Federated continual learning (FCL) has attracted growing attention in achieving collaborative model training among edge clients, each of which learns its local model for a sequence of tasks. Most existing FCL approaches aggregate clients' latest local models to exchange knowledge. This unfortunately deviates from real-world scenarios where each model is optimized independently using the client's own dynamic data and different clients have heterogeneous tasks. These tasks not only have distinct class labels (e.g. animals or vehicles) but also differ in input feature distributions. The aggregated model thus often shifts to a higher loss value and incurs accuracy degradation. In this paper, we depart from the model-grained view of aggregation and transform it into multiple task-grained aggregations. Each aggregation allows a client to learn from other clients to improve its model accuracy on one task. To this end, we propose Loci to provide abstractions for clients' past and peer task knowledge using compact model weights, and develop a communication-efficient approach to train each client's local model by exchanging its tasks' knowledge with the most accuracy relevant one from other clients. Through its general-purpose API, Loci can be used to provide efficient on-device training for existing deep learning applications of graph, image, nature language processing, and multimodal data. Using extensive comparative evaluations, we show Loci improves the model accuracy by 32.48% without increasing training time, reduces communication cost by 83.6%, and achieves more improvements when scale (task/client number) increases.

When running edge intelligence applications with 6G networks, model pipeline effectively reduces inference latency via parallelizing layers across multiple edge devices. Today’s edge inference systems usually employ static architecture of layers in pipeline parallelism but dynamically skip part of layers in early-exit, which may significantly degrade system throughput. In this paper, we introduce DensePipe, an online layer scheduling approach that optimally allocates early-exit layers to edge devices to maximize their throughput in model pipeline. To this end, DensePipe profiles all network layers’ skipping probabilities in early-exit. At run-time, DensePipe maximizes the pipeline throughput by balancing the processing of all unskipped layers among devices according to the current loads and device resource utilizations. We implement DensePipe with Transformer models and demonstrate its effectiveness against state-of-the-art pipeline methods. Comparative experiments show that DensePiple successfully finds the best devices for most of the layers and significantly improves throughput by 3.09x.

Foundation models (FMs) such as large language models are becoming the backbone technology for artificial intelligence systems. It is particularly challenging to deploy multiple FMs on edge devices, which not only have limited computational resources, but also encounter unseen input data from evolving domains or learning tasks. When new data arrives, existing prior art of FM mainly focuses on retraining compressed models of predetermined network architectures, limiting the feasibility of edge devices to efficiently achieve high accuracy for FMs. In this paper, we propose EdgeTA, a neuron-grained FM scaling system to maximize the overall accuracy of FMs promptly in response to their data dynamics. EdgeTA's key design features in scaling are (i) proxy mechanism, which adaptively transforms a FM into a compact architecture retaining the most important neurons to the input data, and (ii) neuron-grained scheduler, which jointly optimizes model sizes and resource allocation for all FMs on edge devices. Under tight retraining window and limited device resources, the design of EdgeTA can achieve most of the original FM's accuracy with much smaller retraining costs. We implement EdgeTA on FMs of natural language processing, computer vision and multimodal applications. Comparison results against state-of-the-art techniques show that our approach improves accuracy by 21.88% and reduces memory footprint and energy consumptions by 27.14% and 65.65%, while further achieving 15.96% overall accuracy improvement via neuron-grained scheduling.

Edge applications are increasingly empowered by deep neural networks (DNN) and face the challenges of adapting or retraining models for the changes in input data domains and learning tasks. The existing techniques to enable DNN retraining on edge devices are to configure the memory-related hyperparameters, termed m-hyperparameters, via batch size reduction, parameter freezing, and gradient checkpoint. While those methods show promising results for static DNNs, little is known about how to online and opportunistically optimize all their m-hyperparameters, especially for retraining tasks of edge applications. In this paper, we propose, MPOptimizer, which jointly optimizes an ensemble of m-hyperparameters according to the input distribution and available edge resources at runtime. The key feature of MPOptimizer is to easily emulate the execution of retraining tasks under different m-hyperparameters and thus effectively estimate their influence on task performance. We implement MPOptimizer on prevalent DNNs and demonstrate its effectiveness against state-of-the-art techniques, i.e. successfully find the best configuration that improves model accuracy by an average of 13% (up to 25.3%) while reducing memory and training time by 4.1x and 5.3x under the same model accuracies.

