Singapore Management University

The recent advances of deep learning in various mobile and Internet-of-Things applications, coupled with the emergence of edge computing, have led to a strong trend of performing deep learning inference on the edge servers located physically close to the end devices. This trend presents the challenge of how to meet the quality-of-service requirements of inference tasks at the resource-constrained network edge, especially under variable or even bursty inference workloads. Solutions to this challenge have not yet been reported in the related literature. In the present paper, we tackle this challenge by means of workload-adaptive inference request scheduling: in different workload states, via adaptive inference request scheduling policies, different models with diverse model sizes can play different roles to maintain high-quality inference services. To implement this idea, we propose a request scheduling framework for general-purpose edge inference serving systems. Theoretically, we prove that, in our framework, the problem of optimizing the inference request scheduling policies can be formulated as a Markov decision process (MDP). To tackle such an MDP, we use reinforcement learning and propose a policy optimization approach. Through extensive experiments, we empirically demonstrate the effectiveness of our framework in the challenging practical case where the MDP is partially observable.

Text classification is a fundamental problem in information retrieval with many real-world applications, such as predicting the topics of online articles and the categories of e-commerce product descriptions. However, low-resource text classification, with no or few labeled samples, presents a serious concern for supervised learning. Meanwhile, many text data are inherently grounded on a network structure, such as a hyperlink/citation network for online articles, and a user-item purchase network for e-commerce products. These graph structures capture rich semantic relationships, which can potentially augment low-resource text classification. In this paper, we propose a novel model called Graph-Grounded Pre-training and Prompting (G2P2) to address low-resource text classification in a two-pronged approach. During pre-training, we propose three graph interaction-based contrastive strategies to jointly pre-train a graph-text model; during downstream classification, we explore handcrafted discrete prompts and continuous prompt tuning for the jointly pre-trained model to achieve zero- and few-shot classification, respectively. Moreover, we explore the possibility of employing continuous prompt tuning for zero-shot inference. Specifically, we aim to generalize continuous prompts to unseen classes while leveraging a set of base classes. To this end, we extend G2P2 into G2P2 $^*$ , hinging on a new architecture of conditional prompt tuning. Extensive experiments on four real-world datasets demonstrate the strength of G2P2 in zero- and few-shot low-resource text classification tasks, and illustrate the advantage of G2P2 $^*$ in dealing with unseen classes.

Federated Learning (FL) enables massive local data owners to collaboratively train a deep learning model without disclosing their private data. The importance of local data samples from various data owners to FL models varies widely. This is exacerbated by the presence of noisy data that exhibit large losses similar to important (hard) samples. Currently, there lacks an FL approach that can effectively distinguish hard samples (which are beneficial) from noisy samples (which are harmful). To bridge this gap, we propose the joint Federated Meta-Weighting based Client and Sample Selection (FedMW-CSS) approach to simultaneously mitigate label noise and hard sample selection. It is a bilevel optimization approach for FL client-and-sample selection and global model construction to achieve hard sample-aware noise-robust learning in a privacy preserving manner. It performs meta-learning based online approximation to iteratively update global FL models, select the most positively influential samples and deal with training data noise. To utilize both the instance-level information and class-level information for better performance improvements, FedMW-CSS efficiently learns a class-level weight by manipulating gradients at the class level, e.g. , it performs a gradient descent step on class-level weights, which only relies on intermediate gradients. Theoretically, we analyze the privacy guarantees and convergence of FedMW-CSS. Extensive experiments comparison against eight state-of-the-art baselines on six real-world datasets in the presence of data noise and heterogeneity shows that FedMW-CSS achieves up to 28.5% higher test accuracy, while saving communication and computation costs by at least 49.3% and 1.2%, respectively.

Accurately estimating route travel time is crucial for intelligent transportation systems. Urban road networks and routes can be viewed from spatial and topological perspectives while existing works typically focus on one view and disregard important information from the other perspective. In this paper, we propose, a novel travel time estimation model. It incorporates an alignment-enhanced spatial-topological aware dual transformer model to adaptively incorporate intra-and inter-view features in the route, guided by cross-view location alignment matrices with clear correspondences between locations in two views. Additionally, we propose a sparsity-aware dual-view traffic feature extraction module to effectively capture temporal traffic state changes. Compared to baseline models, demonstrates improved performance on the MAPE and MAE metrics for Chengdu and Shanghai datasets, achieving improvements of 8.32%, 7.03%, 8.06% and 9.51% respectively, validating the effectiveness of in travel time estimation.

