Emotion recognition (ER) poses a complex multi-class classification challenge, further complicated by significant class imbalances. In natural dialogue corpora, dominant emotions like neutral are prevalent, while minority emotions such as disgust and fear are notably scarce. This imbalance results in models consistently underperforming on less frequent categories. This paper investigates template-guided prompting as a method to improve long-tail emotion recognition using large language models (LLMs). We employ a unified evaluation framework on the MELD dataset to compare various methods: supervised baselines (TextCNN, BiLSTM), a fine-tuned pre-trained model (BERT-base), and training-free LLM inference (DeepSeek) using three structured prompt templates in both zero-shot and few-shot scenarios (K=1, 3, 5, 10). Our findings demonstrate that template-guided LLM prompting achieves the highest overall performance (Acc=0.6573, Macro-F1=0.5268) and significantly enhances minority-class F1 scores compared to all supervised baselines, without requiring parameter updates. A detailed analysis of hard-sample errors shows that 16.9% of test instances are misclassified by all five models, with minority emotions having hard-sample rates up to 48%. This bias remains even with balanced downsampling (Pearson r=0.986) and is linked to a systematic prediction bias toward the neutral class. These results imply that the difficulties in long-tail ER arise from intrinsic semantic ambiguity rather than just data imbalance, and that structured prompting offers a practical and effective solution for achieving more balanced emotion recognition.

Due to rapid deep learning advances, Generative Adversarial Networks (GANs) are still the best at generating images using computers. When performing object synthesis in complex scenarios, the foreground object is placed among numerous different backgrounds, enabling it to seamlessly blend in while maintaining its shape, lighting, and all surrounding elements unchanged. However, there are still some issues, such as inconsistent perspectives, differences in lighting, occlusions, and background elements, which do indeed affect the realism of their appearance. Therefore, this paper provides a review of GAN methods for generating objects in complex scenarios. By studying the foundational GAN theory and its subsequent improvements such as conditional GANs, attention-based GANs, and composite GAN frameworks, this study investigates the current relevant methods in terms of spatial alignment, illumination coordination, and multi-object coordination. And it also discusses how the authenticity and fit of this situation are achieved through the use of the generative adversarial network architecture and training methods. From the above content, it is seen that aspects like spatial attention, semantic guidance, and separating the foreground from the background are indeed very important. However, there are still limitations in terms of its universality, computational capacity, and stability. Further studies are advised to combine both geometric and hybrid generative networks to enable more realism in the synthesis of complex scenes.

String matching is a fundamental problem in computer science with applications in text processing, bioinformatics, and information retrieval. Over the past decades, a vari- ety of algorithms have been developed; however, selecting an appropriate algorithm for spe- cific applications remains challenging. This paper provides a narrative review of major string matching algorithms, including classical approaches, automata-based methods, bit-parallel techniques, and hashing algorithms, to analyze their performance and applicability. Special attention is given to finite automata, which provide a theoretical framework for various string-matching approaches; however, their limitations, such as high preprocessing cost and memory consumption, are also critically examined. By examining existing literature, this paper identifies key characteristics, strengths, and limitations of different algorithms. The review also compares these approaches from a practical perspective and shows that hybrid algorithms, which combine multiple techniques, often achieve better performance in large-scale data environments. The findings suggest that no single algorithm is optimal for all scenarios, and future research should focus on combining different techniques to improve performance.

As a pivotal IoT application, the hardware reliability of Intelligent Connected Vehicles (ICVs) is crucial for functional safety and autonomous driving. This paper systematically analyzes reliability challenges in ICV computing systems through theoretical and mechanistic approaches. Key issues identified include CPU/GPU overheating, single points of failure, and network latency, which lead to performance degradation, systemic risks, and decision-making interference. To address these, a multi-dimensional enhancement framework is proposed, integrating hardware optimization, software regulation, architectural redundancy, and edge computing. The study establishes a comprehensive reliability system covering thermal management, redundant backup, and latency compensation to ensure stable ICV operations.

Although Large Language Models (LLMs) have made significant progress in the field of natural language processing, the hallucination phenomenon has become a core issue in judging their credibility and practicality, which may cause serious consequences, especially in high-risk fields such as medicine and law. This article systematically explains the causes and current situation of LLM hallucinations and focuses on sorting out the current mainstream methods of research on alleviating hallucinations. First, the definition and classification of hallucinations are provided, and then the generation mechanism of hallucinations is explained from various aspects, such as data deviation and training goals. Then, six effective mitigation strategies on the market are introduced in detail and with emphasis. In addition, this article also introduces the main principle of evaluation and detection effectiveness using annotation to quantify hallucination tendency, and briefly describes three practical methods. Finally, he mentioned the security challenges of fabricated or misleading information generated by AI in legal, regulatory, and governance aspects, and put forward new ideas for looking at the problem of AI illusions in a positive way when looking into future research directions. Generally speaking, the problem of hallucinations in large language models consistently exists, and completely solving the hallucination problem still faces huge challenges.

