Higher-order Representation Research Articles

Deconfounded recommendation via causal intervention

Traditional recommenders suffer from hidden confounding factors, leading to the spurious correlations between user/item profiles and user preference prediction, i.e., the confounding bias issue. Most works resort to only one confounding bias, which greatly block their applications on recommendations with mixture confounder, i.e., more than one bias. It is therefore of practical demand to empower the recommender with the capability of debiasing different biases from data. Moreover, the positive effect of bias is neglected in most previous works. We argue that confounding bias is actually beneficial for capturing users’ preferences in some recommendation scenarios. In this paper, we propose a novel deconfounded causal learning method called GCRec (Graph Causal Recommendetion) to debias two confounders: social network confounder and item group confounder. We employ Graph Neural Networks (GNNs) to aggregate user-user connections for social networks and user-item interactions for item groups in order to learn high-order representations that can efficiently debias these two confounders from a causal view. In the inference stage, we use symmetric Kullback–Leibler divergence to measure the user preference drift. If the divergence is large, we perform the causal intervention to alleviate the bias amplification caused by confounders on user preferences. Otherwise, we incorporate the user preferences that can potentially deliver a positive effect on favoring recommendation performance. Extensive experiments are conducted on two benchmark datasets to verify that GCRec outperforms state-of-the-art methods and achieves robust recommendations.

Web API Recommendation Based on Self-Attentional Neural Factorization Machines With Domain Interactions

With the advances of service computing technology, there has been a significant focus on enhancing the relevance or accuracy of Web API recommendation for Mashup creation. In the multiple established works, factorization machines based models portray increased performance compared with traditional content or collaborative filtering techniques with regard to accuracy, which still has great room for improvement. Thus, this article takes a further step to propose a more expressive factorization machines model, i.e., Self-Attentional Neural Factorization Machines with Domain Interactions for Web API recommendation. Specifically, we introduce the concept of domain interaction to reduce the calculation complexity and capture the common characteristics of all features in one domain. Then, we leverage an implicit attention mechanism, that is, self-attention, to identify the importance of different feature interactions, instead of using explicit attention. Moreover, to promote expressive capability, deep neural networks are integrated into the transformed feature interactions following the self-attention propagation, capturing low and high order representations. In this approach, the importance of feature interactions can contribute to both low and high order relations. Multiple studies on a practical dataset indicates the inherent benefits of the method proposed in this study over the baselines for Web API recommendation.

Bivariate multiquadric quasi-interpolation operators of Lidstone type

<abstract><p>In this paper, a kind of bivariate multiquadric quasi-interpolant with the derivatives of a approximated function is studied by combining the known multiquadric quasi-interpolant with the generalized Taylor polynomials that act as the bivariate Lidstone interpolation polynomials. For practical purposes, a kind of improved approximation operator without any derivative of the approximated function is given by using bivariate divided differences to approximate the derivatives. It has the property of high-degree polynomial reproducing. In addition, the improved bivariate quasi-interpolation operators only demand information of the location points rather than the derivatives of the function approximated. Some error bounds in terms of the modulus of continuity of high order and Peano representations for the error are given. Several numerical comparisons with other existing methods are carried out to verify a higher degree of accuracy based on the obtained scheme. Furthermore, the advantage of our method is that the algorithm is very simple and easy to implement.</p></abstract>

Hierarchical Graph Pattern Understanding for Zero-Shot Video Object Segmentation.

The optical flow guidance strategy is ideal for obtaining motion information of objects in the video. It is widely utilized in video segmentation tasks. However, existing optical flow-based methods have a significant dependency on optical flow, which results in poor performance when the optical flow estimation fails for a particular scene. The temporal consistency provided by the optical flow could be effectively supplemented by modeling in a structural form. This paper proposes a new hierarchical graph neural network (GNN) architecture, dubbed hierarchical graph pattern understanding (HGPU), for zero-shot video object segmentation (ZS-VOS). Inspired by the strong ability of GNNs in capturing structural relations, HGPU innovatively leverages motion cues (i.e., optical flow) to enhance the high-order representations from the neighbors of target frames. Specifically, a hierarchical graph pattern encoder with message aggregation is introduced to acquire different levels of motion and appearance features in a sequential manner. Furthermore, a decoder is designed for hierarchically parsing and understanding the transformed multi-modal contexts to achieve more accurate and robust results. HGPU achieves state-of-the-art performance on four publicly available benchmarks (DAVIS-16, YouTube-Objects, Long-Videos and DAVIS-17). Code and pre-trained model can be found at https://github.com/NUST-Machine-Intelligence-Laboratory/HGPU.

