Search

Few-Shot Class-Incremental Learning (FSCIL) techniques are essential for developing Deep Learning (DL) models that can continuously learn new classes with limited samples while retaining existing knowledge. This capability is particularly crucial for DL-based retinal disease diagnosis system, where acquiring large annotated datasets is challenging, and disease phenotypes evolve over time. This paper introduces Re-FSCIL, a novel framework for Few-Shot Class-Incremental Retinal Disease Recognition (FSCIRDR). Re-FSCIL integrates the RETFound model with a fine-grained module, employing a forward-compatible training strategy to improve adaptability, supervised contrastive learning to enhance feature discrimination, and feature fusion for robust representation quality. We convert existing datasets into the FSCIL format and reproduce numerous representative FSCIL methods to create two new benchmarks, RFMiD38 and JSIEC39, specifically for FSCIRDR. Our experimental results demonstrate that Re-FSCIL achieves State-of-the-art (SOTA) performance, significantly surpassing existing FSCIL methods on these benchmarks.

Psychiatric diseases are bringing heavy burdens for both individual health and social stability. The accurate and timely diagnosis of the diseases is essential for effective treatment and intervention. Thanks to the rapid development of brain imaging technology and machine learning algorithms, diagnostic classification of psychiatric diseases can be achieved based on brain images. However, due to divergences in scanning machines or parameters, the generalization capability of diagnostic classification models has always been an issue. We propose Meta-learning with Meta batch normalization and Distance Constraint (M 2 DC) for training diagnostic classification models. The framework can simulate the train-test domain shift situation and promote intra-class cohesion, as well as inter-class separation, which can lead to clearer classification margins and more generalizable models. To better encode dynamic brain graphs, we propose a concatenated spatiotemporal attention graph isomorphism network (CSTAGIN) as the backbone. The network is trained for the diagnostic classification of major depressive disorder (MDD) based on multi-site brain graphs. Extensive experiments on brain images from over 3261 subjects show that models trained by M 2 DC achieve the best performance on cross-site diagnostic classification tasks compared to various contemporary domain generalization methods and SOTA studies. The proposed M 2 DC is by far the first framework for multi-source closed-set domain generalizable training of diagnostic classification models for MDD and the trained models can be applied to reliable auxiliary diagnosis on novel data.

Predicting cognitive load is a crucial issue in the emerging field of human-computer interaction and holds significant practical value, particularly in flight scenarios. Although previous studies have realized efficient cognitive load classification, new research is still needed to adapt the current state-of-the-art multimodal fusion methods. Here, we proposed a feature selection framework based on multiview learning to address the challenges of information redundancy and reveal the common physiological mechanisms underlying cognitive load. Specifically, the multimodal signal features (EEG, EDA, ECG, EOG, & eye movements) at three cognitive load levels were estimated during multiattribute task battery (MATB) tasks performed by 22 healthy participants and fed into a feature selection-multiview classification with cohesion and diversity (FS-MCCD) framework. The optimized feature set was extracted from the original feature set by integrating the weight of each view and the feature weights to formulate the ranking criteria. The cognitive load prediction model, evaluated using real-time classification results, achieved an average accuracy of 81.08% and an average F1-score of 80.94% for three-class classification among 22 participants. Furthermore, the weights of the physiological signal features revealed the physiological mechanisms related to cognitive load. Specifically, heightened cognitive load was linked to amplified δ and θ power in the frontal lobe, reduced α power in the parietal lobe, and an increase in pupil diameter. Thus, the proposed multimodal feature fusion framework emphasizes the effectiveness and efficiency of using these features to predict cognitive load.

In image processing, images are usually composed of partial views due to the uncertainty of collection and how to efficiently process these images, which is called incomplete multi-view learning, has attracted widespread attention. The incompleteness and diversity of multi-view data enlarges the difficulty of annotation, resulting in the divergence of label distribution between the training and testing data, named as label shift. However, existing incomplete multi-view methods generally assume that the label distribution is consistent and rarely consider the label shift scenario. To address this new but important challenge, we propose a novel framework termed as Incomplete Multi-view Learning under Label Shift (IMLLS). In this framework, we first give the formal definitions of IMLLS and the bidirectional complete representation which describes the intrinsic and common structure. Then, a multilayer perceptron which combines the reconstruction and classification loss is employed to learn the latent representation, whose existence, consistency and universality are proved with the theoretical satisfaction of label shift assumption. After that, to align the label distribution, the learned representation and trained source classifier are used to estimate the importance weight by designing a new estimation scheme which balances the error generated by finite samples in theory. Finally, the trained classifier reweighted by the estimated weight is fine-tuned to reduce the gap between the source and target representations. Extensive experimental results validate the effectiveness of our algorithm over existing state-of-the-arts methods in various aspects, together with its effectiveness in discriminating schizophrenic patients from healthy controls.

