Federated learning's poor performance in the presence of heterogeneous data remains one of the most pressing issues in the field. Personalized federated learning departs from the conventional paradigm in which all clients employ the same model, instead striving to discover an individualized model for each client to address the heterogeneity in the data. One of such approach involves personalizing specific layers of neural networks. However, prior endeavors have not provided a dependable rationale, and some have selected personalized layers that are entirely distinct and conflicting. In this work, we take a step further by proposing personalization at the elemental level, rather than the traditional layer-level personalization. To select personalized parameters, we introduce Bayesian neural networks and rely on the uncertainty they offer to guide our selection of personalized parameters. Finally, we validate our algorithm's efficacy on several real-world datasets, demonstrating that our proposed approach outperforms existing baselines.
Performance degradation owing to data heterogeneity and low output interpretability are the most significant challenges faced by federated learning in practical applications. Personalized federated learning diverges from traditional approaches, as it no longer seeks to train a single model, but instead tailors a unique personalized model for each client. However, previous work focused only on personalization from the perspective of neural network parameters and lack of robustness and interpretability. In this work, we establish a novel framework for personalized federated learning, incorporating Bayesian methodology which enhances the algorithm's ability to quantify uncertainty. Furthermore, we introduce normalizing flow to achieve personalization from the parameter posterior perspective and theoretically analyze the impact of normalizing flow on out-of-distribution (OOD) detection for Bayesian neural networks. Finally, we evaluated our approach on heterogeneous datasets, and the experimental results indicate that the new algorithm not only improves accuracy but also outperforms the baseline significantly in OOD detection due to the reliable output of the Bayesian approach.
The online prediction of multivariate signals, existing simultaneously in space and time, from noisy partial observations is a fundamental task in numerous applications. We propose an efficient Neural Network architecture for the online estimation of time-varying graph signals named the Adaptive Least Mean Squares Graph Neural Networks (LMS-GNN). LMS-GNN aims to capture the time variation and bridge the cross-space-time interactions under the condition that signals are corrupted by noise and missing values. The LMS-GNN is a combination of adaptive graph filters and Graph Neural Networks (GNN). At each time step, the forward propagation of LMS-GNN is similar to adaptive graph filters where the output is based on the error between the observation and the prediction similar to GNN. The filter coefficients are updated via backpropagation as in GNN. Experimenting on real-world temperature data reveals that our LMS-GNN achieves more accurate online predictions compared to graph-based methods like adaptive graph filters and graph convolutional neural networks.
Bayesian neural networks use random variables to describe the neural networks rather than deterministic neural networks and are mostly trained by variational inference which updates the mean and variance at the same time. Here, we formulate the Bayesian neural networks as a minimax game problem. We do the experiments on the MNIST data set and the primary result is comparable to the existing closed-loop transcription neural network. Finally, we reveal the connections between Bayesian neural networks and closed-loop transcription neural networks, and show our framework is rather practical, and provide another view of Bayesian neural networks.
We propose the Line Graph Normalized Least Mean Square (LGNLMS) algorithm for online time-varying graph edge signals prediction. LGNLMS utilizes the Line Graph to transform graph edge signals into the node of its edge-to-vertex dual. This enables edge signals to be processed using established GSP concepts without redefining them on graph edges.
An increasingly important brain function analysis modality is functional connectivity analysis which regards connections as statistical codependency between the signals of different brain regions. Graph-based analysis of brain connectivity provides a new way of exploring the association between brain functional deficits and the structural disruption related to brain disorders, but the current implementations have limited capability due to the assumptions of noise-free data and stationary graph topology. We propose a new methodology based on the particle filtering algorithm, with proven success in tracking problems, which estimates the hidden states of a dynamic graph with only partial and noisy observations, without the assumptions of stationarity on connectivity. We enrich the particle filtering state equation with a graph Neural Network called Sequential Monte Carlo Graph Convolutional Network (SMC-GCN), which due to the nonlinear regression capability, can limit spurious connections in the graph. Experiment studies demonstrate that SMC-GCN achieves the superior performance of several methods in brain disorder classification.
This article introduces a novel probability distribution model, namely Complex Isotropic {\alpha}-Stable-Rician (CI{\alpha}SR), for characterizing the data histogram of synthetic aperture radar (SAR) images. Having its foundation situated on the L\'evy {\alpha}-stable distribution suggested by a generalized Central Limit Theorem, the model promises great potential in accurately capturing SAR image features of extreme heterogeneity. A novel parameter estimation method based on the generalization of method of moments to expectations of Bessel functions is devised to resolve the model in a relatively compact and computationally efficient manner. Experimental results based on both synthetic and empirical SAR data exhibit the CI{\alpha}SR model's superior capacity in modelling scenes of a wide range of heterogeneity when compared to other state-of-the-art models as quantified by various performance metrics. Additional experiments are conducted utilizing large-swath SAR images which encompass mixtures of several scenes to help interpret the CI{\alpha}SR model parameters, and to demonstrate the model's potential application in classification and target detection.
SAR technology has been intensively implemented for geo-sensing and mapping purposes due to its advantages of high azimuthal resolution and weather-independent operation compared to other remote sensing technologies. Modelling SAR image data consequently becomes a prominent topic of interest, especially for data populations with impulsive signal features, which are common in SAR images of urban areas. A recently proposed model named Cauchy-Rician has manifested great potential in modelling extremely heterogeneous SAR images, yet the work only provided a MCMC-based parameter estimator that demands considerable computational power. In this work, a novel analytical parameter estimation method based on algebraic moments is proposed to provide stable and accurate estimation of the parameters of the Cauchy-Rician model with significant improvement on computation speed.
In this paper, we introduce an adaptive graph normalized least mean pth power (GNLMP) algorithm for graph signal processing (GSP) that utilizes GSP techniques, including bandlimited filtering and node sampling, to estimate sampled graph signals under impulsive noise. Different from least-squares-based algorithms, such as the adaptive GSP Least Mean Squares (GLMS) algorithm and the normalized GLMS (GNLMS) algorithm, the GNLMP algorithm has the ability to reconstruct a graph signal that is corrupted by non-Gaussian noise with heavy-tailed characteristics. Compared to the recently introduced adaptive GSP least mean pth power (GLMP) algorithm, the GNLMP algorithm reduces the number of iterations to converge to a steady graph signal. The convergence condition of the GNLMP algorithm is derived, and the ability of the GNLMP algorithm to process multidimensional time-varying graph signals with multiple features is demonstrated as well. Simulations show the performance of the GNLMP algorithm in estimating steady-state and time-varying graph signals is faster than GLMP and more robust in comparison to GLMS and GNLMS.