Deep Neural Networks (DNNs) have been ubiquitously adopted in internet of things and are becoming an integral part of our daily life. When tackling the evolving learning tasks in real world, such as classifying different types of objects, DNNs face the challenge to continually retrain themselves according to the tasks on different edge devices. Federated continual learning (FCL) is a promising technique that offers partial solutions but yet to overcome the following difficulties: the significant accuracy loss due to the limited on-device processing, the negative knowledge transfer caused by the limited communication of non-IID (non-Independent and Identically Distributed) data, and the limited scalability on the tasks and edge devices. Moreover, existing FCL techniques are designed for convolutional neural networks (CNNs), which have not utilized the full potential of newly emerged powerful vision transformers (ViTs). Considering ViTs depend heavily on training data diversity and volume, we hypothesize ViTs are well-suited for FCL where data arrives continually. In this paper, we propose FedViT, an accurate and scalable federated continual learning framework for ViT models, via a novel concept of signature task knowledge. FedViT is a client-side solution that continuously extracts and integrates the knowledge of signature tasks which are highly influenced by the current task. Each client of FedViT is composed of a knowledge extractor, a gradient restorer and, most importantly, a gradient integrator. Upon training for a new task, the gradient integrator ensures the prevention of catastrophic forgetting and mitigation of negative knowledge transfer by effectively combining signature tasks identified from the past local tasks and other clients’ current tasks through the global model. We implement FedViT in PyTorch and extensively evaluate it against state-of-the-art techniques using popular federated continual learning benchmarks. Extensive evaluation results on heterogeneous edge devices show that FedViT improves model accuracy by 88.61% without increasing model training time, reduces communication cost by 61.55%, and achieves more improvements under difficult scenarios such as large numbers of tasks or clients, and training different complex ViT models.

Executing deep neural networks (DNN) based vision tasks on edge devices encounters challenging scenarios of significant and continually evolving data domains (e.g. background or subpopulation shift). With limited resources, the state-of-the-art domain adaptation (DA) methods either cause high training overheads on large DNN models, or incur significant accuracy losses when adapting small/compressed models in an online fashion. The inefficient resource scheduling among multiple applications further degrades their overall model accuracy. In this paper, we present ElasticDNN, a framework that enables online DNN remodeling for applications encountering evolving domain drifts at edge. Its first key component is the master-surrogate DNN models, which can dynamically generate a small surrogate DNN by retaining and training the large master DNN's most relevant regions pertinent to the new domain. The second novelty of ElasticDNN is the filter-grained resource scheduling, which allocates GPU resources based on online accuracy estimation and DNN remodeling of co-running applications. We fully implement ElasticDNN and demonstrate its effectiveness through extensive experiments. The results show that, compared to existing online DA methods using the same model sizes, ElasticDNN improves accuracy by 23.31% and reduces adaption time by 35.67x. In the more challenging multi-application scenario, ElasticDNN improves accuracy by an average of 25.91%.

AI applications powered by deep learning models are increasingly run natively at edge. A deployed model not only encounters continuously evolving input distributions (domains) but also faces adversarial attacks from third-party. This necessitates adapting the model to shifting domains to maintain high natural accuracy, while avoiding degrading the model's robust accuracy. However, existing domain adaptation and adversarial attack preventation techniques often have conflicting optimization objectives and they rely on time-consuming training process. This paper presents RobustDA, an on-device lightweight approach that co-optimizes natural and robust accuracies in model retraining. It uses a set of low-rank adapters to retain all learned domains' knowledge with small overheads. In each model retraining, RobustDA constructs an adapter to separate domain-related and robust-related model parameters to avoid their conflicts in updating. Based on the retained knowledge, it quickly generates adversarial examples with high-quality pseudo-labels and uses them to accelerate the retraining process. We demonstrate that, comparing against 14 state-of-the-art DA techniques under 7 prevalent adversarial attacks on edge devices, the proposed co-optimization approach improves natural and robust accuracies by 6.34% and 11.41% simultaneously. Under the same accuracy, RobustDA also speeds up the retraining process by 4.09x.

Edge-cloud applications are rapidly prevailing in recent years and pose the challenge of using both resource-strenuous edge devices and elastic cloud resources under dynamic workloads. Efficient resource allocation on edge-cloud jobs via cluster schedulers (e.g. Kubernetes/Volcano scheduler) is essential to guarantee their performance, e.g. tail latency, and such allocation is sensitive to scheduler configurations such as applied scheduling algorithms and task restart/discard policy. Deep reinforcement learning (DRL) is increasingly applied to optimize scheduling decisions. However, DRL faces the conundrum of achieving high rewards at a dauntingly long training time (e.g. hours or days), making it difficult to tune the scheduler configurations online in accordance to dynamically changing edge-cloud workloads and resources. For such an issue, this paper proposes EdgeTuner, a fast scheduler configuration tuning approach that efficiently leverages DRL to reduce tail latency of edge-cloud jobs. The enabling feature of EdgeTuner is to effectively simulate the execution of edge-cloud jobs under different scheduler configurations and thus quickly estimate these configurations’ influence on job performance. The simulation results allow EdgeTuner to timely train a DRL agent in order to properly tune scheduler configurations in dynamic edge-cloud environment. We implement EdgeTuner in both Kubernetes and Volcano schedulers and extensively evaluate it on real workloads driven by Alibaba production traces. Our results show that EdgeTuner outperforms prevailing scheduling algorithms by achieving much lower tail latency while accelerating DRL training speed by an average of 151.63x.

