Similar Representations Research Articles

Committor functions via tensor networks

We propose a novel approach for computing committor functions, which describe transitions of a stochastic process between metastable states. The committor function satisfies a backward Kolmogorov equation, and in typical high-dimensional settings of interest, it is intractable to compute and store the solution with traditional numerical methods. By parametrizing the committor function in a matrix product state/tensor train format and using a similar representation for the equilibrium probability density, we solve the variational formulation of the backward Kolmogorov equation with linear time and memory complexity in the number of dimensions. This approach bypasses the need for sampling the equilibrium distribution, which can be difficult when the distribution has multiple modes. Numerical results demonstrate the effectiveness of the proposed method for high-dimensional problems.

S507 Esophageal Function Assessment Using Endoscopic Functional Luminal Imaging Probe (ENDOFLIP) in Patients With Different Obesity Classes

Introduction: Patients with obesity have been found to have a higher prevalence of esophageal motor disorders compared to the general population, with abnormal lower esophageal sphincter (LES) function and altered peristalsis on manometry. Patients with class 3 obesity can experience minimal symptoms despite abnormal manometry suggesting abnormal visceral sensation. Endoscopic functional luminal imaging probe (EndoFLIP) assesses luminal distensibility and peristalsis during endoscopy. This study aims to evaluate esophageal characteristics using EndoFLIP in patients with obesity and esophageal symptoms and determine clinical differences between obesity classes. Methods: We conducted retrospective chart review of patients with obesity and esophageal symptoms who had undergone EndoFLIP testing at a single tertiary care center between 2020-2022. Data was collected on patient demographics, symptoms, medical history, and EndoFLIP measurements. Statistical analysis was performed using Fisher exact testing. Results: 33 patients with a BMI ≥30 were included (81% women, mean age 50.2 years, mean BMI 39.4). 9 (27%) had class 3 obesity, 9 (27%) class 2 obesity and 15 (45%) class 1 obesity. Most common symptoms were dysphagia (72%), heartburn (48%), reflux (48%), and regurgitation (42%). On EndoFLIP, 8 (24%) had reduced esophagogastric junction distensibility index (EGJ-DI), of which 2 (13%) were classified as obesity class 1, 4 (44%) obesity class 2 and 2 (22%) obesity class 3 (Table). 17 (52%) had normal repetitive antegrade contractions (RACs), 7 (21%) diminished/disordered contractile response (DDCRs) and 9 (27%) absent contractility. Abnormal contractility patterns were seen in 7 (47%) obesity class 1, 4 (44%) obesity class 2 and 5 (56%) obesity class 3. There were no significant differences between obesity classes with regards to medication use, co-morbidities, symptoms, endoscopy and EndoFLIP findings. Conclusion: In this cohort, symptomatic patients in different obesity classes did not have significantly different LES function or peristalsis on EndoFLIP. EGJ distensibility was reduced in a quarter of the study population and seen more frequently in patients with class 2 obesity. Nearly half of the patients had abnormal peristalsis with DDCRs and absent contractility, with similar representation between obesity classes. Prospective controlled studies with EndoFLIP comparing symptomatic and asymptomatic patients in different obesity classes are needed to validate these findings. Table 1. - Comparison of patient characteristics and EndoFLIP findings by obesity class Class 1 Obesity BMI 30-35 (N=15) Class 2 Obesity BMI 35-40 (N=9) Class 3 Obesity BMI >40 (N=9) p-value MEDICATIONS N (%) PPI 8 (53) 6(67) 5 (56) 0.67 H2 blocker 5 (33) 1 (11) 2 (22) Opiods 2 (13) 0 2 (22) Benzodiazepines 0 1 (11) 0 SYMPTOMS N (%) Reflux 6 (40) 6 (67) 4 (44) 0.57 Dysphagia 12 (80) 7 (78) 5 (56) Heartburn 7 (47) 4 (44) 5 (56) Regurgitation 3 (20) 4 (44) 7 (78) Belching/Nausea 2 (13) 1 (11) 1 (11) Epigastric abdominal pain 4 (27) 2 (22) 2 (22) Vomiting 1 (7) 2 (22) 1 (11) EGD FINDINGS N (%) Esophagitis 4 (27) 2 (22) 3 (33) 0.78 Gastritis 8 (53) 4 (44) 5 (56) Fluid in esophagus 1 (7) 2 (22) 3 (33) ENDOFLIP FINDINGS EGJ-distensibility index N (%) [mean] Normal 13 (87) [5.4] 5 (56) [4.65] 7 (78) [4.98] 0.26 Reduced 2 (13) [2.27] 4 (44) [0.91] 2 (22) [2.82] Contractility pattern N (%) RACs 8 (53) 5 (56) 4 (44) 1 DDCRs 3 (20) 2 (22) 2 (22) Absent contractility 4 (27) 2 (22) 3 (33) BMI = Body Mass Index; PPI = Proton pump inhibitor; H2 blocker = histamine 2 blocker (eg. famotidine); EGD = Esophagogastroduodenoscopy; EGJ = Esophagogastric Junction; RACs = Repetitive Antegrade Contractions; DDCRs = Diminished or Disordered Contractile Responses

