Shared Feature Representation Research Articles

Consistent and specific multi-view multi-label learning with correlation information

In multi-view multi-label (MVML) learning, each sample is represented by several heterogeneous distinct feature representations while associated with a set of class labels simultaneously. To achieve MVML learning, most of the existing methods contribute to the recovery of a consistent subspace, i.e., a shared feature representation, among multiple views. Nevertheless, each view has its inherent specific properties used in the discrimination process of labels. These methods lose sight in the specific information exploitation, and therefore are easily trapped in a sub-optimal result. In this study, we present an optimization framework CSVL to solve the learning problem. The main technical contribution in CSVL is a formulation for MVML learning while consistent subspace across views, specific subspace for each view, and the correlations among labels are taken into account. Specifically, consistent subspace is recovered by imposing a low-rank constraint among multiple views, and specific subspace of each view is extra generated with Frobenius norm. To further improve model generalization capability, we preserve both feature manifolds from multiple views and label correlations from multiple labels. Extensive experiments on 7 benchmark datasets show that our proposal CSVL has the advantages in MVML learning.

LJCD-Net: Cross-Domain Jamming Generalization Diagnostic Network Based on Deep Adversarial Transfer.

Global Navigation Satellite Systems (GNSS) offer comprehensive position, navigation, and timing (PNT) estimates worldwide. Given the growing demand for reliable location awareness in both indoor and outdoor contexts, the advent of fifth-generation mobile communication technology (5G) has enabled expansive coverage and precise positioning services. However, the power received by the signal of interest (SOI) at terminals is notably low. This can lead to significant jamming, whether intentional or unintentional, which can adversely affect positioning receivers. The diagnosis of jamming types, such as classification, assists receivers in spectrum sensing and choosing effective mitigation strategies. Traditional jamming diagnosis methodologies predominantly depend on the expertise of classification experts, often demonstrating a lack of adaptability for diverse tasks. Recently, researchers have begun utilizing convolutional neural networks to re-conceptualize a jamming diagnosis as an image classification issue, thereby augmenting recognition performance. However, in real-world scenarios, the assumptions of independent and homogeneous distributions are frequently violated. This discrepancy between the source and target distributions frequently leads to subpar model performance on the test set or an inability to procure usable evaluation samples during training. In this paper, we introduce LJCD-Net, a deep adversarial migration-based cross-domain jamming generalization diagnostic network. LJCD-Net capitalizes on a fully labeled source domain and multiple unlabeled auxiliary domains to generate shared feature representations with generalization capabilities. Initially, our paper proposes an uncertainty-guided auxiliary domain labeling weighting strategy, which estimates the multi-domain sample uncertainty to re-weight the classification loss and specify the gradient optimization direction. Subsequently, from a probabilistic distribution standpoint, the spatial constraint imposed on the cross-domain global jamming time-frequency feature distribution facilitates the optimization of collaborative objectives. These objectives include minimizing both the source domain classification loss and auxiliary domain classification loss, as well as optimizing the inter-domain marginal probability and conditional probability distribution. Experimental results demonstrate that LJCD-Net enhances the recognition accuracy and confidence compared to five other diagnostic methods.

Source-free domain adaptation for transferable remaining useful life prediction of machine considering source data absence

Data-driven method developed based on deep learning theory has satisfactorily solved the problems of fault classification and health prognosis for industrial equipment. Meanwhile, domain adaptation (DA) further endows the models with the ability to operate effectively across operating scenarios. Unfortunately, current DA methods require the overall participation of source data, which in real-world industrial scenarios is unavailable due to its privacy. In response to this, this paper proposes source-free domain adaptation (SFDA) to realize the transferable remaining useful life (RUL) prognosis of rotating machinery considering source data absence. Specifically, SFDA transforms measurement of inter-domain feature discrepancy into measurement of discrepancy between estimated parameters of corresponding statistical models through implicit statistical distribution generalization. In addition, a design criterion for improving the reliability of pseudo labels is proposed, which generates more robust pseudo labels for source-free domain self-training by minimizing angle offset and distance offset. The proposed framework has good generalization ability for target data and shared feature representations between different multiple domains. Run-to-failure degradation experiments were conducted on the core components of rotating equipment, and the experimental results verified the effectiveness and superiority of the proposed prediction framework.

