Document-level relation extraction aims to discover relations between entities across a whole document. How to build the dependency of entities from different sentences in a document remains to be a great challenge. Current approaches either leverage syntactic trees to construct document-level graphs or aggregate inference information from different sentences. In this paper, we build cross-sentence dependencies by inferring compositional relations between inter-sentence mentions. Adopting aggressive linking strategy, intermediate relations are reasoned on the document-level graphs by mention convolution. We further notice the generalization problem of NA instances, which is caused by incomplete annotation and worsened by fully-connected mention pairs. An improved ranking loss is proposed to attend this problem. Experiments show the connections between different mentions are crucial to document-level relation extraction, which enables the model to extract more meaningful higher-level compositional relations.

In this paper it is shown that $C_\beta$-smooth functions can be approximated by neural networks with parameters $\{0,\pm \frac{1}{2}, \pm 1, 2\}$. The depth, width and the number of active parameters of constructed networks have, up to a logarithimc factor, the same dependence on the approximation error as the networks with parameters in $[-1,1]$. In particular, this means that the nonparametric regression estimation with constructed networks attain the same convergence rate as with the sparse networks with parameters in $[-1,1]$.

Modeling cyber risks has been an important but challenging task in the domain of cyber security. It is mainly because of the high dimensionality and heavy tails of risk patterns. Those obstacles have hindered the development of statistical modeling of the multivariate cyber risks. In this work, we propose a novel approach for modeling the multivariate cyber risks which relies on the deep learning and extreme value theory. The proposed model not only enjoys the high accurate point predictions via deep learning but also can provide the satisfactory high quantile prediction via extreme value theory. The simulation study shows that the proposed model can model the multivariate cyber risks very well and provide satisfactory prediction performances. The empirical evidence based on real honeypot attack data also shows that the proposed model has very satisfactory prediction performances.

Hierarchical forecasting with intermittent time series is a challenge in both research and empirical studies. The overall forecasting performance is heavily affected by the forecasting accuracy of intermittent time series at bottom levels. In this paper, we present a forecasting reconciliation approach that treats the bottom level forecast as latent to ensure higher forecasting accuracy on the upper levels of the hierarchy. We employ a pure deep learning forecasting approach N-BEATS for continuous time series on top levels and a widely used tree-based algorithm LightGBM for the bottom level intermittent time series. The hierarchical forecasting with alignment approach is simple and straightforward to implement in practice. It sheds light on an orthogonal direction for forecasting reconciliation. When there is difficulty finding an optimal reconciliation, allowing suboptimal forecasts at a lower level could retain a high overall performance. The approach in this empirical study was developed by the first author during the M5 Forecasting Accuracy competition ranking second place. The approach is business orientated and could be beneficial for business strategic planning.

We investigate the asymptotic risk of a general class of overparameterized likelihood models, including deep models. The recent empirical success of large-scale models has motivated several theoretical studies to investigate a scenario wherein both the number of samples, $n$, and parameters, $p$, diverge to infinity and derive an asymptotic risk at the limit. However, these theorems are only valid for linear-in-feature models, such as generalized linear regression, kernel regression, and shallow neural networks. Hence, it is difficult to investigate a wider class of nonlinear models, including deep neural networks with three or more layers. In this study, we consider a likelihood maximization problem without the model constraints and analyze the upper bound of an asymptotic risk of an estimator with penalization. Technically, we combine a property of the Fisher information matrix with an extended Marchenko-Pastur law and associate the combination with empirical process techniques. The derived bound is general, as it describes both the double descent and the regularized risk curves, depending on the penalization. Our results are valid without the linear-in-feature constraints on models and allow us to derive the general spectral distributions of a Fisher information matrix from the likelihood. We demonstrate that several explicit models, such as parallel deep neural networks, ensemble learning, and residual networks, are in agreement with our theory. This result indicates that even large and deep models have a small asymptotic risk if they exhibit a specific structure, such as divisibility. To verify this finding, we conduct a real-data experiment with parallel deep neural networks. Our results expand the applicability of the asymptotic risk analysis, and may also contribute to the understanding and application of deep learning.