Improving Automatic Speech Recognition Performance for Low-Resource Languages With Self-Supervised Models

Speech self-supervised learning has attracted much attention due to its promising performance in multiple downstream tasks, and has become a new growth engine for speech recognition in low-resource languages. In this paper, we exploit and analyze a series of wav2vec pre-trained models for speech recognition in 15 low-resource languages in the OpenASR21 Challenge. The investigation covers two important variables during pre-training, three fine-tuning methods, as well as applications in End-to-End and hybrid systems. First, pre-trained models with different pre-training audio data and architectures (wav2vec2.0, HuBERT and WavLM) are explored for their speech recognition performance in low-resource languages. Second, we investigate data utilization, multilingual learning, and the use of a phoneme-level recognition task in fine-tuning. Furthermore, we explore what effect fine-tuning has on the similarity of representations extracted from different transformer layers. The similarity analyses cover different pre-trained architectures and fine-tuning languages. We apply pre-trained representations to End-to-End and hybrid systems to confirm our representation analyses, which have obtained better performances as well.

Noise Robust Face Hallucination Based on Smooth Correntropy Representation.

Face hallucination technologies have been widely developed during the past decades, among which the sparse manifold learning (SML)-based approaches have become the popular ones and achieved promising performance. However, these SML methods always failed in handling noisy images due to the least-square regression (LSR) they used for error approximation. To this end, we propose, in this article, a smooth correntropy representation (SCR) model for noisy face hallucination. In SCR, the correntropy regularization and smooth constraint are combined into one unified framework to improve the resolution of noisy face images. Specifically, we introduce the correntropy induced metric (CIM) rather than the LSR to regularize the encoding errors, which admits the proposed method robust to noise with uncertain distributions. Besides, the fused LASSO penalty is added into the feature space to ensure similar training samples holding similar representation coefficients. This encourages the SCR not only robust to noise but also can well exploit the inherent typological structure of patch manifold, resulting in more accurate representations in noise environment. Comparison experiments against several state-of-the-art methods demonstrate the superiority of SCR in super-resolving noisy low-resolution (LR) face images.

Complementary task representations in hippocampus and prefrontal cortex for generalizing the structure of problems

Humans and other animals effortlessly generalize prior knowledge to solve novel problems, by abstracting common structure and mapping it onto new sensorimotor specifics. To investigate how the brain achieves this, in this study, we trained mice on a series of reversal learning problems that shared the same structure but had different physical implementations. Performance improved across problems, indicating transfer of knowledge. Neurons in medial prefrontal cortex (mPFC) maintained similar representations across problems despite their different sensorimotor correlates, whereas hippocampal (dCA1) representations were more strongly influenced by the specifics of each problem. This was true for both representations of the events that comprised each trial and those that integrated choices and outcomes over multiple trials to guide an animal’s decisions. These data suggest that prefrontal cortex and hippocampus play complementary roles in generalization of knowledge: PFC abstracts the common structure among related problems, and hippocampus maps this structure onto the specifics of the current situation.

