Dynamic Subspace Research Articles

Characterization of the temporal stability of ToM and pain functional brain networks carry distinct developmental signatures during naturalistic viewing

A temporally stable functional brain network pattern among coordinated brain regions is fundamental to stimulus selectivity and functional specificity during the critical period of brain development. Brain networks that are recruited in time to process internal states of others’ bodies (like hunger and pain) versus internal mental states (like beliefs, desires, and emotions) of others’ minds allow us to ask whether a quantitative characterization of the stability of these networks carries meaning during early development and constrain cognition in a specific way. Previous research provides critical insight into the early development of the theory-of-mind (ToM) network and its segregation from the Pain network throughout normal development using functional connectivity. However, a quantitative characterization of the temporal stability of ToM networks from early childhood to adulthood remains unexplored. In this work, reusing a large sample of children (n = 122, 3–12 years) and adults (n = 33) dataset that is available on the OpenfMRI database under the accession number ds000228, we addressed this question based on their fMRI data during a short and engaging naturalistic movie-watching task. The movie highlights the characters’ bodily sensations (often pain) and mental states (beliefs, desires, emotions), and is a feasible experiment for young children. Our results tracked the change in temporal stability using an unsupervised characterization of ToM and Pain networks DFC patterns using Angular and Mahalanobis distances between dominant dynamic functional connectivity subspaces. Our findings reveal that both ToM and Pain networks exhibit lower temporal stability as early as 3-years and gradually stabilize by 5-years, which continues till adolescence and late adulthood (often sharing similarity with adult DFC stability patterns). Furthermore, we find that the temporal stability of ToM brain networks is associated with the performance of participants in the false belief task to access mentalization at an early age. Interestingly, higher temporal stability is associated with the pass group, and similarly, moderate and low temporal stability are associated with the inconsistent group and the fail group. Our findings open an avenue for applying the temporal stability of large-scale functional brain networks during cortical development to act as a biomarker for multiple developmental disorders concerning impairment and discontinuity in the neural basis of social cognition.

Continual learning for seizure prediction via memory projection strategy

Despite extensive algorithms for epilepsy prediction via machine learning, most models are tailored for offline scenarios and cannot handle actual scenarios where data changes over time. Catastrophic forgetting(CF) for learned electroencephalogram(EEG) data occurs when EEG changes dynamically in the clinical setting. This paper implements a continual learning(CL) strategy Memory Projection(MP) for epilepsy prediction, which can be combined with other algorithms to avoid CF. Such a strategy enables the model to learn EEG data from each patient in dynamic subspaces with weak correlation layer by layer to minimize interference and promote knowledge transfer. Regularization Loss Reconstruction Algorithm and Matrix Dimensionality Reduction Algorithm are introduced into the core of MP. Experimental results show that MP exhibits excellent performance and low forgetting rates in sequential learning of seizure prediction. The forgetting rate of accuracy and sensitivity under multiple experiments are below 5%. When learning from multi-center datasets, the forgetting rates for accuracy and sensitivity decrease to 0.65% and 1.86%, making it comparable to state-of-the-art CL strategies. Through ablation experiments, we have analyzed that MP can operate with minimal storage and computational cost, which demonstrates practical potential for seizure prediction in clinical scenarios.

Domain adaptive learning based on equilibrium distribution and dynamic subspace approximation

Nowadays, big data analysis has become an important approach in social information network. However, the social information may not be distributed independently and identically (i.i.d.), which can be addressed using domain adaptation. However, most of the existing domain adaptation methods are designed to align cross-domain distributions. The label information of the samples in the target domain is completely unavailable. Thus, the class-conditional distribution differences cannot be well measured, and the effect of feature distortion on distribution alignment in the original feature space is difficult to handle. This paper proposes a domain adaptation learning based on the Equilibrium Distribution and Dynamic Subspace Approximation (EDDSA) to alleviate these problems. First, EDDSA learns to project the source and target domains into associated feature spaces, and dynamically approximates two subspaces to overcome the feature distortion problem. Second, the balanced distribution alignment term is introduced to dynamically weight the importance of the conditional and marginal distributions. Through many experiments, EDDSA is superior to most traditional methods.