Deep Neural Networks (DNNs) have been ubiquitously adopted in internet of things and are becoming an integral of our daily life. When tackling the evolving learning tasks in real world, such as classifying different types of objects, DNNs face the challenge to continually retrain themselves according to the tasks on different edge devices. Federated continual learning is a promising technique that offers partial solutions but yet to overcome the following difficulties: the significant accuracy loss due to the limited on-device processing, the negative knowledge transfer caused by the limited communication of non-IID data, and the limited scalability on the tasks and edge devices. In this paper, we propose FedKNOW, an accurate and scalable federated continual learning framework, via a novel concept of signature task knowledge. FedKNOW is a client side solution that continuously extracts and integrates the knowledge of signature tasks which are highly influenced by the current task. Each client of FedKNOW is composed of a knowledge extractor, a gradient restorer and, most importantly, a gradient integrator. Upon training for a new task, the gradient integrator ensures the prevention of catastrophic forgetting and mitigation of negative knowledge transfer by effectively combining signature tasks identified from the past local tasks and other clients' current tasks through the global model. We implement FedKNOW in PyTorch and extensively evaluate it against state-of-the-art techniques using popular federated continual learning benchmarks. Extensive evaluation results on heterogeneous edge devices show that FedKNOW improves model accuracy by 63.24% without increasing model training time, reduces communication cost by 34.28%, and achieves more improvements under difficult scenarios such as large numbers of tasks or clients, and training different complex networks.

Vision applications powered by deep neural networks (DNNs) are widely deployed on edge devices and solve the learning tasks of incoming data streams whose class label and input feature continuously evolve, known as domain shift. Despite its prominent presence in real-world edge scenarios, existing benchmarks used by domain adaptation methods overlook evolving domains and under represent their shifts in label and feature distributions. To address this gap, we present EdgeVisionBench, a benchmark seeking to generate evolving domains of various types and reflect their realistic label and feature shifts encountered by edge-based vision applications. To facilitate evaluating domain adaptation methods on edge devices, we provide an open-source package that automates workload generation, contains popular DNN models and compression techniques, and standardizes evaluations with interactive interfaces. Code and datasets are available at https://github.com/LINC-BIT/EdgeVisionBench.

Edge-cloud jobs are rapidly prevailing in many application domains, posing the challenge of using both resource-strenuous edge devices and elastic cloud resources. Efficient resource allocation on such jobs via scheduling algorithms is essential to guarantee their performance, e.g. latency. Deep reinforcement learning (DRL) is increasingly adopted to make scheduling decisions but faces the conundrum of achieving high rewards at a low training overhead. It is unknown if such a DRL can be applied to timely tune the scheduling algorithms that are adopted in response to fast changing workloads and resources. In this paper, we propose EdgeTuner to effectively leverage DRL to select scheduling algorithms online for edge-cloud jobs. The enabling features of EdgeTuner are sophisticated DRL model that captures complex dynamics of Edge-Cloud jobs/tasks and an effective simulator to emulate the response times of short-running jobs in accordance to dynamically changing scheduling algorithms. EdgeTuner trains DRL agents offline by directly interacting with the simulator. We implement EdgeTuner on Kubernetes scheduler and extensively evaluate it on Kubernetes cluster testbed driven by the production traces. Our results show that EdgeTuner outperforms prevailing scheduling algorithms by achieving significant lower job response time while accelerating DRL training speed by more than 180x.