Microlensing model inference with normalising flows and reversible jump MCMC

The estimation of parameters for gravitational microlensing events typically proceeds by fitting proposed lens models to observed data. A problem arises in this approach when alternate parameter values or models produce similar representations, since some comparison must be made between them. This is a challenging problem which has been frequently tackled via Bayesian model selection. Previously this was impractical since the gravitational microlensing joint model posterior is multi-modal and the computational expense required to sample an accurate discretisation prohibitive. Recently Zhang et al. (2021) showed how a neural network can be used to quickly approximate the model dependent posterior distribution of gravitational microlensing data. We show how in combination with Bayesian statistics this can be used to estimate marginal model probabilities from gravitational microlensing data. By using a similar neural network construction to generate approximate posteriors for each candidate lens model, we detail a reversible jump Metropolis–Hastings sampling algorithm which refines these posteriors to the joint-model posterior within a computationally feasible time frame. This is applied to simulated ROMAN events. We intend for this paper to serve as a proof-of-concept for the further development of Bayesian statistics in application to gravitational microlensing model inference.

DGRUnit: Dual graph reasoning unit for brain tumor segmentation

Many fully automatic segmentation models have been created to solve the difficulty of brain tumor segmentation, thanks to the rapid growth of deep learning. However, few approaches focus on the long-range relationships and contextual interdependence in multimodal Magnetic Resonance (MR) images. In this paper, we propose a novel approach for brain tumor segmentation called the dual graph reasoning unit (DGRUnit). Two parallel graph reasoning modules are included in our proposed method: a spatial reasoning module and a channel reasoning module. The spatial reasoning module models the long-range spatial dependencies between distinct regions in an image using a graph convolutional network (GCN). The channel reasoning module uses a graph attention network (GAT) to model the rich contextual interdependencies between different channels with similar semantic representations. Our experimental results clearly demonstrate the superior performance of the proposed DGRUnit. The ablation study shows the flexibility and generalizability of our model, which can be easily integrated into a wide range of neural networks and further improve them. When compared to several state-of-the-art methods, experimental results show that the proposed approach significantly improves both visual inspection and quantitative metrics for brain tumor segmentation tasks.

Evaluation of a Radiation Oncology Microclerkship as a Component of Medical Student Training

