Search

The generative autoencoders, such as the variational autoencoders or the adversarial autoencoders, have achieved great success in lots of real-world applications, including image generation and signal communication. However, little concern has been devoted to their robustness during practical deployment. Due to the probabilistic latent structure, variational autoencoders (VAEs) may confront problems such as a mismatch between the posterior distribution of the latent and real data manifold, or discontinuity in the posterior distribution of the latent. This leaves a back door for malicious attackers to collapse VAEs from the latent space, especially in scenarios where the encoder and decoder are used separately, such as communication and compressed sensing. In this work, we provide the first study on the adversarial robustness of generative autoencoders in the latent space. Specifically, we empirically demonstrate the latent vulnerability of popular generative autoencoders through attacks in the latent space. We also evaluate the difference between variational autoencoders and their deterministic variants and observe that the latter performs better in latent robustness. Meanwhile, we identify a potential trade-off between the adversarial robustness and the degree of the disentanglement of the latent codes. Additionally, we also verify the feasibility of improvement for the latent robustness of generative autoencoders through adversarial training. In summary, we suggest concerning the adversarial latent robustness of the generative autoencoders, analyze several robustness-relative issues, and give some insights into a series of key challenges. [ABSTRACT FROM AUTHOR]

There is a recent trend to leverage the power of graph neural networks (GNNs) for brain-network based psychiatric diagnosis, which, in turn, also motivates an urgent need for psychiatrists to fully understand the decision behavior of the used GNNs. However, most of the existing GNN explainers are either post-hoc in which another interpretive model needs to be created to explain a well-trained GNN, or do not consider the causal relationship between the extracted explanation and the decision, such that the explanation itself contains spurious correlations and suffers from weak faithfulness. In this work, we propose a granger causality-inspired graph neural network (CI-GNN), a built-in interpretable model that is able to identify the most influential subgraph (i.e., functional connectivity within brain regions) that is causally related to the decision (e.g., major depressive disorder patients or healthy controls), without the training of an auxillary interpretive network. CI-GNN learns disentangled subgraph-level representations α and β that encode, respectively, the causal and non-causal aspects of original graph under a graph variational autoencoder framework, regularized by a conditional mutual information (CMI) constraint. We theoretically justify the validity of the CMI regulation in capturing the causal relationship. We also empirically evaluate the performance of CI-GNN against three baseline GNNs and four state-of-the-art GNN explainers on synthetic data and three large-scale brain disease datasets. We observe that CI-GNN achieves the best performance in a wide range of metrics and provides more reliable and concise explanations which have clinical evidence. The source code and implementation details of CI-GNN are freely available at GitHub repository (https://github.com/ZKZ-Brain/CI-GNN/). [ABSTRACT FROM AUTHOR]

The normalized subband adaptive filtering algorithm provides fast convergence rate for colored input signals as compared to the normalized least mean square algorithm, but it suffers from a poor convergence issue in the α -stable noise. In light of this, the normalized subband p-norm (NSPN) algorithm, which is based on the least mean p-power error (MPE) criterion, is proposed in this study. This technique is not only robust against impulsive noise samples, but it also maintains a fast convergence rate when colored input signals are used. In addition to this, we develop both the steady-state and the transient models of the NSPN algorithm and provide some insights. Then, in order to solve the problem of making a choice regarding the order p in the NSPN algorithm, we design an autonomous system and come up with the NSPN algorithm with a variable p-norm (VP-NSPN). In addition, we offer the TFC-based VP-NSPN algorithm with a fast convergence rate and low steady-state misadjustment simultaneously by making use of the tap-weights feedback-based convex combination (TFC) scheme. In conclusion, simulation results on system identification and acoustic echo cancellation are undertaken in order to validate the superiority of the proposed algorithms and validate the usefulness of the theoretical analysis. [ABSTRACT FROM AUTHOR]

