Grading programming assignments manually is a big burden for instructors who teach programming languages for university students due to complexity and subjectivity. The black test approach adopted by online judge systems can only outputs either an answer is correct or incorrect. This study proposes a Large Language Model (LLM) approach to automatically grade answers from students for programming assignments. A LLM mode formed by coder-decoder architecture is utilized to generate summarization from source code, then the summarization is compared to the textual assignment description by semantic similarity. Finally, the output is converted to five-score rating. CodeBERT and a Transformer model serve as coder and decoder respectively. The semantic similarity is computed by MiniLM-L6. The validation test shows that the accuracy of the suggested approach reaches 0.92.
The programmer’s attention have been utilized for the behavioral research in programming education. Existing approaches mainly focused on the programmer’s visual attention measured by gaze trackers. However, for the scenarios of programming and debugging, the measurement on just eye movements is not sufficient for understanding programming attention. The measurements on programmer’s other modals of behaviors, including mouse clicks and keyboard inputs, are also required. In this paper, we focus on the programming attention visualization problem in the debugging test scenario. A series of attention visualization approaches are proposed for the two main types of interactions, i.e. reading source code and editing/debugging by keyboard and mouse. We develop methods to visualize student’s attentions from either reading or debugging interactions, and then merge them into multimodal programming attention. The experiment results show that, the proposed multimodal attention graphs can show the diversity of students’ debugging strategies in terms of reading and editing focus in code line and time dimensions. It can help the teacher infer the students’ behavioral intentions and conduct individual diagnose in debugging tests.
Micro action recognition is an important research area in human motion analysis, which can be applied to the fields of healthcare, motion analysis and human-computer interaction. In this paper, we propose a 3D convolutional neural network-based micro-gesture recognition network to improve the performance, which combines skeletal and semantic embedding losses to enhance network discrimination. Specifically, first, various data enhancement techniques, such as level flipping and random Gaussian noise, are utilized to improve the robustness and generalization of the model. Horizontal flipping aims to help the model better recognise mirrored and stacked features in images by flipping them on the symmetry axis. Meanwhile, Gaussian noise, also known as white noise, is a set of random values with a normal distribution designed to force the model to learn features that are robust to small variations in the input, which can represent smudges or subtle absences in the image. We have evaluated our methods on a micro-posture recognition benchmark dataset, and these methods improve on previous methods. The model was tested on the iMiGUE dataset and achieved a Top1% accuracy of 59.01%-a notable improvement over other models.
Knowledge Tracing (KT) aims to trace students’ knowledge state and predict their future performance from their historical learning records. However, accurately capturing learning process is a challenging due to the complexity of the learning behaviors and learning patterns. In this paper, we propose a Convolution Attentive Knowledge Tracing with comprehensive behavioral features (CAKT). Specially, we propose an ensemble learning-based feature fusion approach that projects learning behavior information, including attempt counts, hint counts, first action, and response time to obtain complex learning behavior features. Furthermore, we propose a convolution attentive layer that mimics the learning patterns of the human brain during learning. This layer utilizes convolution to integrate exercise context information, allowing it to highlight the influence of recent relevant exercises and model forgetting behavior through a temporal decay function. We conduct experiments on several real-world educational datasets and the results demonstrate that CAKT outperforms state-of-the-art knowledge tracing models in predicting student performance. Our work offers an innovative approach to KT that better aligns with the authentic development and formation of learners’ knowledge.
Recent advancements in programming education have transformed it from a niche skill into a fundamental requirement across diverse industries, fostering innovation and problem-solving capabilities. An increasing number of non-computer science(CS) students are now engaging in programming learning. However, they often struggle in early programming courses. Numerous studies have explored the relationship between students' majors and their academic performance to understand the factors influencing their success. Previous studies primarily relied on outcome-based assessments, overlooking the significance of formative evaluations during the learning process. To address this gap, we conducted a study collecting process data from an introductory programming course for freshmen of mixed majors, analyzing learning trajectories using process-oriented methods. Our findings reveal significant differences in score trajectories between CS and non-CS groups, underscoring the need for targeted instructional improvements to enhance non-CS students' learning experiences.