With the massive amount of data generated from mobile devices and the increase of computing power of edge devices, the paradigm of Federated Learning has attracted great momentum. In federated learning, distributed and heterogeneous nodes collaborate to learn model parameters. However, while providing benefits such as privacy by design and reduced latency, the heterogeneous network present challenges to the synchronisation methods, or barrier control methods, used in training, regarding system progress and model convergence etc. The design of these barrier mechanisms is critical for the performance and scalability of federated learning systems. We propose a new barrier control technique called Probabilistic Synchronous Parallel (PSP). In contrast to existing mechanisms, it introduces a sampling primitive that composes with existing barrier control mechanisms to produce a family of mechanisms with improved convergence speed and scalability. Our proposal is supported with a convergence analysis of PSP-based SGD algorithm. In practice, we also propose heuristic techniques that further improve the efficiency of PSP. We evaluate the performance of proposed methods using the federated learning specific FEMNSIT dataset. The evaluation results show that PSP can effectively achieve good balance between system efficiency and model accuracy, mitigating the challenge of heterogeneity in federated learning.

Deep neural networks (DNNs) have been showing significant success in various anomaly detection applications such as smart surveillance and industrial quality control. It is increasingly important to detect anomalies directly on edge devices, because of high responsiveness requirements and tight latency constraints. The accuracy of DNN-based solutions rely on large model capacity and thus long training and inference time, making them inapplicable on resource strenuous edge devices. It is hence imperative to scale DNN model sizes in correspondence to the run-time system requirements, i.e., meeting deadlines with minimal accuracy losses, which are highly dependent on the platforms and real-time system status. Existing scaling techniques either take long training time to pre-generate scaling options or disturb the unsteady training process of anomaly detection DNNs, lacking the adaptability to heterogeneous edge systems and incurring low inference accuracies. In this article, we present LightDNN to scale DNN models for anomaly detection applications at edge, featuring high detection accuracies with lightweight training and inference time. To this end, LightDNN quickly extracts and compresses blocks in a DNN, and provides large scaling space (e.g., 1 million options) by dynamically combining these compressed blocks online. At run-time, LightDNN predicts the DNN's inference latency according to the monitored system status, and optimizes the combination of blocks to maximize its accuracy under deadline constraints. We implement and extensively evaluate LightDNN on both CPU and GPU edge platforms using 8 popular anomaly detection workloads. Comparative experiments with state-of-the-art methods show that our approach provides 145.8 to 0.56 trillion times more scaling options without increasing training and inference overheads, thus achieving as much as 15.05% increase in accuracy under the same deadlines.

Classification algorithms have been widely adopted to detect anomalies for various systems, e.g., IoT, cloud and face recognition, under the common assumption that the data source is clean, i.e., features and labels are correctly set. However, data collected from the wild can be unreliable due to careless annotations or malicious data transformation for incorrect anomaly detection. In this article, we extend a two-layer on-line data selection framework: Robust Anomaly Detector (RAD) with a newly designed ensemble prediction where both layers contribute to the final anomaly detection decision. To adapt to the on-line nature of anomaly detection, we consider additional features of conflicting opinions of classifiers, repetitive cleaning, and oracle knowledge. We on-line learn from incoming data streams and continuously cleanse the data, so as to adapt to the increasing learning capacity from the larger accumulated data set. Moreover, we explore the concept of oracle learning that provides additional information of true labels for difficult data points. We specifically focus on three use cases, (i) detecting 10 classes of IoT attacks, (ii) predicting 4 classes of task failures of big data jobs, and (iii) recognising 100 celebrities faces. Our evaluation results show that RAD can robustly improve the accuracy of anomaly detection, to reach up to 98.95 percent for IoT device attacks (i.e., +7%), up to 85.03 percent for cloud task failures (i.e., +14%) under 40 percent label noise, and for its extension, it can reach up to 77.51 percent for face recognition (i.e., +39%) under 30 percent label noise. The proposed RAD and its extensions are general and can be applied to different anomaly detection algorithms.

Today's available datasets in the wild, e.g., from social media and open platforms, present tremendous opportunities and challenges for deep learning, as there is a significant portion of tagged images, but often with noisy, i.e. erroneous, labels. Recent studies improve the robustness of deep models against noisy labels without the knowledge of true labels. In this paper, we advocate to derive a stronger classifier which proactively makes use of the noisy labels in addition to the original images - turning noisy labels into learning features. To such an end, we propose a novel framework, LABELNET, composed of Amateur and Expert, which iteratively learn from each other. Amateur is a regular image classifier trained by the feedback of Expert, which imitates how human experts would correct the predicted labels from Amateur using the noise pattern learnt from the knowledge of both the noisy and ground truth labels. The trained Amateur and Expert proactively leverage the images and their noisy labels to infer image classes. Our empirical evaluations on noisy versions of MNIST, CIFAR-10, CIFAR-100 and real-world data of Clothing1M show that the proposed model can achieve robust classification against a wide range of noise ratios and with as little as 20-50% training data, compared to state-of-the-art deep models that solely focus on distilling the impact of noisy labels.