Dynamic Active Subspaces for Model Predictive Allocation in Over-Actuated Systems

Dynamic Active Subspaces for Model Predictive Allocation in Over-Actuated Systems

Spatial Dynamic Subspaces Encode Sex-Specific Schizophrenia Disruptions in Transient Network Overlap and Their Links to Genetic Risk

BackgroundSchizophrenia research reveals sex differences in incidence, symptoms, genetic risk factors, and brain function. However, a knowledge gap remains regarding sex-specific schizophrenia alterations in brain function. Schizophrenia is considered a dysconnectivity syndrome, but the dynamic integration and segregation of brain networks are poorly understood. Recent advances in resting-state functional magnetic resonance imaging allow us to study spatial dynamics, the phenomenon of brain networks spatially evolving over time. Nevertheless, estimating time-resolved networks remains challenging due to low signal-to-noise ratio, limited short-time information, and uncertain network identification. MethodsWe adapted a reference-informed network estimation technique to capture time-resolved networks and their dynamic spatial integration and segregation for 193 individuals with schizophrenia and 315 control participants. We focused on time-resolved spatial functional network connectivity, an estimate of network spatial coupling, to study sex-specific alterations in schizophrenia and their links to genomic data. ResultsOur findings are consistent with the dysconnectivity and neurodevelopment hypotheses and with the cerebello-thalamo-cortical, triple-network, and frontoparietal dysconnectivity models, helping to unify them. The potential unification offers a new understanding of the underlying mechanisms. Notably, the posterior default mode/salience spatial functional network connectivity exhibits sex-specific schizophrenia alteration during the state with the highest global network integration and is correlated with genetic risk for schizophrenia. This dysfunction is reflected in regions with weak functional connectivity to corresponding networks. ConclusionsOur method can effectively capture spatially dynamic networks, detect nuanced schizophrenia effects including sex-specific ones, and reveal the intricate relationship of dynamic information to genomic data. The results also underscore the clinical potential of dynamic spatial dependence and weak connectivity.

Dynamic stationary subspace analysis based on Gaussian mixture models for ironmaking process monitoring

AbstractBlast furnace ironmaking process monitoring is an important and challenging task. Due to the influence of hot blast stove switching and large fluctuations in the quality of raw materials, the measurements of ironmaking processes show obvious non‐stationary characteristics, and in addition, the observed data are also characterized by time‐series dynamic and non‐Gaussian characteristics. In this paper, a dynamic stationary subspace analysis method based on the Gaussian mixture model (DSSA–GMM) is proposed to address the difficulties in blast furnace ironmaking process monitoring. The time‐series dynamic relationship of the data is conducted by introducing a sliding time window. The Gaussian mixture model (GMM) is used to deal with the non‐Gaussian characteristics of the data, and the parameters of the GMMs are estimated using the expectation–maximization algorithm. The stationary projection matrix is obtained by optimizing the Kullback–Leibler (K–L) divergence between GMMs of different periods to realize the stationary subspace separation. Finally, the convex hull of the stationary subspace is established for fault detection, thus realizing the monitoring for non‐stationary and non‐Gaussian dynamic processes. The effectiveness of the DSSA–GMM method is verified by a numerical simulation and a dataset collected from an actual blast furnace ironmaking process.