Unsupervised domain adaptation (UDA) aims at learning a classifier for an unlabeled target domain by transferring knowledge from a labeled source domain with a related but different distribution. Most existing approaches learn domain-invariant features by adapting the entire information of the images. However, forcing adaptation of domain-specific variations undermines the effectiveness of the learned features. To address this problem, we propose a novel, yet elegant module, called the deep ladder-suppression network (DLSN), which is designed to better learn the cross-domain shared content by suppressing domain-specific variations. Our proposed DLSN is an autoencoder with lateral connections from the encoder to the decoder. By this design, the domain-specific details, which are only necessary for reconstructing the unlabeled target data, are directly fed to the decoder to complete the reconstruction task, relieving the pressure of learning domain-specific variations at the later layers of the shared encoder. As a result, DLSN allows the shared encoder to focus on learning cross-domain shared content and ignores the domain-specific variations. Notably, the proposed DLSN can be used as a standard module to be integrated with various existing UDA frameworks to further boost performance. Without whistles and bells, extensive experimental results on four gold-standard domain adaptation datasets, for example: 1) Digits; 2) Office31; 3) Office-Home; and 4) VisDA-C, demonstrate that the proposed DLSN can consistently and significantly improve the performance of various popular UDA frameworks.

Current brain cognitive models are insufficient in handling outliers and dynamics of electroencephalogram (EEG) signals. This article presents a novel self-paced dynamic infinite mixture model to infer the dynamics of EEG fatigue signals. The instantaneous spectrum features provided by ensemble wavelet transform and Hilbert transform are extracted to form four fatigue indicators. The covariance of log likelihood of the complete data is proposed to accurately identify similar components and dynamics of the developed mixture model. Compared with its seven peers, the proposed model shows better performance in automatically identifying a pilot's brain workload.

Pilots' brain fatigue status recognition faces two important issues. They are how to extract brain cognitive features and how to identify these fatigue characteristics. In this article, a gamma deep belief network is proposed to extract multilayer deep representations of high-dimensional cognitive data. The Dirichlet distributed connection weight vector is upsampled layer by layer in each iteration, and then the hidden units of the gamma distribution are downsampled. An effective upper and lower Gibbs sampler is formed to realize the automatic reasoning of the network structure. In order to extract the 3-D instantaneous time-frequency distribution spectrum of electroencephalogram (EEG) signals and avoid signal modal aliasing, this article also proposes a smoothed pseudo affine Wigner-Ville distribution method. Finally, experimental results show that our model achieves satisfactory results in terms of both recognition accuracy and stability.

In real-world applications, not all instances in the multiview data are fully represented. To deal with incomplete data, incomplete multiview learning (IML) rises. In this article, we propose the joint embedding learning and low-rank approximation (JELLA) framework for IML. The JELLA framework approximates the incomplete data by a set of low-rank matrices and learns a full and common embedding by linear transformation. Several existing IML methods can be unified as special cases of the framework. More interestingly, some linear transformation-based complete multiview methods can be adapted to IML directly with the guidance of the framework. Thus, the JELLA framework improves the efficiency of processing incomplete multiview data, and bridges the gap between complete multiview learning and IML. Moreover, the JELLA framework can provide guidance for developing new algorithms. For illustration, within the framework, we propose the IML with the block-diagonal representation (IML-BDR) method. Assuming that the sampled examples have an approximate linear subspace structure, IML-BDR uses the block-diagonal structure prior to learning the full embedding, which would lead to more correct clustering. A convergent alternating iterative algorithm with the successive over-relaxation optimization technique is devised for optimization. The experimental results on various datasets demonstrate the effectiveness of IML-BDR.

