Towards Finite-Time Consensus with Graph Convolutional Neural Networks

Iancu, Bianca; Isufi, Elvin

doi:10.23919/eusipco47968.2020.9287610

Cited by 4 publications

(7 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The goal of average consensus is to let each node reach the same average value of all nodes. It can be done in many ways including flooding, distributed linear iterations [ 26 ], deep neural networks [ 27 ], graph filters [ 28 ], and graph convolutional neural networks [ 29 ]. In this paper, we leverage the GNN [ 29 ] to reach average consensus on global model parameters for the following reasons.…”

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

“…It can be done in many ways including flooding, distributed linear iterations [ 26 ], deep neural networks [ 27 ], graph filters [ 28 ], and graph convolutional neural networks [ 29 ]. In this paper, we leverage the GNN [ 29 ] to reach average consensus on global model parameters for the following reasons. First, as for the decentralized learning setting, the network composed of devices can be modeled as a graph with the set of nodes and each link between devices can be seen as an edge.…”

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

“…Last but not least, GNNs do not need high-precision matrix decomposition to get the parameters, as compared to the aforementioned graph filters [ 28 ]. They have shown great scalability to unseen graphs and are robust to the link loss [ 29 ].…”

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

“…The simulation settings are given as follows. We employ the same GNN structure as [ 29 ], which is composed of two graph convolution layers and a fully connected layer. Specifically, we choose the normalized form of the adjacency matrix as the GSO matrix

in Equation ( 7 ) and vary the filter order to see the performance difference.…”

Section: Performance Evaluation On Average Consensusmentioning

confidence: 99%

See 3 more Smart Citations

A Graph Neural Network Based Decentralized Learning Scheme

Gao

Lee

et al. 2022

Sensors

View full text Add to dashboard Cite

As an emerging paradigm considering data privacy and transmission efficiency, decentralized learning aims to acquire a global model using the training data distributed over many user devices. It is a challenging problem since link loss, partial device participation, and non-independent and identically distributed (non-iid) data distribution would all deteriorate the performance of decentralized learning algorithms. Existing work may restrict to linear models or show poor performance over non-iid data. Therefore, in this paper, we propose a decentralized learning scheme based on distributed parallel stochastic gradient descent (DPSGD) and graph neural network (GNN) to deal with the above challenges. Specifically, each user device participating in the learning task utilizes local training data to compute local stochastic gradients and updates its own local model. Then, each device utilizes the GNN model and exchanges the model parameters with its neighbors to reach the average of resultant global models. The iteration repeats until the algorithm converges. Extensive simulation results over both iid and non-iid data validate the algorithm’s convergence to near optimal results and robustness to both link loss and partial device participation.

show abstract

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

Section: Decentralized Learning Tasks Over Networkmentioning

confidence: 99%

in Equation ( 7 ) and vary the filter order to see the performance difference.…”

Section: Performance Evaluation On Average Consensusmentioning

confidence: 99%

See 2 more Smart Citations

A Graph Neural Network Based Decentralized Learning Scheme

Gao

Lee

et al. 2022

Sensors

View full text Add to dashboard Cite

show abstract

“…We consider this scenario in the paper. If one wants to use fewer iterations while sacrificing some MSE, a new finitetime AC method [19] may be considered. The new method extends the idea to nonlinear graph filters and designs filter coefficients using a learning framework of graph convolutional neural networks.…”

Section: Introductionmentioning

confidence: 99%

Distributed Algorithms for Array Signal Processing

Chen

Vaidyanathan

2021

IEEE Trans. Signal Process.

View full text Add to dashboard Cite

Distributed or decentralized estimation of covariance, and distributed principal component analysis have been introduced and studied in the signal processing community in recent years, and applications in array processing have been indicated in some detail. Inspired by these, this paper provides a detailed development of several distributed algorithms for array processing. New distributed algorithms are proposed for DOA estimation methods like root-MUSIC, total least squares-ESPRIT, and FOCUSS. Other contributions include distributed design of the Capon beamformer from data, and distributed implementation of the spatial smoothing method for coherent sources. A distributed implementation of a recently proposed beamspace method called the convolutional beamspace (CBS) is also proposed. The proposed algorithms are fully distributedan average consensus (AC) is used to avoid the need for a fusion center. The algorithms are based on a recently reported finitetime version of AC which converges to the exact solution in a finite number of iterations. Numerical examples are given throughout the paper to show the effectiveness of the proposed algorithms.

show abstract