Multi-modal brain tumor segmentation via disentangled representation learning and region-aware contrastive learning

Brain tumors are threatening the life and health of people in the world. Automatic brain tumor segmentation using multiple MR images is challenging in medical image analysis. It is known that accurate segmentation relies on effective feature learning. Existing methods address the multi-modal MR brain tumor segmentation by explicitly learning a shared feature representation. However, these methods fail to capture the relationship between MR modalities and the feature correlation between different target tumor regions. In this paper, I propose a multi-modal brain tumor segmentation network via disentangled representation learning and region-aware contrastive learning. Specifically, a feature fusion module is first designed to learn the valuable multi-modal feature representation. Subsequently, a novel disentangled representation learning is proposed to decouple the fused feature representation into multiple factors corresponding to the target tumor regions. Furthermore, contrastive learning is presented to help the network extract tumor region-related feature representations. Finally, the segmentation results are obtained using the segmentation decoders. Quantitative and qualitative experiments conducted on the public datasets, BraTS 2018 and BraTS 2019, justify the importance of the proposed strategies, and the proposed approach can achieve better performance than other state-of-the-art approaches. In addition, the proposed strategies can be extended to other deep neural networks.

Transformer-based multimodal feature enhancement networks for multimodal depression detection integrating video, audio and remote photoplethysmograph signals

Depression stands as one of the most widespread psychological disorders and has garnered increasing attention. Currently, how to effectively achieve automatic multimodal depression detection for assisting doctors in early diagnosis of depression, has become an important and challenging issue. To address this issue, this work proposes Transformer-based feature enhancement networks for multimodal depression detection. The proposed method effectively integrates three modalities including video, audio and remote photoplethysmographic (rPPG) signals for multimodal depression detection, in which the rPPG modality is introduced as an additional modality for enhancing the effectiveness of multimodal depression detection. The proposed method consists of three key steps: multimodal feature extraction for video, audio and rPPG modalities, Transformer-based multimodal feature enhancement (TMFE), and graph fusion networks (GFN) based multimodal fusion and depression prediction. More specially, in the stage of multimodal feature extraction, for video and audio modalities we employ deep convolutional neural networks (CNN) to extract the corresponding high-level video and audio features, respectively. For rPPG modality, we adopt a short-time end-to-end rPPG estimation framework to extract the rPPG signal values. The TMFE module stacks multiple Transformers such as the inter-modal, intra-modal, and tri-modal Transformers to jointly capture the dynamics and relationships within and between modalities for each time-step of input sequences. The GFN module is designed to effectively fuse the obtained feature representations from different modalities while maintaining the interactions between them simultaneously. Finally, the obtained shared feature representations of all modalities are fed into a multilayer perceptrons (MLP) network to implement final depression detection tasks. Extensive experiments are conducted on two public datasets such as AVEC2013 and AVEC2014, and experimental results demonstrate the validity of the proposed method on depression detection tasks.

Towards Data-Independent Knowledge Transfer in Model-Heterogeneous Federated Learning

Federated Distillation (FD) extends classic Federated Learning (FL) to a more general training framework that enables model-heterogeneous collaborative learning by Knowledge Distillation (KD) across multiple clients and the server. However, existing KD-based algorithms usually require a set of shared input samples for each client to produce soft-prediction for distillation. Worse still, such a manual selection is accompanied by careful deliberations or prior information on clients' private data distribution, which is not in line with the privacy-preserving characteristic of classic FL. In this paper, we propose a novel training framework to achieve data-independent knowledge transfer by properly designing a distributed generative adversarial network (GAN) between the server and clients that can synthesize shared feature representations to facilitate the FD training. Specifically, we deploy a generator on the server and reuse each local model as a federated discriminator to form a lightweight efficient distributed GAN that can automatically synthesize simulated global feature representations for distillation. Moreover, since the synthesized feature representations are usually more faithful and homologous with global data distribution, faster and better training convergence can be obtained. Extensive experiments on different tasks and heterogeneous models demonstrate the effectiveness of the proposed framework on model accuracy and communication overhead.