Low-rank approximation of Hankel matrices in denoising applications for statistical damage diagnosis of wind turbine blades

Model order selection is a fundamental task in subspace identification for estimation of modal parameters, uncertainty propagation and damage diagnosis. However, the true model order and the related low-rank structure of the dynamic system are generally unknown. In this paper, a statistical methodology to actively select the dynamic signal subspace in covariance-driven subspace identification is developed on the basis of statistical analysis of the eigenvalue condition numbers of the output covariance Hankel matrix. It is shown that the condition numbers highly sensitive to random perturbations characterize the noise subspace. The signal subspace is separated from the noise subspace by analyzing two statistical parameters associated with the condition number sensitivity, whose thresholds are user-defined. A practical algorithm to retrieve the system dynamics is designed and demonstrated on a running example of a simulated wind turbine blade benchmark. The resultant framework is then applied in the context of damage detection on a medium-size wind turbine blade. It is demonstrated that the detectability of small damage is enhanced compared to the classic approaches and robustness of damage diagnosis is increased by reducing the number of false alarms.

A knee point driven Kriging-assisted multi-objective robust fuzzy clustering algorithm for image segmentation

To enhance the segmentation performance and optimization efficiency of multi-objective evolutionary clustering algorithms for noisy images, this study presents a knee point driven Kriging-assisted multi-objective robust fuzzy clustering algorithm for image segmentation (K2 MORFC). First, pixel and region-level fitness functions are developed to comprehensively consider the spatial constraints and region consistency in an image. In the pixel-level fitness function, a Kullback–Leibler (KL) divergence-based spatial constraint term is employed and combined with intra-class compactness to preserve the membership similarity between an arbitrary image pixel and its non-local neighbors. In the region-level fitness function, image edge information is utilized to evaluate the regional consistency of the partition. Moreover, an adaptive determination mechanism of the weight factor for the spatial constraint term is designed in the pixel-level fitness function to control the influence of the spatial constraint for each pixel. Furthermore, considering that the preference toward knee points can effectively improve the convergence performance, a knee point driven Kriging-assisted evolutionary algorithm (K2EA) is proposed to efficiently optimize these two fitness functions, including a dynamic subspace knee point searching strategy, a knee point guided environmental selection strategy, and a Kriging model management mechanism. Finally, we construct a fuzzy clustering validity index with a KL divergence-based non-local spatial constraint term to select the optimal solution from the nondominated solution set. Competitive experimental results on the DTLZ benchmark verify the effectiveness of K2EA. Segmentation results on color images from the Berkeley dataset and magnetic resonance images from the BrainWeb and IBSR datasets confirm the performance of K2MORFC.

Low-rank motion correction for accelerated free-breathing first-pass myocardial perfusion imaging.

Develop a novel approach for accelerated 2D free-breathing myocardial perfusion via low-rank motion-corrected (LRMC) reconstructions. Myocardial perfusion imaging requires high spatial and temporal resolution, despite scan time constraints. Here, we incorporate LRMC models into the reconstruction-encoding operator, together with high-dimensionality patch-based regularization, to produce high quality, motion-corrected myocardial perfusion series from free-breathing acquisitions. The proposed framework estimates beat-to-beat nonrigid respiratory (and any other incidental) motion and the dynamic contrast subspace from the actual acquired data, which are then incorporated into the proposed LRMC reconstruction. LRMC was compared with iterative SENSitivity Encoding(SENSE) (itSENSE) and low-rank plus sparse (LpS) reconstruction in 10 patients based on image-quality scoring and ranking by two clinical expert readers. LRMC achieved significantly improved results relative to itSENSE and LpS in terms of image sharpness, temporal coefficient of variation, and expert reader evaluation. Left ventricle image sharpness was approximately 75%, 79%, and 86% for itSENSE, LpS and LRMC, respectively, indicating improved image sharpness for the proposed approach. Corresponding temporal coefficient of variation results were 23%, 11% and 7%, demonstrating improved temporal fidelity of the perfusion signal with the proposed LRMC. Corresponding clinical expert reader scores (1-5, from poor to excellent image quality) were 3.3, 3.9 and 4.9, demonstrating improved image quality with the proposed LRMC, in agreement with the automated metrics. LRMC produces motion-corrected myocardial perfusion in free-breathing acquisitions with substantially improved image quality when compared with iterative SENSE and LpS reconstructions.