Unsupervised Domain Adaptation (UDA) aims to learn a classifier for the unlabeled target domain by leveraging knowledge from a labeled source domain with a different but related distribution. Many existing approaches typically learn a domain-invariant representation space by directly matching the marginal distributions of the two domains. However, they ignore exploring the underlying discriminative features of the target data and align the cross-domain discriminative features, which may lead to suboptimal performance. To tackle these two issues simultaneously, this paper presents a Joint Clustering and Discriminative Feature Alignment (JCDFA) approach for UDA, which is capable of naturally unifying the mining of discriminative features and the alignment of class-discriminative features into one single framework. Specifically, in order to mine the intrinsic discriminative information of the unlabeled target data, JCDFA jointly learns a shared encoding representation for two tasks: supervised classification of labeled source data, and discriminative clustering of unlabeled target data, where the classification of the source domain can guide the clustering learning of the target domain to locate the object category. We then conduct the cross-domain discriminative feature alignment by separately optimizing two new metrics: 1) an extended supervised contrastive learning, i.e., semi-supervised contrastive learning 2) an extended Maximum Mean Discrepancy (MMD), i.e., conditional MMD, explicitly minimizing the intra-class dispersion and maximizing the inter-class compactness. When these two procedures, i.e., discriminative features mining and alignment are integrated into one framework, they tend to benefit from each other to enhance the final performance from a cooperative learning perspective. Experiments are conducted on four real-world benchmarks (e.g., Office-31, ImageCLEF-DA, Office-Home and VisDA-C). All the results demonstrate that our JCDFA can obtain remarkable margins over state-of-the-art domain adaptation methods. Comprehensive ablation studies also verify the importance of each key component of our proposed algorithm and the effectiveness of combining two learning strategies into a framework.

The micro-Doppler (m-D) effect caused by micro-motion degrades the readability of the inverse synthetic aperture radar (ISAR) image. To achieve well-focused ISAR image of the target with the micro-motion part, this paper proposes a novel approach for the removal of m-D effect of ISAR image. Note that the range profiles of the rigid body are similar to each other, making the respective data matrix low-rank. Those of the micro-motion part, in contrary, generally fluctuate in different range cells, whose data matrix is sparse. Therefore, the removal of m-D effect can be naturally solved by the robust principal component analysis (RPCA)-a convenient convex program to decompose an auxiliary matrix into a low-rank matrix and a sparse one. In RPCA, the rank of a matrix is described by the nuclear norm, which is convex but leads to a suboptimal solution. To address it, we utilize a nonconvex surrogate, i.e., the summation of logistic function of the singular values of a matrix, to approximate the rank. Moreover, the range profiles of the rigid body are generally locally similar. To capture this geometric structured information, we further introduce a Laplacian regularization into the model. Then, the Laplacian regularized nonconvex low-rank (LRNL) model is solved efficiently by the linearized alternating direction method (ADM). Extensive experimental results based on both simulated and measured data demonstrate the effectiveness of the proposed approach on the removal of m-D effect of ISAR image.

Inverse synthetic aperture radar (ISAR) imaging for the sparse aperture data is affected by considerable artifacts, because under-sampling of data produces high-level grating and side lobes. Noting the ISAR image generally exhibits strong sparsity, it is often obtained by sparse signal recovery (SSR) in case of sparse aperture. The image obtained by SSR, however, is often dominated by strong isolated scatterers, resulting in difficulty to recognize the structure of target. This paper proposes a novel approach to enhance the ISAR image obtained from the sparse aperture data. Although the scatterers of target are isolated in the ISAR image, they should be associated with the neighborhood to reflect some intrinsic structural information of the target. A convolutional reweighted l 1 minimization model, therefore, is proposed to model the structural sparsity of ISAR image. Specifically, the ISAR image is reconstructed by solving a sequence of reweighted l 1 problems, where the weight of each pixel used for the next iteration is calculated from the convolution of its neighbor values in the current solution. The problem is solved by the alternating direction of multipliers (ADMM) and linearized approximation, respectively, to improve the computational efficiency. Experimental results based on both simulated and measured data validate that the proposed algorithm is effective to enhance the ISAR image, robust to noise, and more impressively, very efficient to implement.