A Comparative Study of Clinical Risk Prediction Using Limited Patient Electronic Health Records

Predictive modeling of clinical risk using patient electronic health records (EHRs) has the potential to enhance healthcare outcomes by enabling early detection and intervention for high-risk patients. However, dealing with sparse, irregular, and temporal EHR data presents significant challenges. This paper presents a comparative study of clinical risk prediction with limited patient electronic medical records. The related literature is categorized and compared based on research objectives, methods and experimental analysis. Additionally, potential research opportunities for future work in this area are discussed. Meta-learning-based algorithms have the ability to overcome data scarcity challenge by learning shared feature representations. Nevertheless, further research is necessary to address limitations such as the interpretability and generalizability of the model across different patient populations.

A meta-learning network with anti-interference for few-shot fault diagnosis

Considering the changing working conditions of rotating machinery in operation, it is often difficult to collect data accurately in some severe fault states, and the lack of data can lead to poor performance of deep learning-based fault diagnosis models. In the few-shot scenario, traditional meta-learning methods yield superior predictions by identifying ideal initialization parameters. When the update gradients of tasks are in different directions, meta-learning modifies the weights of samples in the current tasks to average them out, which may lead to negative transfer between tasks and poor generalization of biases. Based on the above problems, we propose a meta-learning network with anti-interference (AIML) for few-shot fault diagnosis, which is obtained by combining a dynamic fine-tuning technique to increase the gradient agreement of the tasks. AIML consists of two functions: the feature encoding network (FEN) and the base network. AIML uses the property of shared parameters about transfer learning to learn shared feature representations of different tasks, while FEN dynamically adjusts the loss weights of conflicting tasks due to meta-learning with two-level optimization techniques. First, in the internal loop, AIML updates the gradient of the base network while the network parameters of FEN are fixed. Instead of immediately computing the gradient of the optimal parameters for the new task, the FEN is then integrated in the outer loop to learn the meta-representation, and the weights obtained for the various tasks relate to the gradient for meta-optimization. Only during the outer loop stage of the meta-training is the FEN updated. Three publicly available datasets are used to assess the performance of AIML, and the results show that it is more effective at resolving problems involving few-shot fault diagnosis.

Joint learning-based feature reconstruction and enhanced network for incomplete multi-modal brain tumor segmentation

Multimodal Magnetic Resonance Imaging (MRI) can provide valuable complementary information and substantially enhance the performance of brain tumor segmentation. However, it is common for certain modalities to be absent or missing during clinical diagnosis, which can significantly impair segmentation techniques that rely on complete modalities. Current advanced methods attempt to address this challenge by developing shared feature representations via modal fusion to handle different missing modality situations. Considering the importance of missing modality information in multimodal segmentation, this paper utilize a feature reconstruction method to recover the missing information, and proposes a joint learning-based feature reconstruction and enhancement method for incomplete modality brain tumor segmentation. The method leverages an information learning mechanism to transfer information from the complete modality to a single modality, enabling it to obtain complete brain tumor information, even without the support of other modalities. Additionally, the method incorporates a module for reconstructing missing modality features, which recovers fused features of the absent modality through utilizing the abundant potential information obtained from the available modalities. Furthermore, the feature enhancement mechanism improves shared feature representation by utilizing the information obtained from the missing modalities that have been reconstructed. These processes enable the method to obtain more comprehensive information regarding brain tumors in various missing modality circumstances, thereby enhancing the model’s robustness. The performance of the proposed model was evaluated on BraTS datasets and compared with other deep learning algorithms using Dice similarity scores. On the BraTS2018 dataset, the proposed algorithm achieved a Dice similarity score of 86.28%, 77.02%, and 59.64% for whole tumors, tumor cores, and enhanced tumors, respectively. These results demonstrate the superiority of our framework over state-of-the-art methods in missing modalities situations.

Flexible Fusion Network for Multi-modal Brain Tumor Segmentation.

Automated brain tumor segmentation is crucial for aiding brain disease diagnosis and evaluating disease progress. Currently, magnetic resonance imaging (MRI) is a routinely adopted approach in the field of brain tumor segmentation that can provide different modality images. It is critical to leverage multi-modal images to boost brain tumor segmentation performance. Existing works commonly concentrate on generating a shared representation by fusing multi-modal data, while few methods take into account modality-specific characteristics. Besides, how to efficiently fuse arbitrary numbers of modalities is still a difficult task. In this study, we present a flexible fusion network (termed F 2 Net) for multi-modal brain tumor segmentation, which can flexibly fuse arbitrary numbers of multi-modal information to explore complementary information while maintaining the specific characteristics of each modality. Our F 2 Net is based on the encoder-decoder structure, which utilizes two Transformer-based feature learning streams and a cross-modal shared learning network to extract individual and shared feature representations. To effectively integrate the knowledge from the multi-modality data, we propose a cross-modal feature-enhanced module (CFM) and a multi-modal collaboration module (MCM), which aims at fusing the multi-modal features into the shared learning network and incorporating the features from encoders into the shared decoder, respectively. Extensive experimental results on multiple benchmark datasets demonstrate the effectiveness of our F 2 Net over other state-of-the-art segmentation methods. The implementation code and segmentation maps will be released at https://github.com/taozh2017/F2Net.