Using Dynamic Active Subspaces to Construct Surrogate Models for Calibrating Tumor Growth Models to Data

While the use of complex ODE models describing tumor growth and tumor-immune interactions may be desirable for accurately representing the biological mechanisms that govern tumor growth, the high-dimensionality of the parameter space can make it difficult to uniquely identify optimal parameter estimates during model calibration. Construction of a surrogate model that can accurately approximate the quantity of interest with fewer parameters may enable us to fit this alternate model to the data in place of the original ODE system. In this investigation, we consider how active subspaces may be used to construct time-dependent surrogate models with reduced parameter dimension, and analyze the trade-offs in parameter identifiability, model fit error, and computational run-time to provide a roadmap for ensuring an identifiable parameter set during model calibration.

Neural Cloth Simulation

We present a general framework for the garment animation problem through unsupervised deep learning inspired in physically based simulation. Existing trends in the literature already explore this possibility. Nonetheless, these approaches do not handle cloth dynamics. Here, we propose the first methodology able to learn realistic cloth dynamics unsupervisedly, and henceforth, a general formulation for neural cloth simulation. The key to achieve this is to adapt an existing optimization scheme for motion from simulation based methodologies to deep learning. Then, analyzing the nature of the problem, we devise an architecture able to automatically disentangle static and dynamic cloth subspaces by design. We will show how this improves model performance. Additionally, this opens the possibility of a novel motion augmentation technique that greatly improves generalization. Finally, we show it also allows to control the level of motion in the predictions. This is a useful, never seen before, tool for artists. We provide of detailed analysis of the problem to establish the bases of neural cloth simulation and guide future research into the specifics of this domain.

Optimized clustering method for spectral reflectance recovery.

An optimized method based on dynamic partitional clustering was proposed for the recovery of spectral reflectance from camera response values. The proposed method produced dynamic clustering subspaces using a combination of dynamic and static clustering, which determined each testing sample as a priori clustering center to obtain the clustering subspace by competition. The Euclidean distance weighted and polynomial expansion models in the clustering subspace were adaptively applied to improve the accuracy of spectral recovery. The experimental results demonstrated that the proposed method outperformed existing methods in spectral and colorimetric accuracy and presented the effectiveness and robustness of spectral recovery accuracy under different color spaces.

Attention-Based Dynamic Subspace Learners for Medical Image Analysis.

Learning similarity is a key aspect in medical image analysis, particularly in recommendation systems or in uncovering the interpretation of anatomical data in images. Most existing methods learn such similarities in the embedding space over image sets using a single metric learner. Images, however, have a variety of object attributes such as color, shape, or artifacts. Encoding such attributes using a single metric learner is inadequate and may fail to generalize. Instead, multiple learners could focus on separate aspects of these attributes in subspaces of an overarching embedding. This, however, implies the number of learners to be found empirically for each new dataset. This work, Dynamic Subspace Learners, proposes to dynamically exploit multiple learners by removing the need of knowing apriori the number of learners and aggregating new subspace learners during training. Furthermore, the visual interpretability of such subspace learning is enforced by integrating an attention module into our method. This integrated attention mechanism provides a visual insight of discriminative image features that contribute to the clustering of image sets and a visual explanation of the embedding features. The benefits of our attention-based dynamic subspace learners are evaluated in the application of image clustering, image retrieval, and weakly supervised segmentation. Our method achieves competitive results with the performances of multiple learners baselines and significantly outperforms the classification network in terms of clustering and retrieval scores on three different public benchmark datasets. Moreover, our method also provides an attention map generated directly during inference to illustrate the visual interpretability of the embedding features. These attention maps offer a proxy-labels, which improves the segmentation accuracy up to 15% in Dice scores when compared to state-of-the-art interpretation techniques.