Background Many US medical students gain clinical exposure to oncology specialties such as medical and/or surgical oncology throughout their core clerkships. Radiation oncology (RO), however, lacks similar representation during the standard curriculum, and few students pursue a dedicated RO elective. This substantially limits the broader understanding of the role of radiation therapy (RT) in multidisciplinary cancer care, including the role of palliative RT and survivorship care. The hypothesis of this study is that incorporating a 1-day "microclerkship" in RO as part of related clerkships will be feasible and perceived as valuable to participating students. Methods The RO clerkship director at a single institution partnered with clerkship directors in medical oncology, palliative care, and radiology so that every 3rd or 4th year medical student rotating in those specialties would spend 1 day in RO during each clerkship. Students were given educational materials to review beforehand, including a 30-minute introductory recorded presentation about general RO principles, and a slide deck and article describing how radiation therapy is relevant to patient care in primary care and other medical subspecialties. At minimum, the day in RO included shadowing one radiation oncologist (and any resident working with them) in clinic and observation of patient treatments, including attending tumor board, brachytherapy procedures, or treatment planning sessions if feasible. Students were invited to complete an anonymous electronic survey immediately after their experience, containing multiple choice and Likert-type questions (1=not at all valuable, 5=extremely valuable). Results Pilot data from 15 students are reported. No student had visited a RO department previously, and only 1 had attended a prior lecture from a radiation oncologist. Students reported the experience to be valuable (median 4, [4-4.5 interquartile range (IQR)]). 13 students (87%) rated the experience quite or extremely valuable. 14 students (93%) felt moderately, very, or extremely more knowledgeable about the role of radiation therapy in cancer management. After the microclerkship experience, some students were more interested in a full 2-4 week RO rotation (8 of 13, 62%) and RO career (4 of 13, 31%). Discussion Preliminary data suggest a valuable role of incorporating RO exposure into related medical student clerkships. Future plans include increasing student participation from these and other related clerkships and expanding the program to other institutions. We hypothesize that this approach will improve US medical student exposure to RO, which could make a subsequent positive impact on referral patterns and multidisciplinary patient care. Many US medical students gain clinical exposure to oncology specialties such as medical and/or surgical oncology throughout their core clerkships. Radiation oncology (RO), however, lacks similar representation during the standard curriculum, and few students pursue a dedicated RO elective. This substantially limits the broader understanding of the role of radiation therapy (RT) in multidisciplinary cancer care, including the role of palliative RT and survivorship care. The hypothesis of this study is that incorporating a 1-day "microclerkship" in RO as part of related clerkships will be feasible and perceived as valuable to participating students. The RO clerkship director at a single institution partnered with clerkship directors in medical oncology, palliative care, and radiology so that every 3rd or 4th year medical student rotating in those specialties would spend 1 day in RO during each clerkship. Students were given educational materials to review beforehand, including a 30-minute introductory recorded presentation about general RO principles, and a slide deck and article describing how radiation therapy is relevant to patient care in primary care and other medical subspecialties. At minimum, the day in RO included shadowing one radiation oncologist (and any resident working with them) in clinic and observation of patient treatments, including attending tumor board, brachytherapy procedures, or treatment planning sessions if feasible. Students were invited to complete an anonymous electronic survey immediately after their experience, containing multiple choice and Likert-type questions (1=not at all valuable, 5=extremely valuable). Pilot data from 15 students are reported. No student had visited a RO department previously, and only 1 had attended a prior lecture from a radiation oncologist. Students reported the experience to be valuable (median 4, [4-4.5 interquartile range (IQR)]). 13 students (87%) rated the experience quite or extremely valuable. 14 students (93%) felt moderately, very, or extremely more knowledgeable about the role of radiation therapy in cancer management. After the microclerkship experience, some students were more interested in a full 2-4 week RO rotation (8 of 13, 62%) and RO career (4 of 13, 31%). Preliminary data suggest a valuable role of incorporating RO exposure into related medical student clerkships. Future plans include increasing student participation from these and other related clerkships and expanding the program to other institutions. We hypothesize that this approach will improve US medical student exposure to RO, which could make a subsequent positive impact on referral patterns and multidisciplinary patient care.

Towards the Web of Embeddings: Integrating multiple knowledge graph embedding spaces with FedCoder

The Semantic Web is distributed yet interoperable: Distributed since resources are created and published by a variety of producers, tailored to their specific needs and knowledge; Interoperable as entities are linked across resources, allowing to use resources from different providers in concord. Complementary to the explicit usage of Semantic Web resources, embedding methods made them applicable to machine learning tasks. Subsequently, embedding models for numerous tasks and structures have been developed, and embedding spaces for various resources have been published. The ecosystem of embedding spaces is distributed but not interoperable: Entity embeddings are not readily comparable across different spaces. To parallel the Web of Data with a Web of Embeddings, we must thus integrate available embedding spaces into a uniform space.Current integration approaches are limited to two spaces and presume that both of them were embedded with the same method — both assumptions are unlikely to hold in the context of a Web of Embeddings. In this paper, we present FedCoder— an approach that integrates multiple embedding spaces via a latent space. We assert that linked entities have a similar representation in the latent space so that entities become comparable across embedding spaces. FedCoder employs an autoencoder to learn this latent space from linked as well as non-linked entities.Our experiments show that FedCoder substantially outperforms state-of-the-art approaches when faced with different embedding models, that it scales better than previous methods in the number of embedding spaces, and that it improves with more graphs being integrated whilst performing comparably with current approaches that assumed joint learning of the embeddings and were, usually, limited to two sources. Our results demonstrate that FedCoder is well adapted to integrate the distributed, diverse, and large ecosystem of embeddings spaces into an interoperable Web of Embeddings.

Preprint Highlight: Self-supervised deep learning encodes high-resolution features of protein subcellular localization.