High-Order Semantic Decoupling Network for Remote Sensing Image Semantic Segmentation

Low-order features based on convolution kernel are easy to be distorted when encountering dramatic view angle transformation and atmospheric scattering in remote sensing images. To address this concern, this paper firstly proposes to operate semantic segmentation of remote sensing images based on the high-order information, which can represent the relative relationship of low-order features and is robust and stable when suffering feature distortion. Besides, semantic decouples have recently been well researched and have achieved significant improvement in image understanding. Thus, in this paper, a High-order Semantic Decoupling Network (HSDN) is proposed to disentangle features by semantics based on high-order features. Specifically, HSDN firstly represents each pixel by calculating the pixel-level affinity as a high-order feature and then clusters these pixels into different semantics. Afterward, an attention-like mask generation module is designed for both intra and inter semantic groups, leading to three kinds of masks, including the Semantic Decoupling Mask (SDM), which utilizes each high-order cluster centroid as a mask to compact features intra cluster and expand different inter clusters, so as to improve semantic disentangle performance to a better extent; Semantic Enhancement Mask (SEM), which records pixel-level relative correlation within a class to sufficiently exploit high-order features and could enhance feature robustness; and Boundary Supplementary Mask (BSM) which aims to process borderline pixels to reduce cluster errors. Finally, by applying masks on pixels both within classes and on borderlines, semantic decoupled features are generated and concatenated to realize segmentation. The quantitative and qualitative experiments are conducted on two large-scale fine-resolution remote sensing image datasets to demonstrate the significant performance of adopting high-order representation. Besides, we also implement numerous experiments to validate the effectiveness of the proposed semantic decouple framework in dealing with complicated and distortion-prone remote sensing image segmentation tasks.

A hybrid level-set / embedded boundary method applied to solidification-melt problems

In this paper, we introduce a novel way to represent the interface for two-phase flows with phase change. We combine a level-set method with a Cartesian embedded boundary method and take advantage of both. This is part of an effort to obtain a numerical strategy relying on Cartesian grids allowing the simulation of complex boundaries with possible change of topology while retaining a high-order representation of the gradients on the interface and the capability of properly applying boundary conditions on the interface. This leads to a two-fluid conservative second-order numerical method. The ability of the method to correctly solve Stefan problems, onset dendrite growth with and without anisotropy is demonstrated through a variety of test cases. Finally, we take advantage of the two-fluid representation to model a Rayleigh–Bénard instability with a melting boundary.

A HOROR Theory for Introspective Consciousness

Higher-order theories of consciousness typically account for introspection in terms of one's higher-order thoughts being conscious, which would require a third-order thought — i.e.a thought about a thought about a mental state. In this work, we offer an alternative account of introspection that builds on the recent HigherOrder Representation of a Representation (HOROR) theory of phenomenal consciousness. According to HOROR theory, phenomenal consciousness consists in having the right kind of higher-order representation. We claim that this theory can be extended to introspection by recognizing that there is a distinctive kind of consciousness — i.e. introspective consciousness — which can be accounted for as the theory does for phenomenal consciousness generally. We call this novel view: Higher-Order Representation Intentionally For Introspective Consciousness (HORIFIC). We argue that there are independent reasons for thinking that introspective consciousness can be either 'stimuli-induced' or 'self-triggered' and that one of the benefits of the view we develop is that it can embrace a pluralist approach. Our view also accounts for what specific mental state is represented by a particular higher-order representation, and for the way in which we are aware of changes, transitions, and boundaries between mental states in specific cases of introspective consciousness.

Multimodal Triplet Attention Network for Brain Disease Diagnosis.

