Clinical notes in electronic health records contain highly heterogeneous writing styles, including non-standard terminology or abbreviations. Using these notes in predictive modeling has traditionally required preprocessing (e.g. taking frequent terms or topic modeling) that removes much of the richness of the source data. We propose a pretrained hierarchical recurrent neural network model that parses minimally processed clinical notes in an intuitive fashion, and show that it improves performance for multiple classification tasks on the Medical Information Mart for Intensive Care III (MIMIC-III) dataset, increasing top-5 recall to 89.7% (up by 4.8%) for primary diagnosis classification and AUPRC to 35.2% (up by 2.4%) for multilabel diagnosis classification compared to models that treat the notes as an unordered collection of terms or without pretraining. We also apply an attribution technique to several examples to identify the words and the nearby context that the model uses to make its prediction, and show the importance of the words' context.

Neural models for NLP typically use large numbers of parameters to reach state-of-the-art performance, which can lead to excessive memory usage and increased runtime. We present a structure learning method for learning sparse, parameter-efficient NLP models. Our method applies group lasso to rational RNNs (Peng et al., 2018), a family of models that is closely connected to weighted finite-state automata (WFSAs). We take advantage of rational RNNs' natural grouping of the weights, so the group lasso penalty directly removes WFSA states, substantially reducing the number of parameters in the model. Our experiments on a number of sentiment analysis datasets, using both GloVe and BERT embeddings, show that our approach learns neural structures which have fewer parameters without sacrificing performance relative to parameter-rich baselines. Our method also highlights the interpretable properties of rational RNNs. We show that sparsifying such models makes them easier to visualize, and we present models that rely exclusively on as few as three WFSAs after pruning more than 90% of the weights. We publicly release our code.

We investigate whether it's possible to tighten P AC-Bayes bounds for deep neural networks by utilizing the Hessian of the training loss at the minimum. For the case of Gaussian priors and posteriors we introduce a Hessian-based method to obtain tighter P AC-Bayes bounds that relies on closed form solutions of layerwise sub-problems. We thus avoid commonly used variational inference techniques which can be difficult to implement and time consuming for modern deep architectures. Through careful experiments we analyze the influence of the prior mean, prior covariance, posterior mean and posterior covariance on obtaining tighter bounds. We discuss several limitations in further improving P AC-Bayes bounds through more informative priors. Deep neural networks are by now the established method for tackling a number of machine learning tasks. Despite this usually their performance on out of sample data is extremely difficult to be proven formaly and is usually validated empirically by using a validation set of samples. Classic measures of capacity such as the VC dimension which are uniform across all functions representable by the classification architecture are doomed to fail; DNNs are typically overparameterized and correspondingly the set of representable functions is large enough to make the the bounds vacuous.

Objective: The aim of this study is to develop an automated classification algorithm for polysomnography (PSG) recordings to detect non-apneic and non-hypopneic arousals. Our particular focus is on detecting the respiratory effort-related arousals (RERAs) which are very subtle respiratory events that do not meet the criteria for apnea or hypopnea, and are more challenging to detect. Methods: The proposed algorithm is based on a bidirectional long short-term memory (BiLSTM) classifier and 465 multi-domain features, extracted from multimodal clinical time series. The features consist of a set of physiology-inspired features (n = 75), obtained by multiple steps of feature selection and expert analysis, and a set of physiology-agnostic features (n = 390), derived from scattering transform. Results: The proposed algorithm is validated on the 2018 PhysioNet challenge dataset. The overall performance in terms of the area under the precision-recall curve (AUPRC) is 0.50 on the hidden test dataset. This result is tied for the second-best score during the follow-up and official phases of the 2018 PhysioNet challenge. Conclusions: The results demonstrate that it is possible to automatically detect subtle non-apneic/non-hypopneic arousal events from PSG recordings. Significance: Automatic detection of subtle respiratory events such as RERAs together with other non-apneic/non-hypopneic arousals will allow detailed annotations of large PSG databases. This contributes to a better retrospective analysis of sleep data, which may also improve the quality of treatment.

Self-supervised bidirectional transformer models such as BERT have led to dramatic improvements in a wide variety of textual classification tasks. The modern digital world is increasingly multimodal, however, and textual information is often accompanied by other modalities such as images. We introduce a supervised multimodal bitransformer model that fuses information from text and image encoders, and obtain state-of-the-art performance on various multimodal classification benchmark tasks, outperforming strong baselines, including on hard test sets specifically designed to measure multimodal performance.