The ability to quantitatively characterize protein subcellular spatial organization is key for understanding how cells orchestrate their many functions. This study presents cytoself, a deep‐learning self‐supervised method to compute lower dimensional representations of protein spatial localization distributions by forcing different images of the same protein to have similar representations. Cytoself was applied to the public resource OpenCell to build a protein localization atlas that was validated by identifying known protein complexes that cluster together. The authors demonstrate how to interpret cytoself's output by identifying specific features in the localization representation associated with specific localization patterns in the images and by reverse engineering the neural network. Cytoself could serve as a “discovery tool” to generate hypotheses regarding unknown protein functions and protein‐protein interactions.

Familiarity is familiarity is familiarity: Event-related brain potentials reveal qualitatively similar representations of personally familiar and famous faces.

Humans excel in familiar face recognition, but often find it hard to make identity judgements of unfamiliar faces. Understanding of the factors underlying the substantial benefits of familiarity is at present limited, but the effect is sometimes qualified by the way in which a face is known-for example, personal acquaintance sometimes gives rise to stronger familiarity effects than exposure through the media. Given the different quality of personal versus media knowledge, for example in one's emotional response or level of interaction, some have suggested qualitative differences between representations of people known personally or from media exposure. Alternatively, observed differences could reflect quantitative differences in the level of familiarity. We present 4 experiments investigating potential contributory influences to face familiarity effects in which observers view pictures showing their friends, favorite celebrities, celebrities they dislike, celebrities about whom they have expressed no opinion, and their own face. Using event-related potential indices with high temporal resolution and multiple highly varied everyday ambient images as a strong test of face recognition, we focus on the N250 and the later Sustained Familiarity Effect (SFE). All known faces show qualitatively similar responses relative to unfamiliar faces. Regardless of personal- or media-based familiarity, N250 reflects robust visual representations, successively refined over increasing exposure, while SFE appears to reflect the amount of identity-specific semantic information known about a person. These modulations of visual and semantic representations are consistent with face recognition models which emphasize the degree of familiarity but do not distinguish between different types of familiarity. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Identifying important conceptual areas in a nursing education programme to meet future demands, using group concept mapping

A nursing programme in southern Sweden was revised to meet future demands. The aim of this study was to explore important conceptual areas to be included in a nursing programme in order to meet long-term societal and health care requirements. Group concept mapping (GCM), a mixed-methods approach, was used. Thirty-four experienced teachers participated. Data was collected during brainstorming sessions in focus groups. Following editing and removal of duplicates, 101 statements remained to be sorted into piles that had similar conceptual representation in nursing education. The final step was then to rate each statement from 1 to 5 (higher values = more important/more feasible). Quantitative analysis using non-metric multidimensional scaling and hierarchical cluster analysis resulted in 11 clusters. Assessment and treatment and Person-centred care were rated as the most important and Scientific theories and methods, Basic caregiving in nursing and Person-centred care had the highest feasibility ratings. Further analyses suggested that the content of nursing education can be seen from a systems theory perspective, represented by the macro, meso, and micro levels. These levels may increase the understanding of the complexity of nursing care. Furthermore, the cluster analysis can facilitate the development of a concept-based curriculum for nursing education.

Predicting memory from the network structure of naturalistic events

When we remember events, we often do not only recall individual events, but also the connections between them. However, extant research has focused on how humans segment and remember discrete events from continuous input, with far less attention given to how the structure of connections between events impacts memory. Here we conduct a functional magnetic resonance imaging study in which participants watch and recall a series of realistic audiovisual narratives. By transforming narratives into networks of events, we demonstrate that more central events—those with stronger semantic or causal connections to other events—are better remembered. During encoding, central events evoke larger hippocampal event boundary responses associated with memory formation. During recall, high centrality is associated with stronger activation in cortical areas involved in episodic recollection, and more similar neural representations across individuals. Together, these results suggest that when humans encode and retrieve complex real-world experiences, the reliability and accessibility of memory representations is shaped by their location within a network of events.