Online Structural Change-Point Detection of High-dimensional Streaming Data via Dynamic Sparse Subspace Learning

igh-dimensional streaming data are becoming increasingly ubiquitous in many fields. They often lie in multiple low-dimensional subspaces, and the manifold structures may change abruptly on the time scale due to pattern shift or occurrence of anomalies. However, the problem of detecting the structural changes in a real-time manner has not been well studied. To fill this gap, we propose a dynamic sparse subspace learning approach for online structural change-point detection of high-dimensional streaming data. A novel multiple structural change-point model is proposed and the asymptotic properties of the estimators are investigated. A tuning method based on Bayesian information criterion and change-point detection accuracy is proposed for penalty coefficients selection. An efficient Pruned Exact Linear Time based algorithm is proposed for online optimization and change-point detection. The effectiveness of the proposed method is demonstrated through several simulation studies and a real case study on gesture data for motion tracking. Supplementary materials for this article are available online.

Dynamic subspace angle estimation method for bistatic MIMO radar with system imperfections

The development of high-resolution methods that can withstand unpleasant radar conditions for target localization is challenging. In this paper, a novel dynamic subspace method to address the problem of angle estimation of the bistatic Multiple-input multiple-output (MIMO) radar with system imperfections brought about by the coexistence of mutual coupling and coherent signals with large power differences is proposed. The proposed dynamic subspace angle estimation method comprises two subspace approaches; the dynamic subspace resampling with coupling calibration approach and the dynamic subspace superresolution with coupling calibration approach. The implementation of both approaches is such that they apply a linear transformation technique by way of exploiting the topology of the array structure to attain signal decoupling. In addition, the proposed approaches devise the decimation of the signal samples and the formulation of a beamforming framework to induce decorrelation and achieve signal resolution. Finally, a 1D pairwise spectral estimator which exploits matrix inversion mechanism coupled with a grid refining technique is introduced to resolve the target angle with obvious power differences.In contrast with the existing methods, the proposed method has an extra advantage of its applicability to colored noise scenarios due to its inherent noise suppression ability. This, therefore, augments its high suitability to imperfect systems. Evidence of the proposed algorithm's validity and effectiveness over existing methods is demonstrated in terms of numerical simulations under varying conditions.

Dynamic subspace dual-graph regularized multi-label feature selection

In multi-label learning, feature selection is a topical issue for addressing high-dimension data. However, most of existing methods adopt imperfect labels to perform feature selection. Although some graph-based multi-label feature selection methods are proposed to deal with the problem, they adopt the fixed graph Laplacian matrix so that the performances of these models are under-performing. To this end, this paper proposes a Dynamic Subspace dual-graph regularized Multi-label Feature Selection method named DSMFS. DSMFS decomposes the original label space into a low-dimensional subspace, and then both the dynamic label-level subspace graph and the feature-level graph are used to obtain a high-quality label subspace to conduct feature selection process. Seven state-of-the-art methods are compared to the proposed method on twelve multi-label benchmark data sets in the experiments. Experimental results demonstrate the superiority of DSMFS.

Quality-relevant fault detection based on adversarial learning and distinguished contribution of latent variables to quality

Quality-relevant fault detection is a primary task to reveal the changes of quality variables in process monitoring. Current works mainly focus on learning quality-relevant features, however, how to distinguish quality-relevant and irrelevant information is responsible for the excellent monitoring performance. In this study, a novel quality-relevant fault detection method is proposed on the basis of adversarial learning and distinguished contribution of latent features to quality is originally introduced. First of all, we map the input variables into a gaussian manifold in adversarial and unsupervised manner. Then a fully connected neural network is trained to learn the relationship between latent and quality variables. To distinguish necessary information in such manifold, the Jacobi operator at the corresponding point is calculated to project the latent variables into quality-relevant and quality-irrelevant subspaces. Third, fault detection is implemented in these dynamic subspaces using the probabilities of latent variables. Finally, the proposed method is evaluated by numerical example, the Tennessee-Eastman process and wind turbine blade icing process.