CsAGP: Detecting Alzheimer's disease from multimodal images via dual-transformer with cross-attention and graph pooling

Alzheimer's disease (AD) is a terrible and degenerative disease commonly occurring in the elderly. Early detection can prevent patients from further damage, which is crucial in treating AD. Over the past few decades, it has been demonstrated that neuroimaging can be a critical diagnostic tool for AD, and the feature fusion of different neuroimaging modalities can enhance diagnostic performance. Most previous studies in multimodal feature fusion have only concatenated the high-level features extracted by neural networks from various neuroimaging images simply. However, a major problem of these studies is overlooking the low-level feature interactions between modalities in the feature extraction stage, resulting in suboptimal performance in AD diagnosis. In this paper, we develop a dual-branch vision transformer with cross-attention and graph pooling, namely CsAGP, which enables multi-level feature interactions between the inputs to learn a shared feature representation. Specifically, we first construct a brand-new cross-attention fusion module (CAFM), which processes MRI and PET images by two independent branches of differing computational complexity. These features are fused merely by the cross-attention mechanism to enhance each other. After that, a concise graph pooling algorithm-based Reshape-Pooling-Reshape (RPR) framework is developed for token selection to reduce token redundancy in the proposed model. Extensive experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) database demonstrated that the suggested method obtains 99.04%, 97.43%, 98.57%, and 98.72% accuracy for the classification of AD vs. CN, AD vs. MCI, CN vs. MCI, and AD vs. CN vs. MCI, respectively.

An effective multimodal representation and fusion method for multimodal intent recognition

Intent recognition is a crucial task in natural language understanding. Current research mainly focuses on task-specific unimodal intent recognition. However, in real-world scenes, human intentions are complex and need to be judged by integrating information such as speech, tone, expression, and action. Therefore, this paper proposes an effective multimodal representation and fusion method (EMRFM) for intent recognition in real-world multimodal scenes. First, text, audio, and vision features are extracted based on pre-trained BERT, Wav2vec 2.0, and Faster R-CNN. Then, considering the complementarity and consistency among the modalities, the modality-shared and modality-specific encoders are constructed to learn shared and specific feature representations of the modalities. Finally, an adaptive multimodal fusion method based on an attention-based gated neural network is designed to eliminate noise features. Comprehensive experiments are conducted on the multimodal intent recognition MIntRec benchmark dataset. Our proposed model achieves higher accuracy, precision, recall, and F1-score than state-of-the-art multimodal learning methods. We also conduct multimodal sentiment recognition experiments on the CMU-MOSI dataset, and our model still outperforms state-of-the-art methods. In addition, the experiment demonstrates that the model’s multimodal representation well learned the modality’s shared and specific features. The multimodal fusion of the model achieves adaptive fusion and effectively reduces possible noise interference.

Multimodal emotion recognition based on audio and text by using hybrid attention networks

Multimodal Emotion Recognition (MER) has recently become a popular and challenging topic. The most key challenge in MER is how to effectively fuse multimodal information. Most of prior works may not fully consider the inter-modal and intra-modal attention mechanism to jointly learn intra-modal and inter-modal emotional salient information for further improving the performance of MER. To address this problem, this paper proposes a new MER framework based on audio and text by using Hybrid Attention Networks (MER-HAN). The proposed MER-HAN combines three different attention mechanisms such as the local intra-modal attention, the cross-modal attention, and the global inter-modal attention to effectively learn intra-modal and inter-modal emotional salient features for MER. Specifically, an Audio and Text Encoder (ATE) block equipped with deep learning techniques with the local intra-modal attention mechanism is initially designed to learn high-level audio and text feature representations from the corresponding audio and text sequences, respectively. Then, a Cross-Modal Attention (CMA) block is presented to jointly capture high-level shared feature representations across audio and text modalities. Finally, a Multimodal Emotion Classification (MEC) block with the global inter-modal attention mechanism is provided to obtain final MER results. Extensive experiments conducted on two public multimodal emotional datasets, i.e., IEMOCAP and MELD datasets, show the advantage of the proposed MER-HAN on MER tasks.