Multiview clustering (MVC) has attracted considerable attention owing to its remarkable capacity to reconcile diverse information from multiple perspectives. However, traditional MVC generally has a narrow scope of application owing to its limited efficiency. Consequently, various efficient MVC (EMVC) methods have emerged recently. Despite their promising performance, these EMVC methods still have several unresolved issues: (1) They suffer from reduced effectiveness caused by representation non-alignment across views and information mismatch between stages, and (2) they fail to efficiently resist complex noises and outliers. To address these issues, we propose an efficient correntropy-based multiview clustering method with alignment discretization (ECMCAD). Specifically, a correntropy-based multipartition learning model was developed to efficiently learn view-specific robust partition-level representations. Additionally, a novel alignment discretization strategy was designed to align the learned cross-view representations into a consensus discrete indicator to integrate representation learning, representation alignment, and discrete label acquisition into a unified framework. Furthermore, an efficient alternating optimization method was developed to solve the model. Numerous experiments illustrated the superiority of the proposed method over state-of-the-art baselines. [ABSTRACT FROM AUTHOR]

Concept factorization (CF), unlike nonnegative matrix factorization (NMF), can handle data with negative values by approximating the original data with two low-dimensional nonnegative matrices and itself. Nevertheless, existing CF-based methods continue to suffer from the two issues specified as follows: (1) Their effectiveness is reduced by the high degree of factorization freedom and the two-stage mismatch between factorization and category acquisition, and (2) their robustness drops significantly when dealing with complex noise. In response to the aforementioned issues, we propose a robust spectral-embedded bilateral orthogonal concept factorization (RSOCF) model for clustering. It constrains the factor matrices as orthogonal matrices to decrease the freedom and obtain samples' categories directly after factorization, which can significantly improve clustering effectiveness. Moreover, correntropy is introduced into RSOCF to improve its robustness to complex noise. To optimize the non-convex RSOCF model, a half-quadratic-based algorithm is devised. Numerous experiments demonstrate that RSOCF surpasses other state-of-the-art methods in terms of clustering effectiveness and robustness. • A novel robust concept factorization-based clustering model is proposed. • A fast algorithm is proposed to optimize the non-convex model. • Numerous experiments validate its clustering effectiveness and robustness. [ABSTRACT FROM AUTHOR]

Converging evidence increasingly suggests that psychiatric disorders, such as major depressive disorder (MDD) and autism spectrum disorder (ASD), are not unitary diseases, but rather heterogeneous syndromes that involve diverse, co-occurring symptoms and divergent responses to treatment. This clinical heterogeneity has hindered the progress of precision diagnosis and treatment effectiveness in psychiatric disorders. In this study, we propose BPI-GNN, a new interpretable graph neural network (GNN) framework for analyzing functional magnetic resonance images (fMRI), by leveraging the famed prototype learning. In addition, we introduce a novel generation process of prototype subgraph to discover essential edges of distinct prototypes and employ total correlation (TC) to ensure the independence of distinct prototype subgraph patterns. BPI-GNN can effectively discriminate psychiatric patients and healthy controls (HC), and identify biological meaningful subtypes of psychiatric disorders. We evaluate the performance of BPI-GNN against 11 popular brain network classification methods on three psychiatric datasets and observe that our BPI-GNN always achieves the highest diagnosis accuracy. More importantly, we examine differences in clinical symptom profiles and gene expression profiles among identified subtypes and observe that our identified brain-based subtypes have the clinical relevance. It also discovers the subtype biomarkers that align with current neuro-scientific knowledge. • We propose a new GNN architecture with prototype learning for psychiatric diagnosis. • We successfully identify biological meaningful subtypes of psychiatric disorders. • We observe subtype biomarkers that closely correspond to neuroimaging findings. • Our framework achieves the best performance on three psychiatric datasets. [ABSTRACT FROM AUTHOR]