Deep learning (DL) models are increasingly built on federated edge participants holding local data. To enable insight extractions without the risk of information leakage, DL training is usually combined with differential privacy (DP). The core theme is to tradeoff learning accuracy by adding statistically calibrated noises, particularly to local gradients of edge learners, during model training. However, this privacy guarantee unfortunately degrades model accuracy due to edge learners' local noises, and the global noise aggregated at the central server. Existing DP frameworks for edge focus on local noise calibration via gradient clipping techniques, overlooking the heterogeneity and dynamic changes of local gradients, and their aggregated impact on accuracy. In this article, we present a systematical analysis that unveils the influential factors capable of mitigating local and aggregated noises, and design PrivateDL to leverage these factors in noise calibration so as to improve model accuracy while fulfilling privacy guarantee. PrivateDL features on: (i) sampling-based sensitivity estimation for local noise calibration and (ii) combining large batch sizes and critical data identification in global training. We implement PrivateDL on the popular Laplace/Gaussian DP mechanisms and demonstrate its effectiveness using Intel BigDL workloads, i.e., considerably improving model accuracy by up to 5X when comparing against existing DP frameworks.

With the exponential growth of data created at the network edge, decentralized and Gossip-based training of deep learning (DL) models on edge computing (EC) gains tremendous research momentum, owing to its capability to learn from resource-strenuous edge nodes with limited network connectivity. Today's edge devices are extremely heterogeneous, e.g., hardware and software stacks, and result in high performance variation of training time and inducing extra delay to synchronize and converge. The large body of prior art accelerates DL, being data or model parallelization, via a centralized server, e.g., parameter server scheme, which may easily turn into the system bottleneck or single point of failure. In this artice, we propose EdgeGossip, a framework specifically designed to accelerate the training process of decentralized and Gossip-based DL training for heterogeneous EC platforms. EdgeGossip features on: (i) low performance variation among multiple EC platforms during iterative training, and (ii) accuracy-aware training to fastly obtain best possible model accuracy. We implement EdgeGossip based on popular Gossip algorithms and demonstrate its effectiveness using real-world DL workloads, i.e., considerably reducing model training time by an average of 2.70 times while only incurring accuracy losses of 0.78 percent.

Cluster schedulers provide flexible resource sharing mechanism for best-effort cloud jobs, which occupy a majority in modern datacenters. Properly tuning a scheduler's configurations is the key to these jobs' performance because it decides how to allocate resources among them. Today's cloud scheduling systems usually rely on cluster operators to set the configuration and thus overlook the potential performance improvement through optimally configuring the scheduler according to the heterogeneous and dynamic cloud workloads. In this paper, we introduce AdaptiveConfig, a run-time configurator for cluster schedulers that automatically adapts to the changing workload and resource status in two steps. First, a comparison approach estimates jobs' performances under different configurations and diverse scheduling scenarios. The key idea here is to transform a scheduler's resource allocation mechanism and their variable influence factors (configurations, scheduling constraints, available resources, and workload status) into business rules and facts in a rule engine, thereby reasoning about these correlated factors in job performance comparison. Second, a workload-adaptive optimizer transforms the cluster-level searching of huge configuration space into an equivalent dynamic programming problem that can be efficiently solved at scale. We implement AdaptiveConfig on the popular YARN Capacity and Fair schedulers and demonstrate its effectiveness using real-world Facebook and Google workloads, i.e., successfully finding best configurations for most of scheduling scenarios and considerably reducing latencies by a factor of two with low optimization time.

The core of many large-scale machine learning (ML) applications, such as neural networks (NN), support vector machine (SVM), and convolutional neural network (CNN), is the training algorithm that iteratively updates model parameters by processing massive datasets. From a plethora of studies aiming at accelerating ML, being data parallelization and parameter server, the prevalent assumption is that all data points are equivalently relevant to model parameter updating. In this article, we challenge this assumption by proposing a criterion to measure a data point's effect on model parameter updating, and experimentally demonstrate that the majority of data points are non-critical in the training process. We develop a slim learning framework, termed SlimML, which trains the ML models only on the critical data and thus significantly improves training performance. To such an end, SlimML efficiently leverages a small number of aggregated data points per iteration to approximate the criticalness of original input data instances. The proposed approach can be used by changing a few lines of code in a standard stochastic gradient descent (SGD) procedure, and we demonstrate experimentally, on NN regression, SVM classification, and CNN training, that for large datasets, it accelerates model training process by an average of 3.61 times while only incurring accuracy losses of 0.37 percent.