BoAu: Malicious traffic detection with noise labels based on boundary augmentation

The effectiveness of deep-learning-based malicious traffic detection systems relies on high-quality labeled traffic datasets. However, malicious traffic labeling approaches can easily lead to incorrect labeling, which can have a harmful impact on models. To this end, various methods for learning with noise labels have been proposed. They exclude suspected wrong samples from model updates to ensure accuracy. However, this also removes hard samples, resulting in poor model decision boundaries and the loss of ability to classify hard samples. In this paper, we propose a boundary-augmentation-based approach for malicious traffic identification named BoAu. Unlike other approaches, BoAu treats all samples, including hard samples, equally during training to construct more accurate decision boundaries and thus improve accuracy. Meanwhile, a decision boundary augmentation module is designed to mitigate the impact of mislabeled hard samples on decision boundary generation. The decision boundary augmentation module adaptively adjusts the losses of hard samples based on their distance from the cluster to which their labels belong and other clusters, thus driving the shared feature representation network to fit the true label distribution. We validated BoAu in identifying malicious traffic with noise labels on a dataset covering 22 classes of realistic encrypted malicious traffic. Experimental results showed that even under scenarios with up to 90% noise labels, the classification accuracy was still over 80%, which was better than the state-of-the-art approaches. In addition, we validated the applicability of BoAu on several public datasets, including CIC-IDS-2017 and IoT-23.

Tube-Embedded Transformer for Pixel Prediction

Multi-task pixel-level learning, which aims to exploit the inter-task interactions to improve the learning of each task, is an important but challenging issue in visual reception and multimedia applications. Measuring the inter-task correlation and intra-task specificity, we propose a tube-embedded transformer (TET) framework for robust multi-task pixel prediction. To facilitate inter-task interactions, we aggregate and project all tasks into a shared tube pool to generate the latent multi-task representation during the coarse-to-fine decoding stages. The resulting task-tube interactions replace the two-by-two task-task interactions to reduce the model complexity significantly. In addition, we introduce the transformer mechanism to adaptively transfer tube features to the target task. Concretely, on the one hand, multi-task features aggregate in the tube to generate the shared feature representation bases; on the other hand, based on the task-tube association and complementarity, the tube outputs the query entry and the weighting coefficients of the target task. Experimentally, on the joint learning of semantic segmentation, depth estimation, and surface normal estimation, the comparison experiments show the superiority of the TET multi-task learning method over other state-of-the-art approaches, and the ablation experiments verify the effectiveness of the TET mechanism.

Learnable Multi-View Matrix Factorization With Graph Embedding and Flexible Loss

The goal of multi-view learning is to learn latent patterns from various data sources. Most of previous research focused on fitting feature embedding in target tasks.There is very limited research on the connection between feature representations with hidden layers of neural networks. In this paper, a multi-view deep matrix factorization model is proposed to learn a shared feature representation. The proposed model automatically explores the most discriminative features of multi-view data and makes these features meet the requirements of specific applications.Here we explore the connection between deep learning and feature representations. First, the model constructs a scalable neural network with shared hidden layers for exploring a low-dimensional representations of all views. Second, the quality of representation matrix is evaluated via relaxed graph regularization and evaluators to improve the feature representation capability of matrix factorization. Finally, the effectiveness of the proposed method is verified through comparative experiments with eight state-of-the-art multi-view clustering algorithms on eight real-world datasets.

Context-dependent emotion recognition

Most previous methods for emotion recognition focus on facial emotion and ignore the rich context information that implies important emotion states. To make full use of the contextual information to make up for the facial information, we propose the Context-Dependent Net (CD-Net) for robust context-aware human emotion recognition. Inspired by the long-range dependency of the transformer, we introduce the tubal transformer which forms the shared feature representation space to facilitate the interactions among the face, body, and context features. Besides, we introduce the hierarchical feature fusion to recombine the enhanced multi-scale face, body, and context features for emotion classification. Experimentally, we verify the effectiveness of the proposed CD-Net on the two large emotion datasets, CAER-S and EMOTIC. On the one hand, the quantitative evaluation results demonstrate the superiority of the proposed CD-Net over other state-of-the-art methods. On the other hand, the visualization results show CD-Net can capture the dependencies among the face, body, and context components and focus on the important features related to the emotion.