Embedded control lies at the intersection of elegant control design and strict hardware limitations. Robots, motor drivers and the controllers of industrial machines must operate under conditions such as fixed sampling intervals, constrained computing power and memory, limited computing precision, and sometimes restricted communication conditions. As a result, these limitations can subtly influence closed-loop behavior, thus making deployability as important as performance. This paper explores embedded control algorithms, optimization strategies for implementation, and smart control on embedded platforms. Among the common approaches, algorithm families include classical feedback control, state-space methods with estimation, and robust and adaptive methods for handling uncertainties and system drift. In order to make these methods practical in practice, several implementation strategies are proposed, including offline pre-computation, solver selection in optimization control, fixed-point arithmetic and scaling techniques, and execution methods for energy and memory optimization. Meanwhile, real-time scheduling and time predictability are regarded as core design factors, since delays and jitter directly affect system stability. With regard to concerns such as security during deployment, authentication, and worst-case execution time, this study also discusses rule-based control and learning-based intelligent control methods. Besides, it addresses current challenges and future directions, including hybrid model learning, control-computation integration, and verification for safety-critical embedded systems.

Using phonocardiogram (PCG) analysis to screen congenital heart disease (CHD) at an early stage can help with earlier referral and treatment. This is especially so in low-resource settings where echocardiography is not easily available. Recent transformer-based models have given good results in heart-sound classification, as self-attention is able to capture long-range temporal dependencies across the cardiac cycles. Standard transformers, however, require substantial computation and are hard to deploy on microcontrollers or portable digital stethoscopes, as these have strict limits on memory, power consumption, and latency. This review focuses on lightweight Transformer strategies for embedded CHD screening, such as sparse or linearized attention, CNN-Transformer hybrid architectures, quantization, and knowledge distillation. It also discusses how these methods influence the diagnostic accuracy, robustness to noisy pediatric recordings, model size, and real-time inference. It also identifies several problems that are yet to be solved, such as interpretability, dataset bias, variation in neonatal heart sounds, and the practical difficulty of using multimodal fusion in a point-of-care device. Overall, a lightweight Transformer design provides a path for developing accessible, hardware-efficient AI systems for early CHD screening.

Neural networks have been widely used in image recognition and other practical applications, and have shown excellent performance. However, the problem of insufficient robustness and stability exists in neural networks. Under the disturbance of noise and counter samples, the performance of the model decreases significantly, which affects the safety and reliability of the actual scene. This article focuses on Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM). The application scenarios and core principles of the robust optimization method are combined. Through the comparative analysis of several research cases, this paper obtains the application effect and practical advantages of various robust optimization methods. In this paper, several kinds of robust optimization methods are analyzed combined with defensive distillation, and their limitations in practical application are found. The improvement directions for these limitations are proposed, and the feasibility of these improvement directions is confirmed. By comparing and analyzing the performance of various robust optimization methods in several practical application cases, this paper concludes that these methods have the advantages of resisting noise interference and sample attack. However, there are also some limitations, such as strong specificity of adaptation scenarios, weak cross-domain migration ability, single defense dimension of optimization methods, weak defense board, or inherent limitations of general-purpose defense methods. This paper puts forward some improvement directions, such as using a general regularization method, using an adaptive weighted loss function, using confrontation training for collaborative assistance, and finally demonstrates and analyzes the feasibility of these improvement methods.

In text processing, visual images are usually generated from textual descriptions, and computer systems can perform similar tasks. As a core task in the field of multimodal generation, text-to-image synthesis aims to generate high-quality images that are visually realistic and semantically consistent based on natural language descriptions. Therefore, this paper reviews the application and development of Generative Adversarial Networks (GANs) in text-to-image synthesis, traces their technological evolution, analyzes key breakthroughs, core modules, and current bottlenecks, and proposes future research directions. This paper explores recent literature to outline the practical progress of GANs, covering fundamental principles, technological innovations, and module optimizations. Additionally, it conducts an in-depth analysis of technical bottlenecks such as cross-modal semantic mapping, detail generation, and training stability, and reviews optimization strategies and future research trends. The results indicate that though diffusion models have dominated in recent years in terms of generation quality, handling of complex scenes, and semantic consistency, GANs still perform better in terms of inference speed, fine-grained control, and training efficiency on large-scale datasets. However, GANs face challenges such as training instability, limited comprehension of long texts, attribute binding errors, and insufficient high-resolution detail.

In the field of computer vision, image super-resolution (SR) plays a crucial role as a class of image processing techniques that boost image resolution, with significant research importance and practical applications in medical imaging. Recently, methods based on deep learning have significantly enhanced the super-resolution processing capabilities of medical images, and high-resolution (HR) medical images have become a necessary condition for accurate clinical diagnosis. This paper aims to elaborate on the innovations and progress of deep learning-driven medical image super-resolution technologies. Besides, it reviews the basic theory of SR, outlines the standard metrics for evaluating the performance of super-resolution, discusses the key challenges in medical image super-resolution reconstruction, and points out the possible future development directions. The results show that the medical image super-resolution technology based on deep learning has evolved from convolutional neural networks (CNN) to more advanced architectures such as Transformer, Mamba and Implicit Neural Representation (INR), but due to limited data and insufficient evaluation metrics, challenges remain. Future work should explore unsupervised learning, lightweight networks, and multi-modal fusion for clinical application.