Objective: To introduce a novel event-related potential (ERP)-based brain-computer interface (BCI) paradigm with active mental tasks multiplying precise judgment and visual cognitive capacities and evaluate its performance., Methods: This study employed a paradigm with three types of targets (true-, pseudo-, and non-), double flash codes, colors and color terms, and four test conditions. The primary hypothesis was that active mental tasks combining multiple cognitive capacities and clear judgment for different categories of stimuli increase the BCI performance and evoke stronger or specific ERPs. Classification methods were proposed and evaluated, and two were used in online experiments., Results: The modes containing active mental tasks provided higher accuracy than the control mode (by up to 19.06%). The color-word matching mode had the highest judgment level and achieved the best performance. True-stimuli evoked strong P3b, while pseudotarget signals provided obvious N4, but the control mode seemed less sensitive to both of them. Different types of stimuli evoked distinctive N2 and P3a components., Conclusion: An appropriate boost in the judgment level using multiple stimuli and cognitive approaches could be investigated to improve the BCI performance and evoke or enhance ERPs. Utilizing active mental tasks may be a promising way to promote BCIs., Significance: Active mental tasks combining multiple cognitive capacities and precise judgments were adopted in an ERP-based BCI. Color and color words were introduced as stimuli to construct an alternative paradigm, and the judgment levels of different conditions were calculated. High accuracies and the participants' preferences were obtained, which may promote the effective use of BCIs.

In this work, we conduct comprehensive comparisons between four variants of independent component analysis (ICA) methods and three variants of sparse dictionary learning (SDL) methods, both at the subject-level, by using synthesized fMRI data with ground-truth. Our results showed that ICA methods perform very well and slightly better than SDL methods when functional networks' spatial overlaps are minor, but ICA methods have difficulty in differentiating functional networks with moderate or significant spatial overlaps. In contrast, the SDL algorithms perform consistently well no matter how functional networks spatially overlap, and importantly, SDL methods are significantly better than ICA methods when spatial overlaps between networks are moderate or severe. This work offers empirical better understanding of ICA and SDL algorithms in inferring functional networks from fMRI data and provides new guidelines and caveats when constructing and interpreting functional networks in the era of fMRI-based connectomics.

Salient region detection is a challenging problem and an important topic in computer vision. It has a wide range of applications, such as object recognition and segmentation. Many approaches have been proposed to detect salient regions using different visual cues, such as compactness, uniqueness, and objectness. However, each visual cue-based method has its own limitations. After analyzing the advantages and limitations of different visual cues, we found that compactness and local contrast are complementary to each other. In addition, local contrast can very effectively recover incorrectly suppressed salient regions using compactness cues. Motivated by this, we propose a bottom-up salient region detection method that integrates compactness and local contrast cues. Furthermore, to produce a pixel-accurate saliency map that more uniformly covers the salient objects, we propagate the saliency information using a diffusion process. Our experimental results on four benchmark data sets demonstrate the effectiveness of the proposed method. Our method produces more accurate saliency maps with better precision-recall curve and higher F-Measure than other 19 state-of-the-arts approaches on ASD, CSSD, and ECSSD data sets.

The aim of this study was to design a dynamically optimized steady-state visually evoked potential (SSVEP) brain-computer interface (BCI) system with enhanced performance relative to previous SSVEP BCIs in terms of the number of items selectable on the interface, accuracy, and speed. In this approach, the row/column (RC) paradigm was employed in a SSVEP speller to increase the number of items. The target is detected by subsequently determining the row and column coordinates. To improve spelling accuracy, we added a posterior processing after the canonical correlation analysis (CCA) approach to reduce the interfrequency variation between different subjects and named the new signal processing method CCA-RV, and designed a real-time biofeedback mechanism to increase attention on the visual stimuli. To achieve reasonable online spelling speed, both fixed and dynamic approaches for setting the optimal stimulus duration were implemented and compared. Experimental results for 11 subjects suggest that the CCA-RV method and the real-time biofeedback effectively increased accuracy compared with CCA and the absence of real-time feedback, respectively. In addition, both optimization approaches for setting stimulus duration achieved reasonable online spelling performance. However, the dynamic optimization approach yielded a higher practical information transfer rate (PITR) than the fixed optimization approach. The average online PITR achieved by the proposed adaptive SSVEP speller, including the time required for breaks between selections and error correction, was 41.08 bit/min. These results indicate that our BCI speller is promising for use in SSVEP-based BCI applications.

