Recently, the use of connectional brain templates (CBTs) has revolutionized the field of neurological disorder diagnosis through providing integral representation maps of a population-driven brain connectivity and eff...
详细信息
ISBN:
(纸本)9783030603649;9783030603656
Recently, the use of connectional brain templates (CBTs) has revolutionized the field of neurological disorder diagnosis through providing integral representation maps of a population-driven brain connectivity and effective identification of atypical changes in brain connectivity. Ideally, a reliable CBT should satisfy the following criteria: (1) centeredness as it occupies the center of the brain network population, and (2) discriminativeness as it allows to identify differences in brain connectivity between populations with different brain states (e.g., healthy and disordered). Existing state-of-the-art methods for connectional brain template (CBT) estimation from a population of multi-view brain networks (also called brain multigraphs) learn the integration process in a dichotomized manner, where different learning steps are pieced in together independently. Hence, such frameworks are inherently agnostic to the cumulative estimation error from step to step. This is a key limitation that we addressed by capitalizing on the power of deep learning frameworks residing in learning an end-to-end deep mapping using a single objective function to optimize to transform input data into target output data. In this paper, we propose to learn a many-to-one deep learning mapping by designing a clustering-based multi-graph integrator network (MGINet). Our MGINet inputs population of brain multigraphs (many) and outputs a single CBT graph (one). We first propose to tease apart brain multigraph data heterogeneity by first clustering similar samples together using multi-kernel manifold learning. In this way, we are optimally learning to disentangle the heterogeneity of our population and facilitating the integration task for our MGINet. Next, for each cluster, we first integrate the multigraph of each subject into a single graph, then merge the generated graphs into a cluster-specific CBT. Finally, we simply average the cluster-specific CBTs into a final CBT. Our experimental resu
Federated learning (FL) is a rapidly growing privacy preserving collaborative machine learning paradigm. In practical FL applications, local data from each data silo reflect local usage patterns. Therefore, there exis...
详细信息
ISBN:
(数字)9783031001260
ISBN:
(纸本)9783031001260;9783031001253
Federated learning (FL) is a rapidly growing privacy preserving collaborative machine learning paradigm. In practical FL applications, local data from each data silo reflect local usage patterns. Therefore, there exists heterogeneity of data distributions among data owners (a.k.a. FL clients). If not handled properly, this can lead to model performance degradation. This challenge has inspired the research field of heterogeneous federated learning, which currently remains open. In this paper, we propose a data heterogeneity-robust FL approach, FEDGSP, to address this challenge by leveraging on a novel concept of dynamic Sequential-to-Parallel (STP) collaborative training. FEDGSP assigns FL clients to homogeneous groups to minimize the overall distribution divergence among groups, and increases the degree of parallelism by reassigning more groups in each round. It is also incorporated with a novel Inter-Cluster Grouping (ICG) algorithm to assist in group assignment, which uses the centroid equivalence theorem to simplify the NP-hard grouping problem to make it solvable. Extensive experiments have been conducted on the non-i.i.d. FEMNIST dataset. The results show that FEDGSP improves the accuracy by 3.7% on average compared with seven state-of-the-art approaches, and reduces the training time and communication overhead by more than 90%.
Cloud computing has been widely adopted by many companies and government entities. To ensure high quality computing resource provisioning, cloud platforms should offer smart resource management solutions. An important...
详细信息
ISBN:
(纸本)9781538626672
Cloud computing has been widely adopted by many companies and government entities. To ensure high quality computing resource provisioning, cloud platforms should offer smart resource management solutions. An important step toward better resource management is to accurately predict the workloads of the applications running on the cloud. Many existing workload prediction methods are regression based, which require the workloads of the applications show clear seasonality and trend. However, it is difficult to use these methods for tasks which may not have such recurring workload patterns. From careful analysis of the workloads in a real-world cloud, we found that many tasks have busty workloads that are very difficult to predict using regression-based prediction. Instead, we consider a job-pool based approach, where the knowledge about the workloads of a large pool of tasks is used to help predict the workloads of new tasks. In particular, we develop a clustering-based learning approach to realize the job-pool based concept. The pool of jobs are clustered based on their workloads, and a neuralnet is used to learn the characteristics of the workloads in each cluster. When a new job arrives, we use its initial workload pattern and submission parameters to find the cluster it belongs to. Then, the corresponding neuralnet is used to predict the workload of the new job far into the future. based on this predicted long-term workload, smart resource management decisions can be made to reduce the potential overhead in scaling and migration. We also consider a non-clusteringbasedlearning solution and compare it with the clustering-based learning solution. Experimental results show that the clustering-based learning approach can predict the workload more accurately.
暂无评论