Multi-modal imaging data fusion has attracted much attention in medical data analysis because it can provide complementary information for more accurate analysis. Integrating functional and structural multi-modal imaging data has been increasingly used in the diagnosis of brain diseases, such as epilepsy. Most of the existing methods focus on the feature space fusion of different modalities but ignore the valuable high-order relationships among samples and the discriminative fused features for classification. In this paper, we propose a novel framework by fusing data from two modalities of functional MRI (fMRI) and diffusion tensor imaging (DTI) for epilepsy diagnosis, which effectively captures the complementary information and discriminative features from different modalities by high-order feature extraction with the attention mechanism. Specifically, we propose a triple network to explore the discriminative information from the high-order representation feature space learned from multi-modal data. Meanwhile, self-attention is introduced to adaptively estimate the degree of importance between brain regions, and the cross-attention mechanism is utilized to extract complementary information from fMRI and DTI. Finally, we use the triple loss function to adjust the distance between samples in the common representation space. We evaluate the proposed method on the epilepsy dataset collected from Jinling Hospital, and the experiment results demonstrate that our method is significantly superior to several state-of-the-art diagnosis approaches.

A discontinuous Galerkin method for sequences of earthquakes and aseismic slip on multiple faults using unstructured curvilinear grids

SUMMARYPhysics-based simulations provide a path to overcome the lack of observational data hampering a holistic understanding of earthquake faulting and crustal deformation across the vastly varying space–time scales governing the seismic cycle. However, simulations of sequences of earthquakes and aseismic slip (SEAS) including the complex geometries and heterogeneities of the subsurface are challenging. We present a symmetric interior penalty discontinuous Galerkin (SIPG) method to perform SEAS simulations accounting for the aforementioned challenges. Due to the discontinuous nature of the approximation, the spatial discretization natively provides a means to impose boundary and interface conditions. The method accommodates 2-D and 3-D domains, is of arbitrary order, handles subelement variations in material properties and supports isoparametric elements, that is, high-order representations of the exterior boundaries, interior material interfaces and embedded faults. We provide an open-source reference implementation, Tandem, that utilizes highly efficient kernels for evaluating the SIPG linear and bilinear forms, is inherently parallel and well suited to perform high-resolution simulations on large-scale distributed memory architectures. Additional flexibility and efficiency is provided by optionally defining the displacement evaluation via a discrete Green’s function approach, exploiting advantages of both the boundary integral and volumetric methods. The optional discrete Green’s functions are evaluated once in a pre-computation stage using algorithmically optimal and scalable sparse parallel solvers and pre-conditioners. We illustrate the characteristics of the SIPG formulation via an extensive suite of verification problems (analytic, manufactured and code comparison) for elastostatic and quasi-dynamic problems. Our verification suite demonstrates that high-order convergence of the discrete solution can be achieved in space and time and highlights the benefits of using a high-order representation of the displacement, material properties and geometries. We apply Tandem to realistic demonstration models consisting of a 2-D SEAS multifault scenario on a shallowly dipping normal fault with four curved splay faults, and a 3-D intersecting multifault scenario of elastostatic instantaneous displacement of the 2019 Ridgecrest, CA, earthquake sequence. We exploit the curvilinear geometry representation in both application examples and elucidate the importance of accurate stress (or displacement gradient) representation on-fault. This study entails several methodological novelties. We derive a sharp bound on the smallest value of the SIPG penalty ensuring stability for isotropic, elastic materials; define a new flux to incorporate embedded faults in a standard SIPG scheme; employ a hybrid multilevel pre-conditioner for the discrete elasticity problem; and demonstrate that curvilinear elements are specifically beneficial for volumetric SEAS simulations. We show that our method can be applied for solving interesting geophysical problems using massively parallel computing. Finally, this is the first time a discontinuous Galerkin method is published for the numerical simulations of SEAS, opening new avenues to pursue extreme scale 3-D SEAS simulations in the future.

Higher-order representation of Karamata theorem

As an important result in extreme value theory, Karamata theorem provides the integral properties of regularly varying functions. In this paper, the third-order version of Karamata theorem is derived, which is generalization for the known Karamata theorem. Furthermore, analytic expressions for the second and third-order regularly varying function are established.

