuits.

In recent years, deep learning algorithm models have gradually been applied to fields such as stock prediction, asset pricing, and portfolio management. Among them, Long Short-Term Memory (LSTM) has been widely used to predict stock prices, returns, and directional movements due to its ability to handle long-term dependencies in time series. Compared to traditional linear financial models, LSTM can better capture non-linear correlations and dynamic fluctuations embedded in financial data, which explains its significant attention in financial forecasting research. However, from a practical investment perspective, a lower prediction error does not necessarily translate into better portfolio performance. This paper reviews the application of LSTM in stock prediction, long-short strategies, and market-neutral portfolio construction. We first introduce the theoretical foundations of LSTM, alpha signals, long-short strategies, and market-neutral portfolios. Subsequently, it delineates the complete research pipeline spanning stock price forecasting, stock alpha ranking and portfolio construction, before analyzing the practical value, existing limitations and prospective research directions of LSTM-based frameworks. Core Viewpoint: The optimal positioning of LSTM is not to predict absolute stock prices, but to serve as an alpha signal generation tool for stock ranking and portfolio optimization. Future research should place more emphasis on point-in-time data, out-of-sample validation, risk neutralization, transaction cost modeling, and comprehensive investment workflow design to improve the usability of LSTM strategies in real-world trading environments.

As integrated circuit scaling approaches atomic limits, traditional process nodes face severe quantum tunneling and thermal constraints, which decelerates the progression of Moore's law. To satisfy high-performance computing demands for greater power efficiency and integration density, Chiplet technology—based on "partition-first, integration-later" and heterogeneous integration—has emerged to extend Moore's Law by improving yield, reducing design complexity, and enabling module reuse. Advanced packaging provides the physical foundation for Chiplets across three dimensions: 2D planar interconnection offers low-cost, mature manufacturing and uses high-density bridges for basic chiplet-to-chiplet electrical connections. 2.5D passive silicon interposers utilize ultra-fine routing and through-silicon vias to bypass conventional layout routing limits, which enhances signal integrity and reduces RC delay. 3D vertical stacking employs advanced hybrid bonding to eliminate solder bumps, which achieves molecular-level integration that minimizes compute-to-memory physical distance while maximizing bandwidth and power efficiency. Finally, applying heterogeneous automotive SoC to intelligent driving systems is critical for integrating multimodal environmental sensing data from HD cameras, LiDAR, and radar. The transition from monolithic ICs to 3D heterogeneous integration represents a key pathway to overcoming physical limitations and realizing highly integrated computing for smart automotive control. Addressing the physical, manufacturing, and latency challenges associated with this integration will provide vital guidelines for future chip design.

Neural networks are often employed in artificial intelligence because they are capable of learning patterns and approximating nonlinear functions from input data. But the performance of a neural network is not only decided by the model structure, but also by the optimization technique utilized in the training process. In this research, we compare full-batch Gradient Descent versus mini-batch Stochastic Gradient Descent in a neural network function approximation problem. A tiny feedforward neural network is trained to approximate the nonlinear function. The experiment is implemented in Python and the comparison is made based on training loss curves, final training means squared error, final test mean squared error and prediction visualization. The results reveal that the full-batch Gradient Descent gives a smoother loss curve since in each update the entire training dataset is used. Mini-batch Stochastic Gradient Descent on the other hand shows more noticeable oscillations but reaches a lower ultimate training MSE and test MSE in this experimental context. This implies that stochastic updates can be less stable at each step but can still assist the model in achieving greater approximation performance in the same number of epochs. The study shows how optimization techniques influence neural network training behavior and it combines mathematical optimization theory with real model performance.

The Internet of Things (IoT) and artificial intelligence have advanced quickly, and with shifting consumer behaviors, the smart home has become the most promising field of IoT application. Rapid growth, however, has exposed persistent weaknesses in interoperability, security, and reliability that slow adoption. This paper analyses the deployment of smart home IoT and the obstacles it faces. The paper first clarifies the concept of the smart home and its underlying IoT infrastructure, then reviews the industry's evolution, and finally assesses the domestic market from four angles: overall scale, penetration rate, major platforms and ecosystems, and typical application scenarios. It then identifies five recurring problems—limited interoperability and ecosystem fragmentation, security and privacy hazards, conflicting automation logic and unreliable operation, a weak user experience, and high cost and energy use alongside lagging standards—and proposes strategies built around unified standards, security-oriented design, conflict detection and mitigation, edge computing with AI, and aging-friendly design. Methodologically, the study combines literature review, case analysis, and industry statistics. It concludes that China's smart home market, though large and fast-growing, remains in a growth phase, and that its tensions are best resolved through stronger standards, a security-first stance, and a systems perspective. Finally, since closed commercial ecosystems provide no usable means of detecting rule conflicts, the study put forward a black-box, log-driven approach to conflict identification and severity assessment that requires neither platform source code nor proprietary interfaces. Unlike techniques that depend on open APIs or runtime instrumentation, it can inform risk governance in multi-brand smart homes and the drafting of technical standards.