Machine unlearning in learned cardinality estimation (CE) systems presents unique challenges due to the complex distributional dependencies in multi-table relational data. Specifically, data deletion, a core component of machine unlearning, faces three critical challenges in learned CE models: attribute-level sensitivity, inter-table propagation and domain disappearance leading to severe overestimation in multi-way joins. We propose Cardinality Estimation Pruning (CEP), the first unlearning framework specifically designed for multi-table learned CE systems. CEP introduces Distribution Sensitivity Pruning, which constructs semi-join deletion results and computes sensitivity scores to guide parameter pruning, and Domain Pruning, which removes support for value domains entirely eliminated by deletion. We evaluate CEP on state-of-the-art architectures NeuroCard and FACE across IMDb and TPC-H datasets. Results demonstrate CEP consistently achieves the lowest Q-error in multi-table scenarios, particularly under high deletion ratios, often outperforming full retraining. Furthermore, CEP significantly reduces convergence iterations, incurring negligible computational overhead of 0.3%-2.5% of fine-tuning time.

While Large Language Models have significantly advanced Text2SQL generation, a critical semantic gap persists: syntactically valid queries can still misinterpret user intent.To mitigate this challenge, we propose GBV-SQL, a multi-agent framework that introduces Guided Generation with SQL2Text Back-translation Validation.In particular, a dedicated validator translates generated SQL back into natural language and checks whether its logic is aligned with the original question.Beyond the method itself, we also conduct a systematic audit of benchmark quality and introduce a typology of "Gold Errors" in Text2SQL datasets.Our analysis shows that benchmark issues can coexist with strong execution accuracy and can substantially affect evaluation outcomes.On the challenging BIRD benchmark, GBV-SQL achieves 63.23% execution accuracy, a 5.8% absolute improvement over the Deepseek-v3-based MAC-SQL setting.Under manually audited benchmark corrections, GBV-SQL reaches 96.5% (dev) and 97.6% (test) on Spider, and 90.42% on repaired BIRD dev, providing a diagnostic view of model behavior under improved gold quality.Our work contributes both a practical framework for semantic validation and an empirical analysis of benchmark integrity in Text2SQL evaluation.

Federated Learning (FL) is typically deployed in a client-server architecture, which makes the Edge-Cloud architecture an ideal backbone for FL. A significant challenge in this setup arises from the diverse data feature distributions across different edge locations (i.e., non-IID data). In response, Personalized Federated Learning (PFL) approaches have been developed. Network segmentation-based PFL is an important approach to achieving PFL, in which the training network is divided into a global segment for server aggregation and a local segment maintained client-side. Existing methods determine the segmentation before the training, and the segmentation remains fixed throughout the PFL training. However, our investigation reveals that model representations vary as PFL progresses and the fixed segmentation may not deliver best performance across various training settings. To address this, we propose PFed-NS, a PFL framework based on adaptive network segmentation. This adaptive segmentation technique is composed of two elements: a mechanism for assessing divergence of clients’ probability density functions constructed from network layers’ outputs, and a model for dynamically establishing divergence thresholds, beyond which server aggregation is deemed detrimental. Further optimization strategies are proposed to reduce the computation and communication costs incurred by divergence modeling. Moreover, we propose a divergence-based BN strategy to optimize BN performance for network segmentation-based PFL. Extensive experiments have been conducted to compare PFed-NS against recent PFL models. The results demonstrate its superiority in enhancing model accuracy and accelerating convergence.

Deploying quantized deep neural network (DNN) models with resource adaptation capabilities on ubiquitous Internet of Things (IoT) devices to provide high-quality AI services can leverage the benefits of compression and meet multi-scenario resource requirements. However, existing dynamic/mixed precision quantization requires retraining or special hardware, whereas post-training quantization (PTQ) has two limitations for resource adaptation: (i) The state-of-the-art PTQ methods only provide one fixed bitwidth model, which makes it challenging to adapt to the dynamic resources of IoT devices; (ii) Deploying multiple PTQ models with diverse bitwidths consumes large storage resources and switching overheads. To this end, this paper introduces a resource-friendly post-training integer-nesting quantization, i.e., NestQuant, for on-device quantized model switching on IoT devices. The proposed NestQuant incorporates the integer weight decomposition, which bit-wise splits quantized weights into higher-bit and lower-bit weights of integer data types. It also contains a decomposed weights nesting mechanism to optimize the higher-bit weights by adaptive rounding and nest them into the original quantized weights. In deployment, we can send and store only one NestQuant model and switch between the full-bit/part-bit model by paging in/out lower-bit weights to adapt to resource changes and reduce consumption. Experimental results on the ImageNet-1K pretrained DNNs demonstrated that the NestQuant model can achieve high performance in top-1 accuracy, and reduce in terms of data transmission, storage consumption, and switching overheads. In particular, the ResNet-101 with INT8 nesting INT6 can achieve 78.1% and 77.9% accuracy for full-bit and part-bit models, respectively, and reduce switching overheads by approximately 78.1% compared with diverse bitwidths PTQ models. Code:https://github.com/jianhayes/NESTQUANT.

