Background and objectivesIntradialytic hypotension has high clinical significance. However, predicting it using conventional statistical models may be difficult because several factors have interactive and complex effects on the risk. Herein, we applied a deep learning model (recurrent neural network) to predict the risk of intradialytic hypotension using a timestamp-bearing dataset.Design, setting, participants, & measurementsWe obtained 261,647 hemodialysis sessions with 1,600,531 independent timestamps (i.e., time-varying vital signs) and randomly divided them into training (70%), validation (5%), calibration (5%), and testing (20%) sets. Intradialytic hypotension was defined when nadir systolic BP was <90 mm Hg (termed intradialytic hypotension 1) or when a decrease in systolic BP ≥20 mm Hg and/or a decrease in mean arterial pressure ≥10 mm Hg on the basis of the initial BPs (termed intradialytic hypotension 2) or prediction time BPs (termed intradialytic hypotension 3) occurred within 1 hour. The area under the receiver operating characteristic curves, the area under the precision-recall curves, and F1 scores obtained using the recurrent neural network model were compared with those obtained using multilayer perceptron, Light Gradient Boosting Machine, and logistic regression models.ResultsThe recurrent neural network model for predicting intradialytic hypotension 1 achieved an area under the receiver operating characteristic curve of 0.94 (95% confidence intervals, 0.94 to 0.94), which was higher than those obtained using the other models (P<0.001). The recurrent neural network model for predicting intradialytic hypotension 2 and intradialytic hypotension 3 achieved area under the receiver operating characteristic curves of 0.87 (interquartile range, 0.87–0.87) and 0.79 (interquartile range, 0.79–0.79), respectively, which were also higher than those obtained using the other models (P≤0.001). The area under the precision-recall curve and F1 score were higher using the recurrent neural network model than they were using the other models. The recurrent neural network models for intradialytic hypotension were highly calibrated.ConclusionsOur deep learning model can be used to predict the real-time risk of intradialytic hypotension.

We propose the position-based scaled gradient (PSG) that scales the gradient depending on the position of a weight vector to make it more compression-friendly. First, we theoretically show that applying PSG to the standard gradient descent (GD), which is called PSGD, is equivalent to the GD in the warped weight space, a space made by warping the original weight space via an appropriately designed invertible function. Second, we empirically show that PSG acting as a regularizer to the weight vectors is favorable for model compression domains such as quantization, pruning, and knowledge distillation. PSG reduces the gap between the weight distributions of a full-precision model and its compressed counterpart. This enables the versatile deployment of a model either as an uncompressed mode or as a compressed mode depending on the availability of resources. The experimental results on CIFAR-10/100 and ImageNet datasets show the effectiveness of the proposed PSG in model compression including an iterative pruning method and the knowledge distillation.

Word-level adversarial attacks have shown success in NLP models, drastically decreasing the performance of transformer-based models in recent years. As a countermeasure, adversarial defense has been explored, but relatively few efforts have been made to detect adversarial examples. However, detecting adversarial examples may be crucial for automated tasks (e.g. review sentiment analysis) that wish to amass information about a certain population and additionally be a step towards a robust defense system. To this end, we release a dataset for four popular attack methods on four datasets and four models to encourage further research in this field. Along with it, we propose a competitive baseline based on density estimation that has the highest AUC on 29 out of 30 dataset-attack-model combinations. 1

Word-level adversarial attacks have shown success in NLP models, drastically decreasing the performance of transformer-based models in recent years. As a countermeasure, adversarial defense has been explored, but relatively few efforts have been made to detect adversarial examples. However, detecting adversarial examples may be crucial for automated tasks (e.g. review sentiment analysis) that wish to amass information about a certain population and additionally be a step towards a robust defense system. To this end, we release a dataset for four popular attack methods on four datasets and four models to encourage further research in this field. Along with it, we propose a competitive baseline based on density estimation that has the highest AUC on 29 out of 30 dataset-attack-model combinations. The source code is released. 1

Recent advances in federated learning have demonstrated its promising capability to learn on decentralized datasets. However, a considerable amount of work has raised concerns due to the potential risks of adversaries participating in the framework to poison the global model for an adversarial purpose. This paper investigates the feasibility of model poisoning for backdoor attacks through rare word embeddings of NLP models in text classification and sequence-to-sequence tasks. In text classification, less than 1% of adversary clients suffices to manipulate the model output without any drop in the performance of clean sentences. For a less complex dataset, a mere 0.1% of adversary clients is enough to poison the global model effectively. We also propose a technique specialized in the federated learning scheme called gradient ensemble, which enhances the backdoor performance in all experimental settings.