Federated recommender systems (FedRSs) effectively tackle the trade-off between recommendation accuracy and privacy preservation. However, recent studies have revealed severe vulnerabilities in FedRSs, particularly against untargeted attacks seeking to undermine their overall performance. Defense methods employed in traditional recommender systems are not applicable to FedRSs, and existing robust aggregation schemes for other federated learning-based applications have proven ineffective in FedRSs. Building on the observation that malicious clients contribute negatively to the training process, we design a novel contribution-aware robust aggregation scheme to defend FedRSs against untargeted attacks, named contribution-aware Bayesian knowledge distillation aggregation (ConDA), comprising two key components for the defense. In the first contribution estimation component, we decentralize the estimation from the server side to the client side and propose an ensemble-based Shapley value to enable the efficient calculation of contributions, addressing the limitations of lacking auxiliary validation data and high computational complexity. In the second contribution-aware aggregation component, we merge the decentralized contributions via a majority voting mechanism and integrate the merged contributions into a Bayesian knowledge distillation aggregation scheme for robust aggregation, mitigating the impact of unreliable contributions induced by attacks. We evaluate the effectiveness and efficiency of ConDA on two real-world datasets from movie and music service providers. Through extensive experiments, we demonstrate the superiority of ConDA over the baseline robust aggregation schemes.

An epicentre of ethnicity‐based identity politics and exclusions, India's northeastern region has long been politically imagined as both a ‘frontier’ and ‘borderland’ for state control, territorialization, extractive economies and (re)developments by colonial and subsequent post‐colonial regimes. In post‐independence India, the region experienced many territorial assertions made on ethnic grounds. With the rise of religio‐national and neoliberal politics, new urban ‘developments’ and connectivity projects in the region have deepened exclusions based on class, caste, ethnic and religious borders. These have, in turn, led to further displacement, enclavization, and ghettoization of minority communities who are deemed ‘non‐native’ along ethnocentric lines. Taking two cases from Shillong and Guwahati, we introduce the concept of entangled exclusions to analyse the layered, intertwined, multi‐scalar nature of divides within these frontier cities. We argue these cities are critical sites for understanding urban geopolitics where concentrated entanglements of exclusions and (re)borderings become manifest as these cities experience new infrastructural investments.

Cancer image analysis is a crucial element in the field of modern oncology, offering insights into tumor characteristics that are vital for accurate diagnosis and treatment planning. Conventional diagnostic techniques are frequently constrained by discrepancies in interpretation and delays in processing. In recent years, deep learning techniques, in particular convolutional neural networks (CNNs), have demonstrated remarkable capabilities in processing large-scale medical images, enabling the rapid detection and classification of cancerous tissue with high accuracy and precision. This paper reviews the application of CNNs in cancer image classification, introduces common CNN models, and discusses advances such as transfer learning and multimodal data fusion. Furthermore, this paper examines the current challenges confronting convolutional neural networks (CNNs) in the domain of cancer classification. It proposes promising avenues for future research and underscores the potential of CNNs in enhancing cancer diagnosis and even revolutionizing the healthcare system.

We develop a rating system to evaluate the quality of individual non-GAAP exclusions. Our perspective is that high-quality exclusions reflect nonrecurring economic transactions, are transitory accounting adjustments, or have little usefulness in forecasting cash flows. We use four approaches to rate exclusions. We evaluate the serial correlation of the exclusion, survey accounting academics’ views, obtain practitioner ratings from the CFA Institute, and identify the exclusions approved by the Chinese securities regulator. A firm’s exclusion quality score is the weighted average rating of its individual exclusions. For our sample of S&P 500 firms, we document that exclusion quality varies by industry, captures trends in non-GAAP reporting, and is reasonably stable at the firm level. To validate the rating, we show that firms with lower exclusion quality scores receive more SEC comment letters, incur more Regulation G violations, exhibit greater analyst forecast dispersion, and have slower price discovery following earnings announcements.