Multi-feature sparse similar representation for person identification

Person identification with a single feature (e.g., face recognition, speaker verification, person re-identification, etc.) has been studied extensively for many years, while few works focus on multi-feature person identification. Though promising performance has been achieved by only using the information of facial images, voice, or pedestrian appearance, it is still challenging to recognize a person with only a single feature in some situations (e.g., a person at a distance or occluded by other objects, and a partial person out of view). In this paper, we present a multi-feature sparse similar representation (MFSSR) method to effectively fuse face features, body features, and global image features for the task of person identification. In MFSSR, we designed a reconstructed deep spatial feature for representing the appearance of human body by using the spatial correlation coding of partial deep spatial features. Then we presented a multi-feature sparse similar representation model for jointly using different features, e.g., face, body, and the global image. Besides, considering that the coding coefficients associated with good samples but not outliers should be more similar among different features, we jointly represent different features by imposing a weighted ℓ1-norm distance regularization, instead of the conventional ℓ2-norm regularization, on the coefficients. Experimental results on several multi-feature person identification databases have clearly shown the superior performance of the proposed model.

Mental representations of speech and musical pitch contours reveal a diversity of profiles in autism spectrum disorder.

As a key auditory attribute of sounds, pitch is ubiquitous in our everyday listening experience involving language, music and environmental sounds. Given its critical role in auditory processing related to communication, numerous studies have investigated pitch processing in autism spectrum disorder. However, the findings have been mixed, reporting either enhanced, typical or impaired performance among autistic individuals. By investigating top-down comparisons of internal mental representations of pitch contours in speech and music, this study shows for the first time that, while autistic individuals exhibit diverse profiles of pitch processing compared to non-autistic individuals, their mental representations of pitch contours are typical across domains. These findings suggest that pitch-processing mechanisms are shared across domains in autism spectrum disorder and provide theoretical implications for using music to improve speech for those autistic individuals who have language problems.

Text Mining to Alleviate the Cold-Start Problem of Adaptive Comparative Judgments

Comparative judgments permit the assessment of open-ended student works by constructing a latent quality scale through repeated pairwise comparisons (i.e., which works “win” or “lose”). Adaptive comparative judgments speed up the judgment process by maximizing the Fisher information of the next comparison. However, at the start of a judgment process, such an adaptive algorithm will not perform well. In order to reliably approximate the Fisher Information of possible pairs well, multiple comparisons are needed. In addition, adaptive comparative judgments have been shown to inflate the scale separation coefficient, which is a reliability estimator for the quality estimates. Current methods to solve the inflation issue increase the number of required comparisons. The goal of this study is to alleviate the cold-start problem of adaptive comparative judgments for essays or other textual assignments, but also to minimize the bias of the scale separation coefficient. By using text-mining techniques, which can be performed before the first judgment, essays can be adaptively compared from the start. More specifically, we propose a selection rule that is based both on a high (1) cosine similarity of the vector representations and (2) Fisher Information of essay pairs. At the start of the judgment process, the cosine similarity has the highest weight in the selection rule. With more judgments, this weight decreases progressively, whereas the weight of the Fisher Information increases. Using simulated data, the proposed strategy is compared with existing approaches. The results indicate that the proposed selection rule can mitigate both the cold-start. That is, fewer judgments are needed to obtain accurate and reliable quality estimates. In addition, the selection rule was found to reduce the inflation of the scale separation reliability.

Knowledge Distillation via Constrained Variational Inference.

Knowledge distillation has been used to capture the knowledge of a teacher model and distill it into a student model with some desirable characteristics such as being smaller, more efficient, or more generalizable. In this paper, we propose a framework for distilling the knowledge of a powerful discriminative model such as a neural network into commonly used graphical models known to be more interpretable (e.g., topic models, autoregressive Hidden Markov Models). Posterior of latent variables in these graphical models (e.g., topic proportions in topic models) is often used as feature representation for predictive tasks. However, these posterior-derived features are known to have poor predictive performance compared to the features learned via purely discriminative approaches. Our framework constrains variational inference for posterior variables in graphical models with a similarity preserving constraint. This constraint distills the knowledge of the discriminative model into the graphical model by ensuring that input pairs with (dis)similar representation in the teacher model also have (dis)similar representation in the student model. By adding this constraint to the variational inference scheme, we guide the graphical model to be a reasonable density model for the data while having predictive features which are as close as possible to those of a discriminative model. To make our framework applicable to a wide range of graphical models, we build upon the Automatic Differentiation Variational Inference (ADVI), a black-box inference framework for graphical models. We demonstrate the effectiveness of our framework on two real-world tasks of disease subtyping and disease trajectory modeling.