Causality-based CTR prediction using graph neural networks

As a prevalent problem in online advertising, CTR prediction has attracted plentiful attention from both academia and industry. Recent studies have been reported to establish CTR prediction models in the graph neural networks (GNNs) framework. However, most of GNNs-based models handle feature interactions in a complete graph, while ignoring causal relationships among features, which results in a huge drop in the performance on out-of-distribution data. This paper is dedicated to developing a causality-based CTR prediction model in the GNNs framework (Causal-GNN) integrating representations of feature graph, user graph and ad graph in the context of online advertising. In our model, a structured representation learning method (GraphFwFM) is designed to capture high-order representations on feature graph based on causal discovery among field features in gated graph neural networks (GGNNs), and GraphSAGE is employed to obtain graph representations of users and ads. Experiments conducted on three public datasets demonstrate the superiority of Causal-GNN in AUC and Logloss and the effectiveness of GraphFwFM in capturing high-order representations on causal feature graph.

High-Order Interaction Learning for Image Captioning

Image captioning aims at understanding various semantic concepts (e.g., objects and relationships) from an image and integrating them in a sentence-level description. Hence, it is necessary to learn the interaction among these concepts. If we define the context of the interaction to be involved in the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">subject-predicate-object</i> triplet, most current methods only focus on the single triplet for the first-order interaction to generate sentences. Intuitively, we humans are able to perceive the high-order interaction among concepts from two or more triplets to describe an image. For example, when we see the triplets <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">man-cutting-sandwich</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">man-with-knife</i> , it is natural to integrate and predict the sentence <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">man cutting sandwich with knife</i> . This depends on the high-order interaction between <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">cutting</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">knife</i> in different triplets. Therefore, exploiting high-order interaction is expected to benefit image captioning and focus on reasoning. In this paper, we introduce the novel high-order interaction learning method over detected objects and relationships for image captioning under the umbrella of the encoder-decoder framework. We first extract a set of object and relationship features in an image. During the encoding stage, the interactive refining network is proposed to learn high-order representations by modeling intra- and inter-object feature interaction in the self-attention fashion. During the decoding stage, the interactive fusion network is proposed to integrate object and relationship information by strengthening their high-order interaction based on language context for sentence generation. In this way, we learn the object-relationship dependencies in different stages, which can provide abundant cues for both visual understanding and caption generation. Extensive experiments show that the proposed method can achieve competitive performances against the state-of-the-art methods on MSCOCO dataset. Additional ablation studies further validate its effectiveness.

Mastering the Explicit Opinion-Role Interaction: Syntax-Aided Neural Transition System for Unified Opinion Role Labeling

Unified opinion role labeling (ORL) aims to detect all possible opinion structures of 'opinion-holder-target' in one shot, given a text. The existing transition-based unified method, unfortunately, is subject to longer opinion terms and fails to solve the term overlap issue. Current top performance has been achieved by employing the span-based graph model, which however still suffers from both high model complexity and insufficient interaction among opinions and roles. In this work, we investigate a novel solution by revisiting the transition architecture, and augmenting it with a pointer network (PointNet). The framework parses out all opinion structures in linear-time complexity, meanwhile breaks through the limitation of any length of terms with PointNet. To achieve the explicit opinion-role interactions, we further propose a unified dependency-opinion graph (UDOG), co-modeling the syntactic dependency structure and the partial opinion-role structure. We then devise a relation-centered graph aggregator (RCGA) to encode the multi-relational UDOG, where the resulting high-order representations are used to promote the predictions in the vanilla transition system. Our model achieves new state-of-the-art results on the MPQA benchmark. Analyses further demonstrate the superiority of our methods on both efficacy and efficiency.