JSH-Net: joint semantic segmentation and height estimation using deep convolutional networks from single high-resolution remote sensing imagery

ABSTRACT Semantic segmentation for high-resolution remote sensing imagery is a pivotal component of land use and land cover categorization, and height estimation is essential for rebuilding the 3D information of an image. Because of the higher intra-class variation and smaller inter-class dissimilarity, these two challenging tasks are generally treated separately. This paper proposes a fully convolutional network that can tackle these problems simultaneously by estimating the land-cover categories and height values of pixels from a single aerial image. To handle these tasks, we develop a multi-task learning architecture (JSH-Net) that employs a shared feature representation and exploits their potential consistency across tasks, resulting in robust features and better prediction accuracy. Specifically, we propose a novel skip connection module that aggregates the contexts from the encoder part to the decoder part, bridging the semantic gap between them. In addition, we propose a progressive refinement strategy to recover detailed information about the objects. Moreover, we also proposed a height estimation branch on the head of the model to utilize shared features. The experiments we conducted on ISPRS 2D Labelling dataset verified that our network provided precise results of semantic segmentation and height estimation from two output branches and outperformed other state-of-the-art approaches.

Evidence for a Selective Influence of Short-Term Experiences on the Retrieval of Item-Specific Long-Term Bindings.

Human behavior is guided by prior experience such as bindings between stimuli and responses. Experimentally, this is evident in performance changes when features of the stimulus-response episode reoccur either in the short-term or in the long-term. So far, effects of short-term and long-term bindings are assumed to be independent from one another. In a large-scale re-analysis of eight item-specific stimulus-response priming experiments that orthogonally varied task-specific classifications and actions in the short-term (trial N-1 to trial N) and, item-specifically, in the long-term (lag of several trials), we tested this independence assumption. In detail, we tested whether short-term experiences (repetitions of classification and action features in two consecutive trials) affected the retrieval of item-specific long-term stimulus-classification (S-C) and stimulus-action (S-A) bindings as well as potential long-term C-A bindings. The retrieval of item-specific long-term S-C bindings (i.e., the size of item-specific S-C priming effects) was affected by the persisting activation of classifications from trial N-1 (short-term priming). There were no further interactions between short-term experiences and long-term bindings. These results suggest a feature-specific, selective influence of short-term priming on long-term binding retrieval (e.g., based on shared feature representations). In contrast, however, we found evidence against an influence of short-term C-A bindings on long-term binding retrieval. This finding suggests that the processes contributing to short-term priming and long-term binding retrieval are dissociable from short-term binding and retrieval processes. Our results thus inform current theories on how short-term and long-term bindings are bound and retrieved (e.g., the BRAC framework).

Multi-task multi-level feature adversarial network for joint Alzheimer’s disease diagnosis and atrophy localization using sMRI

Capitalizing on structural magnetic resonance imaging (sMRI), existing deep learning methods (especially convolutional neural networks, CNNs) have been widely and successfully applied to computer-aided diagnosis of Alzheimer’s disease (AD) and its prodromal stage (i.e. mild cognitive impairment, MCI). But considering the generalization capability of the obtained model trained on limited number of samples, we construct a multi-task multi-level feature adversarial network (M2FAN) for joint diagnosis and atrophy localization using baseline sMRI. Specifically, the linear-aligned T1 MR images were first processed by a lightweight CNN backbone to capture the shared intermediate feature representations, which were then branched into a global subnet for preliminary dementia diagnosis and a multi instance learning network for brain atrophy localization in multi-task learning manner. As the global discriminative information captured by the global subnet might be unstable for disease diagnosis, we further designed a module of multi-level feature adversarial learning that accounts for regularization to make global features robust against the adversarial perturbation synthesized by the local/instance features to improve the diagnostic performance. Our proposed method was evaluated on three public datasets (i.e. ADNI-1, ADNI-2, and AIBL), demonstrating competitive performance compared with several state-of-the-art methods in both tasks of AD diagnosis and MCI conversion prediction.