This article compares China’s stance during the UNCLOS negotiations – the starting point of contemporary law of the sea, with its engagement in the latest development of negotiations on the United Nations agreement on biodiversity beyond national jurisdiction (BBNJ). It answers the question, how does China participate in these two important rules-making processes of the international law of the sea? By identifying salient positions China took in each set of lengthy negotiations and explaining the reasons behind, the article also aims to reflect what a rising China may bring to the international legal maritime order in the foreseeable future. The first part of this article, on the nature of China’s engagement in the UNCLOS negotiations, draws on archival study of official records of the UNCLOS III (1973–1982), as has been digitalized by the UN Office of Legal Affairs. The second part examines the period between the adoption of the UNCLOS (1982) and the start of the BBNJ process (2004), paying attention to China’s shifting practice towards the exploration and exploitation of the deep seabed mineral resources, and its concerns over the ratification of the UNCLOS and the 1995 FSA. Then the article focuses on China in the BBNJ negotiations – Working Groups, Preparatory Committee Meetings and Intergovernmental Conferences. Drawing upon the evolution of China’s positions over the past five decades, the article concludes with some insights on the likely future directions and implications of China’s engagement with the international law of the sea.

Context Software development creates and relies on a large volume of information, yet the volume of this information can make it challenging for developers to maintain an overview of all goings-on that a team and external actors contribute to a project. We posit that unexpected or “surprising” events could serve as important signposts amidst this information overload. These unexpected events may indicate underlying anomalies or emergent situations that require immediate attention. To explore this premise, our study leverages the concept of ‘surprisal’ from information theory to identify and quantify these unusual occurrences from the issues and pull requests of popular open-source software repositories. Objective Drawing from a previously published research protocol, our study investigates whether a correlation exists between the ‘surprisal’ of issues and their perceived importance or difficulty within software repositories. Results We performed a comprehensive analysis of approximately two million issues and pull requests, gathered from 1,270 repositories. Their ‘surprisal’ was then examined in relation to several indicative metrics of difficulty and perceived importance. Our results indicate only a weak correlation. This outcome underscores the need for further research to devise more effective strategies for helping developers prioritise issues.

Over the past two decades, deep learning has received tremendous success in developing software systems across various domains. Deep learning frameworks have been proposed to facilitate the development of such software systems, among which, PyTorch and TensorFlow stand out as notable examples. Considerable attention focuses on exploring software engineering practices and addressing diverse technical aspects in developing and deploying deep learning frameworks and software systems. Despite these efforts, little is known about the open-source software communities involved in the development of deep learning frameworks. In this paper, we perform a comparative investigation into the open-source software communities of the two representative deep learning frameworks, PyTorch and TensorFlow. To facilitate the investigation, we compile a dataset of 2,792 and 3,288 code commit authors, along with 9,826 and 19,750 participants engaged in issue events on GitHub, from the two communities, respectively. With the dataset, we first characterize the structures of the two communities by employing four operationalizations to classify contributors into various roles and inspect the contributions made by common contributors across the two communities. We then conduct a longitudinal analysis to characterize the evolution of the two communities across various releases, in terms of the numbers of contributors with various roles and role transitions among contributors. Finally, we explore the causal effects between community characteristics and the popularity of the two frameworks. We find that the TensorFlow community harbors a larger base of contributors, encompassing a higher proportion of core developers and a more extensive cohort of active users compared to the PyTorch community. In terms of the technical background of the developers, 64.4% and 56.1% developers in the PyTorch and TensorFlow communities are employed by the leading companies of the corresponding open-source software projects, Meta and Google, respectively. 25.9% and 21.9% core developers in the PyTorch and TensorFlow communities possess Ph.D. degrees, while 77.2% and 77.7% contribute to other machine learning or deep learning open-source projects, respectively. Developers contributing to both communities demonstrate spatial and temporal similarities to some extent in their pull requests across the respective projects. The evolution of contributors with various roles exhibits a consistent upward trend over time in the PyTorch community. Conversely, a noticeable turning point in the growth of contributors characterizes the evolution of the TensorFlow community. Both communities show a statistically significant decreasing trend in the inflow rates of core developers. Furthermore, we observe statistically significant causal effects between the expansion of communities and retention of core developers and the popularity of deep learning frameworks. Based on our findings, we discuss implications, provide recommendations for sustaining open-source software communities of deep learning frameworks, and outline directions for future research.