Owing to its ability to handle negative data and promising clustering performance, concept factorization (CF), an improved version of non-negative matrix factorization, has been incorporated into multi-view clustering recently. Nevertheless, existing CF-based multi-view clustering methods still have the following issues: (1) they directly conduct factorization in the original data space, which means its efficiency is sensitive to the feature dimension; (2) they ignore the high degree of factorization freedom of standard CF, which may lead to non-uniqueness factorization thereby causing reduced effectiveness; (3) traditional robust norms they used are unable to handle complex noises, significantly challenging their robustness. To address these issues, we establish a fast multi-view clustering via correntropy-based orthogonal concept factorization (FMVCCF). Specifically, FMVCCF executes factorization on a learned consensus anchor graph rather than directly decomposing the original data, lessening the dimensionality sensitivity. Then, a lightweight graph regularization term is incorporated to refine the factorization process with a low computational burden. Moreover, an improved multi-view correntropy-based orthogonal CF model is developed, which can enhance the effectiveness and robustness under the orthogonal constraint and correntropy criterion, respectively. Extensive experiments demonstrate that FMVCCF can achieve promising effectiveness and robustness on various real-world datasets with high efficiency. [ABSTRACT FROM AUTHOR]

Accurate cluster photovoltaic power prediction (CPPP) is crucial for the operation and control of renewable energy grid-connected power systems. The traditional modeling strategies for CPPP such as direct aggregation (DA) and statistical upscaling (SU) have limitations such as error accumulation and upscaling factor uncertainty. To address these issues, this paper proposed a novel hybrid approach for CPPP by combining machine learning models with an improved SU technique. Firstly, a robust broad learning system (BLS) model, in which the Generalized Maximum Correntropy Criterion (GMCC) is used to replace the original mean square error (MSE) loss in BLS, is proposed to solve the problem of multiple outliers affecting the prediction accuracy of regional cluster stations, and it is called GBLS. Then, the Relevance Vector Machine (RVM) as an effective nonlinear regression model is further utilized to compensate for the prediction errors obtained by the GBLS to form the hybrid prediction model, namely GBLS-RVM. Moreover, to mitigate the uncertainty associated with scaling factors in traditional SU strategy, a new SU strategy is developed to refine the relationship between the reference station and the cluster sub-region, enabling direct modeling for regional power prediction. Finally, data from two PV clusters in different regions of China are used to validate the effectiveness of the proposed model, and the results show that under the improved SU strategy, the GBLS-RVM model, reduced RMSE by approximately 33.7% compared to the traditional BLS model, and the RMSE decreased by 12.89% and 30.2% when compared to traditional DA and traditional SU strategies. • A novel hybrid method for CPPP is developed by combining ML and improved SU technique. • A BLS with GMCC (GMCC-BLS) as a prediction model is developed to handle the outliers in measurement data. • An innovative error correction strategy based on RVM is proposed to improve prediction reliability. • A new SU technique is used to establish direct relationships between reference stations and subregions. [ABSTRACT FROM AUTHOR]

Concept factorization (CF) has attracted widespread attention for its promising performance in document clustering. Among various CF variants, graph-regularized CF is the most impressive type, which can improve clustering effectiveness by exploring structural information. Nevertheless, their clustering efficiency is restricted by the following considerations: (1) the introduction of the full-sample graph is accompanied by an increase in computational complexity; (2) most of them require intensive multiplications in optimization, which impair the optimization efficiency. To address these issues, in this work, we propose an anchor-graph regularized orthogonal concept factorization (AROCF) method to enhance the clustering efficiency and effectiveness in document clustering tasks. Firstly, AROCF approximates the full-sample graph with a small-scale anchor graph to reduce the complexity of graph construction from quadratic to linear. Then, one of the factor matrices is constrained as the cluster indicator matrix in our method, which can avoid extra efficiency loss in K-means after optimization. Finally, an orthogonal constraint is employed to restrict the freedom of factorization to increase the clustering effectiveness. To optimize the AROCF model, we develop a fast optimization strategy by combining the trace and orthogonality of matrices. Extensive experiments on various document datasets demonstrate the effectiveness and efficiency of AROCF. [ABSTRACT FROM AUTHOR]