When analyzing brain imaging data, blind source separation (BSS) techniques critically depend on the level of dimensional reduction. If the reduction level is too slight, the BSS model would be overfitted and become unavailable. Thus, the reduction level must be set relatively heavy. This approach risks discarding useful information and crucially limits the performance of BSS techniques. In this study, a new BSS method that can work well even at a slight reduction level is presented. We proposed the concept of "signal intensity" which measures the significance of the source. Only picking the sources with significant intensity, the new method can avoid the overfitted solutions which are nonexistent artifacts. This approach enables the reduction level to be set slight and retains more useful dimensions in the preliminary reduction. Comparisons between the new and conventional algorithms were performed on both simulated and real data.

This study proposes a novel hybrid brain-computer interface (BCI) approach for increasing the spelling speed. In this approach, the P300 and steady-state visually evoked potential (SSVEP) detection mechanisms are devised and integrated so that the two brain signals can be used for spelling simultaneously. Specifically, the target item is identified by 2-D coordinates that are realized by the two brain patterns. The subarea/location and row/column speedy spelling paradigms were designed based on this approach. The results obtained for 14 healthy subjects demonstrate that the average online practical information transfer rate, including the time of break between selections and error correcting, achieved using our approach was 53.06 bits/min. The pilot studies suggest that our BCI approach could achieve higher spelling speed compared with the conventional P300 and SSVEP spellers.

Reinforcement learning is a powerful mechanism for enabling agents to learn in an unknown environment, and most reinforcement learning algorithms aim to maximize some numerical value, which represents only one long-term objective. However, multiple long-term objectives are exhibited in many real-world decision and control problems; therefore, recently, there has been growing interest in solving multiobjective reinforcement learning (MORL) problems with multiple conflicting objectives. The aim of this paper is to present a comprehensive overview of MORL. In this paper, the basic architecture, research topics, and naive solutions of MORL are introduced at first. Then, several representative MORL approaches and some important directions of recent research are reviewed. The relationships between MORL and other related research are also discussed, which include multiobjective optimization, hierarchical reinforcement learning, and multi-agent reinforcement learning. Finally, research challenges and open problems of MORL techniques are highlighted.

We present a novel artery-vein separation method using 0.1-Hz oscillation at two wavelengths with optical imaging of intrinsic signals (OIS). The 0.1-Hz oscillation at a green light wavelength of 546 nm exhibits greater amplitude in arteries than in veins and is primarily caused by vasomotion, whereas the 0.1-Hz oscillation at a red light wavelength of 630 nm exhibits greater amplitude in veins than in arteries and is primarily caused by changes of deoxyhemoglobin concentration. This spectral feature enables cortical arteries and veins to be segmented independently. The arteries can be segmented on the 0.1-Hz amplitude image at 546 nm using matched filters of a modified dual Gaussian model combining with a single Gaussian model. The veins are a combination of vessels segmented on both amplitude images at the two wavelengths using multiscale matched filters of single Gaussian model. Our method can separate most of the thin arteries and veins from each other, especially the thin arteries with low contrast in raw gray images. In vivo OIS experiments demonstrate the separation ability of the 0.1-Hz based segmentation method in cerebral cortex of eight rats. Two validation studies were undertaken to evaluate the performance of the method by quantifying the arterial and venous length based on a reference standard. The results indicate that our 0.1-Hz method is very effective in separating both large and thin arteries and veins regardless of vessel crossover or overlapping to great extent in comparison with previous methods.

In recent years, approximate policy iteration (API) has attracted increasing attention in reinforcement learning (RL), e.g., least-squares policy iteration (LSPI) and its kernelized version, the kernel-based LSPI algorithm. However, it remains difficult for API algorithms to obtain near-optimal policies for Markov decision processes (MDPs) with large or continuous state spaces. To address this problem, this paper presents a hierarchical API (HAPI) method with binary-tree state space decomposition for RL in a class of absorbing MDPs, which can be formulated as time-optimal learning control tasks. In the proposed method, after collecting samples adaptively in the state space of the original MDP, a learning-based decomposition strategy of sample sets was designed to implement the binary-tree state space decomposition process. Then, API algorithms were used on the sample subsets to approximate local optimal policies of sub-MDPs. The original MDP was decomposed into a binary-tree structure of absorbing sub-MDPs, constructed during the learning process, thus, local near-optimal policies were approximated by API algorithms with reduced complexity and higher precision. Furthermore, because of the improved quality of local policies, the combined global policy performed better than the near-optimal policy obtained by a single API algorithm in the original MDP. Three learning control problems, including path-tracking control of a real mobile robot, were studied to evaluate the performance of the HAPI method. With the same setting for basis function selection and sample collection, the proposed HAPI obtained better near-optimal policies than previous API methods such as LSPI and KLSPI.