Recent research on reinforcement learning has shown that trained agents are vulnerable to maliciously crafted adversarial samples. In this work, we show how adversarial samples against RL agents can be generalised from White-box and Grey-box attacks to a strong Black-box case, namely where the attacker has no knowledge of the agents and their training methods. We use sequence-to-sequence models to predict a single action or a sequence of future actions that a trained agent will make. Our approximation model, based on time-series information from the agent, successfully predicts agents' future actions with consistently above 80% accuracy on a wide range of games and training methods. Second, we find that although such adversarial samples are transferable, they do not outperform random Gaussian noise as a means of reducing the game scores of trained RL agents. This highlights a serious methodological deficiency in previous work on such agents; random jamming should have been taken as the baseline for evaluation. Third, we do find a novel use for adversarial samples in this context: they can be used to trigger a trained agent to misbehave after a specific delay. This appears to be a genuinely new type of attack; it potentially enables an attacker to use devices controlled by RL agents as time bombs.

We discuss training techniques, objectives and metrics toward mass personalization of deep learning models. In machine learning, personalization refers to the fact that every trained model should be targeted towards an individual by optimizing one or several performance metrics and often obeying additional constraints. We investigate three methods for personalized training of neural networks. They constitute three forms of curriculum learning. The methods are partially inspired by the "shaping" concept from psychology. Interestingly, we discover that extensive exposure to a limited set of training data in terms of class diversity \emph{early} in the training can lead to an irreversible reduction of the capability of a network to learn from more diverse training data. This is in close alignment with existing theories in human development. In contrast, training on a small data set covering all classes \emph{early} in the training can lead to better performance.

Prediction and interpolation for long-range video data involves the complex task of modeling motion trajectories for each visible object, occlusions and dis-occlusions, as well as appearance changes due to viewpoint and lighting. Optical flow based techniques generalize but are suitable only for short temporal ranges. Many methods opt to project the video frames to a low dimensional latent space, achieving long-range predictions. However, these latent representations are often non-interpretable, and therefore difficult to manipulate. This work poses video prediction and interpolation as unsupervised latent structure inference followed by a temporal prediction in this latent space. The latent representations capture foreground semantics without explicit supervision such as keypoints or poses. Further, as each landmark can be mapped to a coordinate indicating where a semantic part is positioned, we can reliably interpolate within the coordinate domain to achieve predictable motion interpolation. Given an image decoder capable of mapping these landmarks back to the image domain, we are able to achieve high-quality long-range video interpolation and extrapolation by operating on the landmark representation space.

Fine-tuning a deep network trained with the standard cross-entropy loss is a strong baseline for few-shot learning. When fine-tuned transductively, this outperforms the current state-of-the-art on standard datasets such as Mini-Imagenet, Tiered-Imagenet, CIFAR-FS and FC-100 with the same hyper-parameters. The simplicity of this approach enables us to demonstrate the first few-shot learning results on the Imagenet-21k dataset. We find that using a large number of meta-training classes results in high few-shot accuracies even for a large number of test classes. We do not advocate our approach as the solution for few-shot learning, but simply use the results to highlight limitations of current benchmarks and few-shot protocols. We perform extensive studies on benchmark datasets to propose a metric that quantifies the "hardness" of a test episode. This metric can be used to report the performance of few-shot algorithms in a more systematic way.

Attention mechanisms have become ubiquitous in NLP. Recent architectures, notably the Transformer, learn powerful context-aware word representations through layered, multi-headed attention. The multiple heads learn diverse types of word relationships. However, with standard softmax attention, all attention heads are dense, assigning a non-zero weight to all context words. In this work, we introduce the adaptively sparse Transformer, wherein attention heads have flexible, context-dependent sparsity patterns. This sparsity is accomplished by replacing softmax with $\alpha$-entmax: a differentiable generalization of softmax that allows low-scoring words to receive precisely zero weight. Moreover, we derive a method to automatically learn the $\alpha$ parameter -- which controls the shape and sparsity of $\alpha$-entmax -- allowing attention heads to choose between focused or spread-out behavior. Our adaptively sparse Transformer improves interpretability and head diversity when compared to softmax Transformers on machine translation datasets. Findings of the quantitative and qualitative analysis of our approach include that heads in different layers learn different sparsity preferences and tend to be more diverse in their attention distributions than softmax Transformers. Furthermore, at no cost in accuracy, sparsity in attention heads helps to uncover different head specializations.