Deploying a single deep neural network (DNN) model on ubiquitous Internet of Things (IoT) devices to provide multiple vision services (e.g., semantic segmentation service, facial attribute recognition service) has attracted significant interest from industry and academia. However, the most relevant studies have two limitations: (i) the multiple services DNN commonly has low performance due to the gradient interference between different vision services; (ii) the multiple services DNN has low inference speed on IoT devices. To this end, this paper introduces a multiple vision services acceleration framework for IoT devices. The proposed MCBNet contains a novel multi-controllable branching structure to control different services explicitly for improving multiple services performance and a service balance layer with the per-channel trainable masks in specific branches for mitigating gradient interference. In deployment, the proposed framework converts the multi-branch topology MCBNet model into the plain topology Rep-MCBNet model by re-parameterization technique for accelerating services inference speed. Experimental results on the CelebA and Cityscapes datasets show that our proposed framework outperforms the state-of-the-art approaches in terms of performance and inference speed. In particular, we deploy the Rep-MCBNet on the laptop with Intel mobile processors demonstrating that the proposed framework achieves 1.9× faster processing frames per second than ETR-NLP, with 1.5× fewer parameters and 1.3× fewer floating-point operations.

Abstract Big data analysis requires the speedup of parallel computing. However, in the virtualized systems, the power of parallel computing is not fully exploited due to the limit of current VMM schedulers. Xen, one of the most popular virtualization platforms, has been widely used by industry to host parallel job. In practice, the virtualized systems are expected to accommodate both parallel jobs and serial jobs, and resource contention between virtual machines results in severe performance degradation of the parallel jobs. Moreover, the physical resource is vastly wasted during the communication process due to the ineffective scheduling of parallel jobs. Unfortunately, the existing schedulers of Xen are initially targeting at serial jobs, which are not capable of correctly scheduling the parallel jobs. This paper presents vChecker, an application-level co-scheduler which mitigates the performance degradation of the parallel job and optimizes the utilization of the hardware resource. Our co-scheduler takes number of available CPU cores in one hand, and satisfies need of the parallel jobs in other hand, which helps the credit scheduler of Xen to appropriately schedule the parallel job. As our co-scheduler is implemented at application level, no modifications on the hypervisor is required. The experimental result shows that the vChecker optimizes the performance of the parallel job in Xen and enhances the utilization of the system.

Abstract In this paper, we investigate the influential factors that impact on the performance when the tasks are co-running on a multicore computers. Further, we propose the machine learning-based prediction framework to predict the performance of the co-running tasks. In particular, two prediction frameworks are developed for two types of task in our model: repetitive tasks (i.e., the tasks that arrive at the system repetitively) and new tasks (i.e., the task that are submitted to the system the first time). The difference between which is that we have the historical running information of the repetitive tasks while we do not have the prior knowledge about new tasks. Given the limited information of the new tasks, an online prediction framework is developed to predict the performance of co-running new tasks by sampling the performance events on the fly for a short period and then feeding the sampled results to the prediction framework. We conducted extensive experiments with the SPEC2006 benchmark suite to compare the effectiveness of different machine learning methods considered in this paper. The results show that our prediction model can achieve the accuracy of 99.38% and 87.18% for repetitive tasks and new tasks, respectively.

Training Deep Neural Network is a computation-intensive and time-consuming task. Asynchronous Stochastic Gradient Descent (ASGD) is an effective solution to accelerate the training process since it enables the network to be trained in a distributed fashion, but with a main issue of the delayed gradient update. A recent notable work called DC-ASGD improves the performance of ASGD by compensating the delay using a cheap approximation of the Hessian matrix. DC-ASGD works well with a short delay; however, the performance drops considerably with an increasing delay between the workers and the server. In real-life large-scale distributed training, such gradient delay experienced by the worker is usually high and volatile. In this paper, we propose a novel algorithm called LC-ASGD to compensate for the delay, basing on Loss Prediction. It effectively extends the tolerable delay duration for the compensation mechanism. Specifically, LC-ASGD utilizes additional models that reside in the parameter server and predict the loss to compensate for the delay, basing on historical losses collected from each worker. The algorithm is evaluated on the popular networks and benchmark datasets. The experimental results show that our LC-ASGD significantly improves over existing methods, especially when the networks are trained with a large number of workers.

Summary Cloud computing emerges as one of the most important technologies for interconnecting people and building the so‐called Internet of People (IoP). In such a cloud‐based IoP, the virtualization technique provides the key supporting environments for running the IoP jobs such as performing data analysis and mining personal information. Nowadays, energy consumption in such a system is a critical metric to measure the sustainability and eco‐friendliness of the system. This paper develops three power‐aware scheduling strategies in virtualized systems managed by Xen, which is a popular virtualization technique. These three strategies are the Least performance Loss Scheduling strategy, the No performance Loss Scheduling strategy, and the Best Frequency Match scheduling strategy. These power‐aware strategies are developed by identifying the limitation of Xen in scaling the CPU frequency and aim to reduce the energy waste without sacrificing the jobs running performance in the computing systems virtualized by Xen. Least performance Loss Scheduling works by re‐arranging the execution order of the virtual machines (VMs). No performance Loss Scheduling works by setting a proper initial CPU frequency for running the VMs. Best Frequency Match reduces energy waste and performance loss by allowing the VMs to jump the queue so that the VM that is put into execution best matches the current CPU frequency. Scheduling for both single core and multicore processors is considered in this paper. The evaluation experiments have been conducted, and the results show that compared with the original scheduling strategy in Xen, the developed power‐aware scheduling algorithm is able to reduce energy consumption without reducing the performance for the jobs running in Xen. Copyright © 2016 John Wiley & Sons, Ltd.