Learning to Transfer with von Neumann Conditional Divergence

The similarity of feature representations plays a pivotal role in the success of problems related to domain adaptation. Feature similarity includes both the invariance of marginal distributions and the closeness of conditional distributions given the desired response y (e.g., class labels). Unfortunately, traditional methods always learn such features without fully taking into consideration the information in y, which in turn may lead to a mismatch of the conditional distributions or the mixup of discriminative structures underlying data distributions. In this work, we introduce the recently proposed von Neumann conditional divergence to improve the transferability across multiple domains. We show that this new divergence is differentiable and eligible to easily quantify the functional dependence between features and y. Given multiple source tasks, we integrate this divergence to capture discriminative information in y and design novel learning objectives assuming those source tasks are observed either simultaneously or sequentially. In both scenarios, we obtain favorable performance against state-of-the-art methods in terms of smaller generalization error on new tasks and less catastrophic forgetting on source tasks (in the sequential setup).

A novel explainable neural network for Alzheimer’s disease diagnosis

Visual classification for medical images has been dominated by convolutional neural networks (CNNs) for years. Though they have shown great performance on accuracy, some of them provide decisions that are hard to explain while others encode information from irrelevant or noisy regions. In this work, we try to close this gap by proposing an explainable framework which consists of a predictor and an explainable tool, so as to provide accurate diagnoses with intuitive visualization maps and prediction basis. Specifically, the predictor is designed by applying attention mechanisms to multi-scale features so as to learn and discover class discriminative latent representations that are close to each brain volume’s label. Meanwhile, to explain our predictor, we propose the novel explainable tool which includes a high-resolution visualization method and a prediction-basis creation and retrieval module. The former effectively integrates the feature maps of intermediate layers as well as the last convolutional layer, which surpasses state-of-the-art visualization approaches in producing high-resolution representations with more accurate localization of discriminative areas. While the latter provides prediction basis evidence via retrieved volumes with similar latent representations which are accessible to neurologists. Extensive experiments show that the proposed framework achieves higher level of accuracy and explainability over other state-of-the-art solutions. More importantly, it localizes crucial brain areas with clearer boundaries, less noises, which matches background knowledge in the neuroscience literature.

Powerful Graph Convolutional Networks with Adaptive Propagation Mechanism for Homophily and Heterophily

Graph Convolutional Networks (GCNs) have been widely applied in various fields due to their significant power on processing graph-structured data. Typical GCN and its variants work under a homophily assumption (i.e., nodes with same class are prone to connect to each other), while ignoring the heterophily which exists in many real-world networks (i.e., nodes with different classes tend to form edges). Existing methods deal with heterophily by mainly aggregating higher-order neighborhoods or combing the immediate representations, which leads to noise and irrelevant information in the result. But these methods did not change the propagation mechanism which works under homophily assumption (that is a fundamental part of GCNs). This makes it difficult to distinguish the representation of nodes from different classes. To address this problem, in this paper we design a novel propagation mechanism, which can automatically change the propagation and aggregation process according to homophily or heterophily between node pairs. To adaptively learn the propagation process, we introduce two measurements of homophily degree between node pairs, which is learned based on topological and attribute information, respectively. Then we incorporate the learnable homophily degree into the graph convolution framework, which is trained in an end-to-end schema, enabling it to go beyond the assumption of homophily. More importantly, we theoretically prove that our model can constrain the similarity of representations between nodes according to their homophily degree. Experiments on seven real-world datasets demonstrate that this new approach outperforms the state-of-the-art methods under heterophily or low homophily, and gains competitive performance under homophily.