The two most common paradigms to identify records of preference in a multi-objective setting rely either on dominance (e.g., the skyline operator) or on a utility function defined over the records’ attributes (typically, using a top- k query). Despite their proliferation, each of them has its own palpable drawbacks. Motivated by these drawbacks, we identify three hard requirements for practical decision support, namely, personalization, controllable output size, and flexibility in preference specification. With these requirements as a guide, we combine elements from both paradigms and propose two new operators, $\mathsf {ORD}$ and $\mathsf {ORU}$ . We present a suite of algorithms for their efficient processing, dedicating more technical effort to $\mathsf {ORU}$ , whose nature is inherently more challenging. Specifically, besides a sophisticated algorithm for $\mathsf {ORD}$ , we describe two exact methods for $\mathsf {ORU}$ , and one approximate. We perform a qualitative study to demonstrate how our operators work, and evaluate the performance of our algorithms against adaptations of previous work that mimic their output.

Members of the WTO have long regarded e-commerce as an issue of lesser priority. Though seeking to tackle e-commerce issues since 1998 with a Work Programme, progress on enacting WTO law on the matter has remained largely stagnant. To combat two decades of relative inaction, the Joint Statement Initiative on e-commerce was introduced in 2017. Now, within a decade, the JSI has drafted and released a finalised agreement which is now looked towards being integrated into the WTO legal framework. This article explores the history of the JSI, why it has been successful, and how it has overcome previous indifference of the WTO on e-commerce issues from the 1998 Work Programme. It will also highlight how the JSI does and should do to remain significant in today’s legal landscape where many plurilateral trade agreements concurrently emerge in tackling e-commerce issues.

This paper considers the problem of deriving heteroskedasticity and autocorrelation robust (HAR) inference about a scalar parameter of interest. The main assumption is that there is a known upper bound on the degree of persistence in data. I derive finite‐sample optimal tests in the Gaussian location model and show that the robustness‐efficiency tradeoffs embedded in the optimal tests are essentially determined by the maximal persistence. I find that with an appropriate adjustment to the critical value, it is nearly optimal to use the so‐called equal‐weighted cosine (EWC) test, where the long‐run variance is estimated by projections onto q type II cosines. The practical implications are an explicit link between the choice of q and assumptions on the underlying persistence, as well as a corresponding adjustment to the usual Student‐t critical value. I illustrate the results in two empirical examples.

The increasing use of earbuds in applications like immersive entertainment and health monitoring necessitates effective implicit user authentication systems to preserve the privacy of sensitive data and provide personalized experiences. Existing approaches, which leverage physiological cues (e.g., jawbone structure) and behavioral cues (e.g., gait), face challenges such as limited usability, high delay and energy overhead, and significant computational demands, rendering them impractical for resource-constrained earbuds. To address these issues, we present LR-Auth, a lightweight, user-friendly implicit authentication system designed for various earbud usage scenarios. LR-Auth utilizes the modulation of sound frequencies by the user's unique occluded ear canal, generating user-specific templates through linear correlations between two audio streams instead of complex machine-learning models. Our prototype, evaluated with 30 subjects under diverse conditions, demonstrates over 99% balanced accuracy with five 100 ms audio segments, even in noisy environments and during music playback. LR-Auth significantly reduces system overhead, achieving a 20 × to 404 × decrease in latency and a 24 × to 410 × decrease in energy consumption compared to existing methods. These results highlight LR-Auth's potential for accurate, robust, and efficient user authentication on resource-constrained earbuds.