Separating brain imaging signals by maximizing their autocorrelations is an important component of blind source separation (BSS). Canonical correlation analysis (CCA), one of leading BSS techniques, has been widely used for analyzing optical imaging (OI) and functional magnetic resonance imaging (fMRI) data. However, because of the need to reduce dimensionality and ignore spatial autocorrelation, CCA is problematic for separating temporal signal sources. To solve the problems of CCA, "straightforward image projection" (SIP) has been incorporated into temporal BSS. This novel method, termed low-dimensional canonical correlation analysis (LD-CCA), relies on the spatial and temporal autocorrelations of all genuine signals of interest. Incorporating both spatial and temporal information, here we introduce a "generalized timecourse" technique in which data are artificially reorganized prior to separation. The quantity of spatial plus temporal autocorrelations can then be defined. By maximizing temporal and spatial autocorrelations in combination, LD-CCA is able to obtain expected "real" signal sources. Generalized timecourses are low-dimensional, eliminating the need for dimension reduction. This removes the risk of discarding useful information. The new method is compared with temporal CCA and temporal independent component analysis (tICA). Comparison of simulated data showed that LD-CCA was more effective for recovering signal sources. Comparisons using real intrinsic OI and fMRI data also supported the validity of LD-CCA.

This paper is concerned with the global convergence of the solutions of a nonautonomous system with variable delays, arising from the description of the states of neurons in delayed Cohen-Grossberg type in a time-varying situation. By exploring intrinsic features between nonautonomous system and its asymptotic equation, several novel sufficient conditions are established to ensure that all solutions of the networks converge to a periodic function or a constant vector for delayed Cohen-Grossberg-type neural network (NN) models in time-varying situation. The results can be applied directly to group of NNs models including Hopfield NNs, bidirectional association memory NNs, and cellular NNs. Our results are not only presented in terms of system parameters and can be easily verified but also are less restrictive than previously known criteria. Numerical simulations have also been presented to demonstrate the theoretical analysis.

In this paper, we present a kernel-based least squares policy iteration (KLSPI) algorithm for reinforcement learning (RL) in large or continuous state spaces, which can be used to realize adaptive feedback control of uncertain dynamic systems. By using KLSPI, near-optimal control policies can be obtained without much a priori knowledge on dynamic models of control plants. In KLSPI, Mercer kernels are used in the policy evaluation of a policy iteration process, where a new kernel-based least squares temporal-difference algorithm called KLSTD-Q is proposed for efficient policy evaluation. To keep the sparsity and improve the generalization ability of KLSTD-Q solutions, a kernel sparsification procedure based on approximate linear dependency (ALD) is performed. Compared to the previous works on approximate RL methods, KLSPI makes two progresses to eliminate the main difficulties of existing results. One is the better convergence and (near) optimality guarantee by using the KLSTD-Q algorithm for policy evaluation with high precision. The other is the automatic feature selection using the ALD-based kernel sparsification. Therefore, the KLSPI algorithm provides a general RL method with generalization performance and convergence guarantee for large-scale Markov decision problems (MDPs). Experimental results on a typical RL task for a stochastic chain problem demonstrate that KLSPI can consistently achieve better learning efficiency and policy quality than the previous least squares policy iteration (LSPI) algorithm. Furthermore, the KLSPI method was also evaluated on two nonlinear feedback control problems, including a ship heading control problem and the swing up control of a double-link underactuated pendulum called acrobot. Simulation results illustrate that the proposed method can optimize controller performance using little a priori information of uncertain dynamic systems. It is also demonstrated that KLSPI can be applied to online learning control by incorporating an initial controller to ensure online performance.