In the era of generative artificial intelligence, the accuracy of data is paramount. Erroneous data often leads to faulty outcomes and economic detriments. Previous cleaning methods employ a sequential detect-repair paradigm, leaving over half of the errors unsolved in real scenarios. We introduce IterClean, an iterative data cleaning framework leveraging large language models (LLMs). Utilizing an iterative mechanism, the framework employs a two-step process: data labeling and iterative data cleaning. With few labeled data, IterClean leverages an iterative cleaning process involving an error detector, an error verifier, and an error repairer to significantly enhance the cleaning performance. Extensive experiments across four datasets demonstrate that, IterClean achieves an F1 score that is up to three times higher than the best state-of-the-art approaches requiring only 5 labeled tuples.
Considerable attention has been devoted to the use of automated assessment tools in introductory programming instruction. However, the efficacy of automated feedback hinges on the quality of the content it provides and the student’s ability to apply this information to their learning. The process of learners’ continuous debugging of their code exemplifies the principles of self-regulated learning (SRL). This study aims to investigate the effectiveness of automated feedback using SRL theory. The findings indicate that there was no difference in code improvement rates between students with different SRL abilities when using automated feedback. However, automated feedback was effective in helping students debug common compilation errors, which has significant implications for improving the efficiency of programming education. We recommend that researchers further design the feedback content and the presentation to investigate the connection between SRL, feedback usage, and the programming learning process.
In this paper, a newdataset is constructed for intelligent multi-modal sensing-communication integration in low-altitude scenarios. To construct the dataset, we utilize AirSim and WaveFarer to acquire multi-modal sensing data and exploit Wireless InSite to acquire communication data. Currently, the constructed dataset covers two typical scenarios, i.e., urban crossroad and mountain road, and contains 3,000 RGB images, 3,000 depth maps, 30,000 light detection and ranging (LiDAR) point clouds, 30,000 millimeter wave (mmWave) radar point clouds, 30,000 channel impulse response (CIR) matrices, and 30,000 path loss values.
The growing demand for memory systems with larger capacities and faster data transfer speeds has driven progress in the widespread adoption of multi-socket machines and memory expansion through Compute eXpress Link (CXL). However, processes running on such multi-socket machines suffer non-uniform bandwidth and latency when accessing physical memory. Despite prior efforts to propose data allocation and placement strategies in NUMA environments over the years, they still fall short due to the semantic gap between the process scheduling and memory access pattern - the process scheduler has limited knowledge of its running processes' memory access latency. Actually, the latency of memory access is influenced not only by the distance between NUMA nodes but also by the memory bandwidth pressure, especially in scenarios involving co-located workloads. We propose Tiresias, a feedback-based controller that migrates NUMA effects on data access latency by transparently employing memory locality-aware process scheduling and provisioning differentiated memory bandwidth allocations with assistance from CXL memory. Tiresias exploits multiple resource optimization techniques, including (1) workload-aware and software-based memory bandwidth management, (2) a memory page migration strategy to alleviate memory bandwidth contention by leveraging CXL memory, and (3) page-table self-replication (PTSR) based locality-aware process scheduling. To evaluate the impact of Tiresias on performance, we conduct an analysis that focuses on the temporal and spatial correlation of memory access patterns.
Pedestrians usually change their clothing features in real scenarios due to external factors (e.g., weather, seasons, etc.), which makes methods that rely on consistent clothing features no longer applicable. A knowledge-driven cross-period network for clothes changing Re-ID is proposed in this paper, which includes: (1) A knowledge-driven topology inference network is proposed, which adopts knowledge graph and graph convolution network to capture spatio-temporal information between camera nodes, and finally introduces knowledge embedding into the graph convolution network for topology inference. (2) A cross-period clothing change network based on spatio-temporal information aggregation is designed for clothing generation. The overall pedestrian clothing characteristics under the logical topology cameras are utilized as auxiliary information to achieve matching of pedestrians, thus effectively reducing the matching error caused by external factors. (3) A joint optimization mechanism is proposed. Multi-camera logical topology provides auxiliary information and retrieval order for the cross-period clothing change network. Meanwhile, pedestrian re-identification results are utilized as feedback to modify the logical topology.