Bayesian learning of stochastic dynamical models

A new methodology for rigorous Bayesian learning of high-dimensional stochastic dynamical models is developed. The methodology performs parallelized computation of marginal likelihoods for multiple candidate models, integrating over all state variable and parameter values, and enabling a principled Bayesian update of model distributions. This is accomplished by leveraging the dynamically orthogonal (DO) evolution equations for uncertainty prediction in a dynamic stochastic subspace and the Gaussian Mixture Model-DO filter for inference of nonlinear state variables and parameters, using reduced-dimension state augmentation to accommodate models featuring uncertain parameters. Overall, the joint Bayesian inference of the state, model equations, geometry, boundary conditions, and initial conditions is performed. Results are exemplified using two high-dimensional, nonlinear simulated fluid and ocean systems. For the first, limited measurements of fluid flow downstream of an obstacle are used to perform joint inference of the obstacle’s shape, the Reynolds number, and the O(105) fluid velocity state variables. For the second, limited measurements of the concentration of a microorganism advected by an uncertain flow are used to perform joint inference of the microorganism’s reaction equation and the O(105) microorganism concentration and ocean velocity state variables. When the observations are sufficiently informative about the learning objectives, we find that our posterior model probabilities correctly identify either the true model or the most plausible models, even in cases where a human would be challenged to do the same.

Cardiopulmonary coupling indices to assess weaning readiness from mechanical ventilation

The ideal moment to withdraw respiratory supply of patients under Mechanical Ventilation at Intensive Care Units (ICU), is not easy to be determined for clinicians. Although the Spontaneous Breathing Trial (SBT) provides a measure of the patients’ readiness, there is still around 15–20% of predictive failure rate. This work is a proof of concept focused on adding new value to the prediction of the weaning outcome. Heart Rate Variability (HRV) and Cardiopulmonary Coupling (CPC) methods are evaluated as new complementary estimates to assess weaning readiness. The CPC is related to how the mechanisms regulating respiration and cardiac pumping are working simultaneously, and it is defined from HRV in combination with respiratory information. Three different techniques are used to estimate the CPC, including Time-Frequency Coherence, Dynamic Mutual Information and Orthogonal Subspace Projections. The cohort study includes 22 patients in pressure support ventilation, ready to undergo the SBT, analysed in the 24 h previous to the SBT. Of these, 13 had a successful weaning and 9 failed the SBT or needed reintubation –being both considered as failed weaning. Results illustrate that traditional variables such as heart rate, respiratory frequency, and the parameters derived from HRV do not differ in patients with successful or failed weaning. Results revealed that HRV parameters can vary considerably depending on the time at which they are measured. This fact could be attributed to circadian rhythms, having a strong influence on HRV values. On the contrary, significant statistical differences are found in the proposed CPC parameters when comparing the values of the two groups, and throughout the whole recordings. In addition, differences are greater at night, probably because patients with failed weaning might be experiencing more respiratory episodes, e.g. apneas during the night, which is directly related to a reduced respiratory sinus arrhythmia. Therefore, results suggest that the traditional measures could be used in combination with the proposed CPC biomarkers to improve weaning readiness.

Dynamic process monitoring using dynamic latent‐variable and canonical correlation analysis model

AbstractDynamic latent‐variable (DLV) modelling is a very effective method for dynamic process monitoring. However, the DLV method only focuses on auto‐correlation in process data but ignores the cross‐correlation between inputs and outputs. To overcome this shortcoming, a novel dynamic process monitoring method using dynamic‐latent variable and canonical correlation analysis (DLV‐CCA) is proposed. Considering the dynamics in the process data, the proposed DLV‐CCA method first utilizes the DLV method to decompose the input space into input dynamic and static subspaces. The output space is also decomposed into output dynamic and static subspaces by DLV. Then, canonical correlation analysis (CCA) is used to explore the cross‐correlation between the dynamic subspaces (including input dynamic and output dynamic subspaces) and the static subspaces (including input static and output static subspaces). According to the CCA results, residual signals are generated and corresponding Hotelling's T2 statistics are established to detect variations in these residual signals. A numerical example and a closed‐loop continuous stirred‐tank reactor (CSTR) are employed to demonstrate the superior performance of the DLV‐CCA based process monitoring compared with other relevant methods.