Graph-based clustering commonly provides promising clustering effectiveness as it can preserve samples' local geometric information. Inspired by it, multi-view graph clustering was developed to integrate complementary information among graphs of diverse views and it has received intensive attention recently. Nevertheless, on the one hand, most existing methods require extra k-means after obtaining embedding representations to generate a discrete cluster indicator, which reduces effectiveness due to the two-stage mismatch. On the other hand, numerous complex noises in real-world multi-view data challenge the robustness of existing clustering methods. In this paper, we established a discrete correntropy-based multi-view anchor-graph clustering (DCMAC) model that not only emphasizes the aforementioned issues but also makes use of anchor graphs to improve the efficiency of the graph construction stage. To optimize this non-convex model, we propose a fast half-quadratic-based coordinate descent strategy to acquire the discrete cluster indicator directly without extra k-means. Furthermore, we extend the DCMAC model to a single-view form and provide optimization strategies for it. Extensive experiments illustrate that the proposed method is effective and robust compared to those advanced baselines. • A new discrete correntropy-based multi-view anchor-graph clustering (DCMAC) model is proposed. • A novel half-quadratic-based coordinate descent algorithm is proposed to solve the DCMAC model. • To handle single-view clustering instances, DCMAC is extended to a single-view form named DCAC. • Numerous experiments demonstrate the outstanding effectiveness and robustness of DCMAC. [ABSTRACT FROM AUTHOR]

Recently, the kernel adaptive filtering algorithm with the generalized maximum correntropy criterion (GMCC) has gained attention due to its excellent nonlinear system modeling ability and robustness to non-Gaussian noise. However, since the default center of the GMCC is located at zero, it hampers its performance in the non-zero mean noise environment. To overcome the problem, this paper proposes a kernel adaptive filtering algorithm based on the GMCC with variable center (GMCC-VC) called generalized kernel maximum correntropy criterion with variable center (GKMCC-VC). In addition, the online vector quantization (VQ) method is introduced to the GKMCC-VC algorithm for suppressing the increasing network size. The probability of divergence, the mean-square steady-state behavior, and the convergence condition of the GKMCC-VC algorithm are also obtained under some assumptions. The potential relationship between the GMCC-VC and the generalized minimum error entropy (GMEE) criteria is discussed in the paper. Finally, the simulation results verify the superiority of the proposed algorithm in modeling nonlinear systems and the correctness of the theoretical model. • This paper proposes the GKMCC-VC algorithm for non-zero mean and non-Gaussian noise. • The steady-state performance of the proposed algorithm is analyzed. • Simulation results show the effectiveness and feasibility of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Randomized autoencoder (RAE) has attracted much attention due to its strong capability of representation with fast learning speed. However, the mainstream RAEs are still designed for scalar/vector data, which inevitably destroys the structure information of tensor data. To alleviate this deficiency, a novel convolutions based matrix randomized autoencoder (MRAE) is developed for two-dimensional (2D) data in this paper, including a one-side MRAE (OMRAE) exploiting the row or column information and a double-side MRAE (DMRAE) that simultaneously extracts the row and column information by 2 parallel OMRAEs. To reduce meaningless encoded features, the within-class scatter matrix (WSI) and within-class interaction distance (WID) constraints are added into OMRAE resulting WSI-OMRAE and WID-OMRAE, respectively. To demonstrate the superiority, stacked MRAEs are embedded into hierarchical regularized least squares for one-class classification and comparisons with several state-of-the-art methods are provided. The source code would be available at https://github.com/ML-HDU/MRAE. • MRAE can learn from 2D data directly without breaking structure information. • WSI-OMRAE minimizes both reconstruction error and within-class scatter. • WID-OMRAE can learn intrinsic manifold structure with compact representation. • Adopting two-sides OMRAEs can perform row/column feature learning in parallel. [ABSTRACT FROM AUTHOR]