Non-partitioned hash join on GPU suffers from significant overhead from two types of synchronizations, i.e., (i) synchronizing concurrent writes in build phase and (ii) memory stall of synchronous random memory access in probe phase. To address these problems, we introduce a synchronization-free hash join implementation. The hash table is constructed in a lock-free way in build phase and searched with asynchronous memory access in probe phase. Through these innovations, we demonstrate a substantial improvement in hash join performance on GPUs, achieving up to a 1.19x speedup compared to traditional approaches.
Multi-view spectral clustering (MVSC) is a typical unsupervised data analysis method in literature. It aims to integrate the complementary information of different data views for higher clustering accuracy. Under the assumption that all views share a unified clustering structure, the base Laplacian matrices can be regarded as different perturbations of a consensus Laplacian matrix. On this basis, a set of practical MVSC algorithms are designed. However, almost all of them suffer from high computational complexity due to the construction and processing of the matrices whose size is square of sample number. To address the above issues, we propose a scalable multi-view spectral clustering based on spectral perturbation theory to handle large-scale datasets and learn diverse information from each base views. Specifically, we first construct bipartite graphs for all base views and aim to learn a consensus Laplacian matrix of these base bipartite graphs. Based on a perturbation theory of singular subspace, we design an objective function that can minimize the discrepancy of the canonical angles between the consensus Laplacian matrix and base Laplacian matrices. Moreover, we impose a matrix-induced regularization item to increase the diversity of base views. We design a simple but efficient method to solve the resultant problem. The complexity of the optimization method can be proven linear to the sample number. Finally, we conduct extensive experiments in benchmark datasets to verify the efficiency and effectiveness of the proposed method.
Applying existing face swapping algorithms independently to each video frame typically leads to temporal inconsistency. We analyze the inconsistency in the generated results and model inter-frame inconsistency as time-domain noise. We propose a face swapping mapper network to inherit identity and suppress noise. Training strategies include primary perceptual loss to learn the face swapping information of the reference face, optical flow loss to impose temporal constraints, and identity loss to transfer identity information. In addition, we introduce a 3D face disentanglement model to regress FLAME parameters and guide the optimization direction precisely for facial detail consistency. Only a pair of original and swapped videos is used for training, eliminating the need for a large dataset. Experiments demonstrate that we improve the temporal consistency and detail consistency of the results, and enhance the generation quality of face swapping methods at the video level.
A scientific disciplinary assessment system is crucial for nurturing high-quality disciplines within Computer Science. Computer Science Education (CSE) emphasizes the need for a scientific and comprehensive assessment method that guides the development of the discipline, with a particular focus on practical contributions. However, traditional assessment systems tend to prioritize the theoretical outcomes. Moreover, data expansion demands significant effort and time from educational professionals, making it challenging to conduct a thorough evaluation of the disciplines. To tackle these issues, we introduce a data-driven, contribution-based disciplinary assessment system. This system takes into account both theoretical and practical contributions to provide a holistic evaluation. Our proposed system employs a contribution-based assessment approach to establish a correct evaluative direction, steering discipline construction to align with societal needs. It also incorporates intelligent algorithms and a Large Language Model (LLM), leveraging their substantial computational power in the evaluation process. This integration alleviates the workload of educational professionals by automating the collection and analysis of information. The paper outlines a detailed implementation plan that integrates contribution evaluation theory with intelligent technologies, aiming to foster the ongoing advancement of CSE education.