PB1800: DEEPDECON: A DEEP-LEARNING METHOD FOR DETECTING MINIMAL RESIDUAL DISEASE OF ACUTE MYELOID LEUKEMIA

Background: Acute myeloid leukemia (AML) is a heterogeneous disease that the myeloid lineages of hematopoietic progenitor cells (blasts) lose the ability of normal differentiation and build up in the bone marrow. The ability to identify minimal residual disease (MRD) below the morphology-based 5% blast threshold has significant impact in clinical outcome. Flow cytometry and real-time quantitative polymerase chain reaction [qPCR]) are the most common methods for MRD detection while next-generation sequencing (NGS) is emerging as a more sensitive method. Aims: Here, we report DeepDecon, a deep neural network model that utilizes NGS RNAseq data to detect MRD of AML. Methods: DeepDecon makes use of the non-linear transformation among the latent layers and learns to extract high-order representations of the input gene expression. Meanwhile, DeepDecon doesn’t need to select a subset of genes. The neural network will optimize and assign different weights to all genes to minimize the prediction error. A challenging issue with deep learning models is the need for large data sets of diverse examples for model training to obtain robust and good performance. With single-cell RNAseq data, we digitally create a large number of bulk RNAseq samples with known AML fraction for model training. Results: When applied to clinical data TARGET-AML from GDC Data Portal (https://portal.gdc. cancer.gov/), DeepDecon demonstrated better performance than other machine learning algorithms, including MuSic, NNLS, Bisque, RNA-Sieve, and Scaden, for detecting MRD of AML. Summary/Conclusion: Deepdecon can be applied to clinical NGS data for accurate and sensitive detection of MRD in AML treatment. It will significantly improve the clinical outcomes of current AML therapy.

How distinct functional insular subdivisions mediate interacting neurocognitive systems.

Recent neurocognitive models propose that the insula serves as a hub of interoceptive awareness system, modulating 2 interplaying neurocognitive systems: The posterior insula (PI) receives and integrates various interoceptive signals; these signals are then transmitted to the anterior insula for processing higher-order representations into awareness, where the dorsal anterior insula (dAI) modulates the prefrontal self-control system and the ventral anterior insula (vAI) modulates the amygdala (AMG)-striatal reward-seeking circuit. We sought to test this view using a multimodal approach. We first used a resting-state functional magnetic resonance imaging (fMRI) approach with a sample of 120 undergraduate students. Then, we unpacked the neuro-cognitive association between insular connectivity and cognitive performance during an Iowa gambling fMRI task. Lastly, an independent Open Southwest University Longitudinal Imaging Multimodal dataset was used to validate the results. Findings suggested that the dAI was predominantly connected to the prefrontal regions; the vAI was primarily connected to the AMG-ventral-striatum system; and the PI was mainly connected to the visceral-sensorimotor system. Moreover, cognitive scores were positively correlated with FC between dAI and the self-control process of ventrolateral prefrontal cortex and were negatively correlated with FC between vAI and the reward-seeking process of orbitofrontal cortex and subgenual anterior cingulate cortex. The findings highlight the roles of our theorized subinsular functionality in the overall operation of the neural cognitive systems.

Delamination prediction in composite panels using unsupervised-feature learning methods with wavelet-enhanced guided wave representations

With the introduction of damage tolerance-based design philosophies, the demand for reliable and robust structural health monitoring (SHM) procedures for aerospace composite structures is increasing rapidly. The performance of supervised learning algorithms for SHM depends on the amount of labeled and balanced datasets. Apart from this, collecting datasets accommodating all possible damage scenarios is cumbersome, costly, and inaccessible for aerospace applications. In this paper, we have proposed two different unsupervised-feature learning approaches where the algorithms are trained only on the baseline scenarios to learn the distribution of baseline signals. The trained unsupervised feature learner is used for delamination prediction with an anomaly detection philosophy. In the first approach, we have combined dimensionality reduction techniques (principal component analysis and independent component analysis) with a one-class support vector machine. In another approach, we have utilized deep learning-based deep convolutional autoencoders (CAE). These state-of-the-art algorithms are applied on three different guided wave-based experimental datasets. The raw guided wave signals present in the datasets are converted into wavelet-enhanced higher-order representations for training unsupervised feature-learning algorithms. We have also compared different techniques, and it is seen that CAE generates better reconstructions with lower mean squared error and can provide higher accuracy on all the datasets.