Conditional quantile estimation is a key statistical learning challenge motivated by the need to quantify uncertainty in predictions or to model a diverse population without being overly reductive. As such, many models have been developed for this problem. Adopting a meta viewpoint, we propose a general framework (inspired by neural network optimization) for aggregating any number of conditional quantile models in order to boost predictive accuracy. We consider weighted ensembling strategies of increasing flexibility where the weights may vary over individual models, quantile levels, and feature values. An appeal of our approach is its portability: we ensure that estimated quantiles at adjacent levels do not cross by applying simple transformations through which gradients can be backpropagated, and this allows us to leverage the modern deep learning toolkit for building quantile ensembles. Our experiments confirm that ensembling can lead to big gains in accuracy, even when the constituent models are themselves powerful and flexible.

Opioid Use Disorder (OUD) is a public health crisis costing the US billions of dollars annually in healthcare, lost workplace productivity, and crime. Analyzing longitudinal healthcare data is critical in addressing many real-world problems in healthcare. Leveraging the real-world longitudinal healthcare data, we propose a novel multi-stream transformer model called MUPOD for OUD prediction. MUPOD is designed to simultaneously analyze multiple types of healthcare data streams, such as medications and diagnoses, by finding the attentions within and across these data streams. Our model tested on the data from 392,492 patients with long-term back pain problems showed significantly better performance than the traditional models and recently developed deep learning models.

Hate speech on social media is a growing concern, and automated methods have so far been sub-par at reliably detecting it. A major challenge lies in the potentially evasive nature of hate speech due to the ambiguity and fast evolution of natural language. To tackle this, we introduce a vectorisation based on a crowd-sourced and continuously updated dictionary of hate words and propose fusing this approach with standard word embedding in order to improve the classification performance of a CNN model. To train and test our model we use a merge of two established datasets (110,748 tweets in total). By adding the dictionary-enhanced input, we are able to increase the CNN model's predictive power and increase the F1 macro score by seven percentage points.

With the rise and ever-increasing potential of deep learning techniques in recent years, publicly available medical data sets became a key factor to enable reproducible development of diagnostic algorithms in the medical domain. Medical data contains sensitive patient-related information and is therefore usually anonymized by removing patient identifiers, e.g., patient names before publication. To the best of our knowledge, we are the first to show that a well-trained deep learning system is able to recover the patient identity from chest X-ray data. We demonstrate this using the publicly available large-scale ChestX-ray14 dataset, a collection of 112,120 frontal-view chest X-ray images from 30,805 unique patients. Our verification system is able to identify whether two frontal chest X-ray images are from the same person with an AUC of 0.9940 and a classification accuracy of 95.55%. We further highlight that the proposed system is able to reveal the same person even ten and more years after the initial scan. When pursuing a retrieval approach, we observe an mAP@R of 0.9748 and a precision@1 of 0.9963. Based on this high identification rate, a potential attacker may leak patient-related information and additionally cross-reference images to obtain more information. Thus, there is a great risk of sensitive content falling into unauthorized hands or being disseminated against the will of the concerned patients. Especially during the COVID-19 pandemic, numerous chest X-ray datasets have been published to advance research. Therefore, such data may be vulnerable to potential attacks by deep learning-based re-identification algorithms.

Conventional Endoscopy (CE) and Wireless Capsule Endoscopy (WCE) are known tools for diagnosing gastrointestinal (GI) tract disorders. Localizing frames provide valuable information about the anomaly location and also can help clinicians determine a more appropriate treatment plan. There are many automated algorithms to detect the anomaly. However, very few of the existing works address the issue of localization. In this study, we present a combination of meta-learning and deep learning for localizing both endoscopy images and video. A dataset is collected from 10 different anatomical positions of human GI tract. In the meta-learning section, the system was trained using 78 CE and 27 WCE annotated frames with a modified Siamese Neural Network (SNN) to predict the location of one single image/frame. Then, a postprocessing section using bidirectional long short-term memory is proposed for localizing a sequence of frames. Here, we have employed feature vector, distance and predicted location obtained from a trained SNN. The postprocessing section is trained and tested on 1,028 and 365 seconds of CE and WCE videos using hold-out validation (50%), and achieved F1-score of 86.3% and 83.0%, respectively. In addition, we performed subjective evaluation using nine gastroenterologists. The results show that the computer-aided methods can outperform gastroenterologists assessment of localization. The proposed method is compared with various approaches, such as support vector machine with hand-crafted features, convolutional neural network and the transfer learning-based methods, and showed better results. Therefore, it can be used in frame localization, which can help in video summarization and anomaly detection.