Eye-tracking technology has gained significant attention in recent years due to its wide range of applications in human-computer interaction, virtual and augmented reality, and wearable health. Traditional RGB camera-based eye-tracking systems often struggle with poor temporal resolution and computational constraints, limiting their effectiveness in capturing rapid eye movements. To address these limitations, we propose EyeTrAES, a novel approach using neuromorphic event cameras for high-fidelity tracking of natural pupillary movement that shows significant kinematic variance. One of EyeTrAES's highlights is the use of a novel adaptive windowing/slicing algorithm that ensures just the right amount of descriptive asynchronous event data accumulation within an event frame, across a wide range of eye movement patterns. EyeTrAES then applies lightweight image processing functions over accumulated event frames from just a single eye to perform pupil segmentation and tracking (as opposed to gaze-based techniques that require simultaneous tracking of both eyes). We show that these two techniques boost pupil tracking fidelity by 6+%, achieving IoU~=92%, while incurring at least 3x lower latency than competing pure event-based eye tracking alternatives [38]. We additionally demonstrate that the microscopic pupillary motion captured by EyeTrAES exhibits distinctive variations across individuals and can thus serve as a biometric fingerprint. For robust user authentication, we train a lightweight per-user Random Forest classifier using a novel feature vector of short-term pupillary kinematics, comprising a sliding window of pupil (location, velocity, acceleration) triples. Experimental studies with two different datasets (capturing eye movement across a range of environmental contexts) demonstrate that the EyeTrAES-based authentication technique can simultaneously achieve high authentication accuracy (~=0.82) and low processing latency (~=12ms), and significantly outperform multiple state-of-the-art competitive baselines.

Stack Overflow is one of the most influential Software Question & Answer (SQA) websites, hosting millions of programming-related questions and answers. Tags play a critical role in efficiently organizing the contents on Stack Overflow and are vital to support various site operations, such as querying relevant content. Poorly chosen tags often lead to issues such as tag ambiguity and tag explosion. Therefore, a precise and accurate automated tag recommendation technique is needed. Inspired by the recent success of pre-trained models (PTMs) in natural language processing (NLP), we present PTM4Tag+, a tag recommendation framework for Stack Overflow posts that utilize PTMs in language modeling. PTM4Tag+ is implemented with a triplet architecture, which considers three key components of a post, i.e., Title, Description, and Code, with independent PTMs. We utilize a number of popular pre-trained models, including BERT-based models (e.g., BERT, RoBERTa, CodeBERT, BERTOverflow, and ALBERT), and encoder-decoder models (e.g., PLBART, CoTexT, and CodeT5). Our results show that leveraging CodeT5 under the PTM4Tag+ framework achieves the best performance among the eight considered PTMs and outperforms the state-of-the-art Convolutional Neural Network-based approach by a substantial margin in terms of average Precision@k, Recall@k, and F1-score@k (k ranges from 1 to 5). Specifically, CodeT5 improves the performance of F1-score@1-5 by 8.8%, 12.4%, 15.3%, 16.4%, and 16.6%, respectively. Moreover, to address the concern with inference latency, we experimented PTM4Tag+ using smaller PTM models (i.e., DistilBERT, DistilRoBERTa, CodeBERT-small, and CodeT5-small). We find that although smaller PTMs cannot outperform larger PTMs, they still maintain over 93.96% of the performance on average while reducing the mean inference time by more than 47.2%.

This paper explores understudied issues surrounding accessions to shareholder and partnership agreements: the process by which such accessions take effect; the survival of equities following an accession; and the enforcement of a condition for incoming shareholders to have to execute and deliver a deed of accession. Accessions happen extremely often in modern commercial life, which renders surprising the dearth of academic and judicial discussion, but more disconcerting is the unsettledness of some of the complex issues implicated. The repurposing of unilateral contracts to explain how deeds of accession operate is not fully tested in English law; the conception of partial novation as adumbrated in Unitech Global Ltd v Deutsche Bank AG , which is not even law – much less bad law – has already generated academic controversy; and the enforcement of a condition precedent, in the form of prior accession to a shareholder agreement, for registration of membership in a company interacts in an uncertain way with the Companies Act 2006, lending impetus to the adoption of new methods for attaining relief.