Brain-wide visual habituation networks in wild type and fmr1 zebrafish

Habituation is a form of learning during which animals stop responding to repetitive stimuli, and deficits in habituation are characteristic of several psychiatric disorders. Due to technical challenges, the brain-wide networks mediating habituation are poorly understood. Here we report brain-wide calcium imaging during larval zebrafish habituation to repeated visual looming stimuli. We show that different functional categories of loom-sensitive neurons are located in characteristic locations throughout the brain, and that both the functional properties of their networks and the resulting behavior can be modulated by stimulus saliency and timing. Using graph theory, we identify a visual circuit that habituates minimally, a moderately habituating midbrain population proposed to mediate the sensorimotor transformation, and downstream circuit elements responsible for higher order representations and the delivery of behavior. Zebrafish larvae carrying a mutation in the fmr1 gene have a systematic shift toward sustained premotor activity in this network, and show slower behavioral habituation.

Distinct higher-order representations of natural sounds in human and ferret auditory cortex.

Little is known about how neural representations of natural sounds differ across species. For example, speech and music play a unique role in human hearing, yet it is unclear how auditory representations of speech and music differ between humans and other animals. Using functional ultrasound imaging, we measured responses in ferrets to a set of natural and spectrotemporally matched synthetic sounds previously tested in humans. Ferrets showed similar lower-level frequency and modulation tuning to that observed in humans. But while humans showed substantially larger responses to natural vs. synthetic speech and music in non-primary regions, ferret responses to natural and synthetic sounds were closely matched throughout primary and non-primary auditory cortex, even when tested with ferret vocalizations. This finding reveals that auditory representations in humans and ferrets diverge sharply at late stages of cortical processing, potentially driven by higher-order processing demands in speech and music.

MBRep: Motif-based representation learning in heterogeneous networks

In recent years, there has been a surge of interest in applying machine learning to graphs and networks that already exist in the world around us. The approach has been successfully used for domains as diverse as traffic management, e-commerce recommendation and public opinion monitoring. A critical aspect of representation learning for applied machine learning is feature engineering. Deep learning-based research in representation learning has developed methods for automatically learning a large number of potentially correlated features from original networks. However, most of these methods cannot be applied to heterogeneous networks, which are true expressions of the real-world. This is because they do not adequately capture the structure and semantics of different types of nodes in heterogeneous networks and the links between them. They also struggle to represent higher-order heterogeneous patterns of connection. This paper proposes a generalized motif-based higher-order representation learning method, MBRep, that learns triangle motif embedding in a network, on the basis of which it can obtain the embedding and representation of nodes in a heterogeneous network. Statistically, significant motif structures are extracted from the original heterogeneous network and combined with the heterogeneity of the nodes. A weight-biased random walk is then applied to the motif level higher-order network, using a SkipGram model to embed the motifs. After this, the embedding of the original network nodes is calculated using weighted averages and feature alignment. This can then be used for link prediction. We confirmed the effectiveness of MBRep by comparing its AUC and MRR performance with other state-of-the-art methods on three real-world datasets. Its adaptability was also validated by conducting a cold-start test.

Partially Preserved Processing of Musical Rhythms in REM but Not in NREM Sleep.

The extent of high-level perceptual processing during sleep remains controversial. In wakefulness, perception of periodicities supports the emergence of high-order representations such as the pulse-like meter perceived while listening to music. Electroencephalography (EEG) frequency-tagged responses elicited at envelope frequencies of musical rhythms have been shown to provide a neural representation of rhythm processing. Specifically, responses at frequencies corresponding to the perceived meter are enhanced over responses at meter-unrelated frequencies. This selective enhancement must rely on higher-level perceptual processes, as it occurs even in irregular (i.e., syncopated) rhythms where meter frequencies are not prominent input features, thus ruling out acoustic confounds. We recorded EEG while presenting a regular (unsyncopated) and an irregular (syncopated) rhythm across sleep stages and wakefulness. Our results show that frequency-tagged responses at meter-related frequencies of the rhythms were selectively enhanced during wakefulness but attenuated across sleep states. Most importantly, this selective attenuation occurred even in response to the irregular rhythm, where meter-related frequencies were not prominent in the stimulus, thus suggesting that neural processes selectively enhancing meter-related frequencies during wakefulness are weakened during rapid eye movement (REM) and further suppressed in non-rapid eye movement (NREM) sleep. These results indicate preserved processing of low-level acoustic properties but limited higher-order processing of auditory rhythms during sleep.