Hash tables can efficiently determine whether an element exists in a given set and have been widely used in computer networks, the Internet of Things (IoT), data centers, and stream data mining. With the continuous generation of massive data, the memory consumption of hash tables keeps increasing. The emerging Compute Express Link (CXL) technique can significantly expand memory capacity. Porting hash tables from DRAM to CXL memory can alleviate the issue that hash tables occupy significant amounts of DRAM space. However, porting hash tables to CXL memory is not a trivial task. This paper analyzes the challenges of porting hash tables to CXL memory and shows opportunities to address these challenges.
The emergence of large language models (LLMs) offers a new opportunity to build LLMs-based applications, such as smart home, as these models have demonstrated general-purpose language understanding by generating coherent and contextually relevant text. However, LLMs are trained on massive amounts of text data to predict tokens, so these models have limitations and it is difficult for them performing physical world tasks directly. To further exploit the potential of LLMs to solve the challenge of integrating them with the physical world, LLMs enhanced and augmented techniques should be addressed, especially reinforcement learning based techniques. In this paper, we study the issue of integrating LLMs with physical world. We first describe the large language models and limitations. Then, we revisit LLMs enhanced and augmented techniques. After that, we present methods of interaction LLMs with physical world, such as integration IoT sensing with LLMs, embodied agent post-training with LLMs, and robot task planning with LLMs. Finally, we provide a case study of smart home powered by LLMs to discuss future research directions of next-generation intelligent smart home, personal health assistant, and LLM-based household robot.
With the expanding mileage of China’s railroads, efficiently managing and maintaining railroad operation data, particularly for heavy railroads, has become increasingly crucial. However, the mainstream railroad data management system is based on the “server-client" Web 2.0 model, which has some unsolvable security problems. This paper introduces the advantages of the blockchain-based Web 3.0 railroad data management system in managing and maintaining railroad facilities and operation data, such as decentralization to improve security, comparability to ensure data security, and traceability to support data traceability. Additionally, this paper introduces a grade recommendation mechanism for evaluating and recommending various safety and performance indicators for heavy railroads and ports to ensure that the operation of each link meets safety standards.
As deep learning progresses, programming language generation models such as CodeLlama, GitHub Copilot, and ChatGPT have been widely applied to intelligent code development. However, this also reduces the cost of code plagiarism, posing challenges to copyright and academic integrity. In response to the specific needs for human-machine code detection, this paper introduces a comprehensive automated benchmark CodeWMBench for active detection of human-machine code through watermarking. With a meticulous evaluation of eight code watermarking methods, we demonstrated their performance in terms of harmlessness, robustness, and transparency. Specifically, for the first time, we introduced watermark removal techniques based on large language models and conducted the first assessment of these watermarking methods against code rewriting and retranslating attacks. In the discussion, we delved into the critical issues currently facing code watermarking, including why existing code watermarking methods struggle to resist removal by large language models and potential future methods that could withstand such removals.
The data quality is a critical factor that affects the accuracy of feature selection algorithms. Meanwhile, structured knowledge information is embedded in complex feature correlations. However, the uncertainty of data and the complex correlation of features pose challenges to feature selection. For these issues, we propose a Multi-Correlation Collaborative Computation for Feature Selection with Uncertainty Measures in the fuzzy rough set (MCCFS). The proposed MCCFS algorithm is compared to some representative feature selection algorithms on sixteen datasets. The experimental results and an ablation study demonstrate that MCCFS achieves significant superiority compared to the comparative algorithms due to the consideration of multi-correlation collaboration.
As a new communication protocol, Remote Direct Memory Access (RDMA), although providing satisfactory average latency, greatly reduces tail bandwidth fluctuations caused by frequency and density levels. In this paper, we propose a dual-terminal collaborative congestion control strategy where both the network source and node participate in control simultaneously, to solve the problem of random delays in data flow transmission that prevent network nodes from achieving stable control. Experimental results show that the proposed congestion control has higher throughput.
