The challenge of recognizing named entities in a given text has been a very dynamic field in recent years. This is due to the advances in neural network architectures, increase of computing power and the availability of diverse labeled datasets, which deliver pre-trained, highly accurate models. These tasks are generally focused on tagging common entities, but domain-specific use-cases require tagging custom entities which are not part of the pre-trained models. This can be solved by either fine-tuning the pre-trained models, or by training custom models. The main challenge lies in obtaining reliable labeled training and test datasets, and manual labeling would be a highly tedious task. In this paper we present PharmKE, a text analysis platform focused on the pharmaceutical domain, which applies deep learning through several stages for thorough semantic analysis of pharmaceutical articles. It performs text classification using state-of-the-art transfer learning models, and thoroughly integrates the results obtained through a proposed methodology. The methodology is used to create accurately labeled training and test datasets, which are then used to train models for custom entity labeling tasks, centered on the pharmaceutical domain. The obtained results are compared to the fine-tuned BERT and BioBERT models trained on the same dataset. Additionally, the PharmKE platform integrates the results obtained from named entity recognition tasks to resolve co-references of entities and analyze the semantic relations in every sentence, thus setting up a baseline for additional text analysis tasks, such as question answering and fact extraction. The recognized entities are also used to expand the knowledge graph generated by DBpedia Spotlight for a given pharmaceutical text.

Knowledge distillation is classically a procedure where a neural network is trained on the output of another network along with the original targets in order to transfer knowledge between the architectures. The special case of self-distillation, where the network architectures are identical, has been observed to improve generalization accuracy. In this paper, we consider an iterative variant of self-distillation in a kernel regression setting, in which successive steps incorporate both model outputs and the ground-truth targets. This allows us to provide the first theoretical results on the importance of using the weighted ground-truth targets in self-distillation. Our focus is on fitting nonlinear functions to training data with a weighted mean square error objective function suitable for distillation, subject to $\ell_2$ regularization of the model parameters. We show that any such function obtained with self-distillation can be calculated directly as a function of the initial fit, and that infinite distillation steps yields the same optimization problem as the original with amplified regularization. Finally, we examine empirically, both in a regression setting and with ResNet networks, how the choice of weighting parameter influences the generalization performance after self-distillation.

Background: Pre-training large-scale neural language models on raw texts has made a significant contribution to improving transfer learning in natural language processing (NLP). With the introduction of transformer-based language models, such as bidirectional encoder representations from transformers (BERT), the performance of information extraction from a free text by NLP has significantly improved for both the general domain and medical domain; however, it is difficult to train specific BERT models that perform well for domains in which there are few publicly available databases of high quality and large size. We hypothesized that this problem can be addressed by up-sampling a domain-specific corpus and using it for pre-training with a larger corpus in a balanced manner. Method: Our proposed method consists of a single intervention with one option: simultaneous pretraining after up-sampling, which is intended to encourage masked language modeling and next-sentence prediction on the small corpus, and amplified vocabulary, which helps to suit the smaller corpus when building the customized vocabulary by byte-pair encoding. We conducted three experiments and evaluated the resulting products; (1) Japanese medical BERT from a small medical corpus, (2) English biomedical BERT from a small biomedical corpus, and (3) Enhanced biomedical BERT from whole PubMed abstracts. Results: We confirmed that our Japanese medical BERT outperformed conventional baselines and the other BERT models in terms of the medical document classification task and that our English BERT pre-trained using both the general and medical-domain corpora performed sufficiently well for practical use in terms of the biomedical language understanding evaluation (BLUE) benchmark. Moreover, our enhanced biomedical BERT model, in which clinical notes were not used during pre-training, showed that both the clinical and biomedical scores of the BLUE benchmark were 0.3 points above that of the ablation model trained without our proposed method. Conclusions: Well-balanced pre-training by up-sampling instances derived from a corpus appropriate for the target task allows us to construct a high-performance BERT model.

A number of machine learning tasks entail a high degree of invariance: the data distribution does not change if we act on the data with a certain group of transformations. For instance, labels of images are invariant under translations of the images. Certain neural network architectures -- for instance, convolutional networks -- are believed to owe their success to the fact that they exploit such invariance properties. With the objective of quantifying the gain achieved by invariant architectures, we introduce two classes of models: invariant random features and invariant kernel methods. The latter includes, as a special case, the neural tangent kernel for convolutional networks with global average pooling. We consider uniform covariates distributions on the sphere and hypercube and a general invariant target function. We characterize the test error of invariant methods in a high-dimensional regime in which the sample size and number of hidden units scale as polynomials in the dimension, for a class of groups that we call `degeneracy $\alpha$', with $\alpha \leq 1$. We show that exploiting invariance in the architecture saves a $d^\alpha$ factor ($d$ stands for the dimension) in sample size and number of hidden units to achieve the same test error as for unstructured architectures. Finally, we show that output symmetrization of an unstructured kernel estimator does not give a significant statistical improvement; on the other hand, data augmentation with an unstructured kernel estimator is equivalent to an invariant kernel estimator and enjoys the same improvement in statistical efficiency.

Traditional statistics forbids use of test data (a.k.a. holdout data) during training. Dwork et al. 2015 pointed out that current practices in machine learning, whereby researchers build upon each other's models, copying hyperparameters and even computer code -- amounts to implicitly training on the test set. Thus error rate on test data may not reflect the true population error. This observation initiated {\em adaptive data analysis}, which provides evaluation mechanisms with guaranteed upper bounds on this difference. With statistical query (i.e. test accuracy) feedbacks, the best upper bound is fairly pessimistic: the deviation can hit a practically vacuous value if the number of models tested is quadratic in the size of the test set. In this work, we present a simple new estimate, {\em Rip van Winkle's Razor}. It relies upon a new notion of \textquotedblleft information content\textquotedblright\ of a model: the amount of information that would have to be provided to an expert referee who is intimately familiar with the field and relevant science/math, and who has been just been woken up after falling asleep at the moment of the creation of the test data (like \textquotedblleft Rip van Winkle\textquotedblright\ of the famous fairy tale). This notion of information content is used to provide an estimate of the above deviation which is shown to be non-vacuous in many modern settings.

Integrating physics models within machine learning holds considerable promise toward learning robust models with improved interpretability and abilities to extrapolate. In this work, we focus on the integration of incomplete physics models into deep generative models, variational autoencoders (VAEs) in particular. A key technical challenge is to strike a balance between the incomplete physics model and the learned components (i.e., neural nets) of the complete model, in order to ensure that the physics part is used in a meaningful manner. To this end, we propose a VAE architecture in which a part of the latent space is grounded by physics. We couple it with a set of regularizers that control the effect of the learned components and preserve the semantics of the physics-based latent variables as intended. We not only demonstrate generative performance improvements over a set of synthetic and real-world datasets, but we also show that we learn robust models that can consistently extrapolate beyond the training distribution in a meaningful manner. Moreover, we show that we can control the generative process in an interpretable manner.

With a better understanding of the loss surfaces for multilayer networks, we can build more robust and accurate training procedures. Recently it was discovered that independently trained SGD solutions can be connected along one-dimensional paths of near-constant training loss. In this paper, we show that there are mode-connecting simplicial complexes that form multi-dimensional manifolds of low loss, connecting many independently trained models. Inspired by this discovery, we show how to efficiently build simplicial complexes for fast ensembling, outperforming independently trained deep ensembles in accuracy, calibration, and robustness to dataset shift. Notably, our approach only requires a few training epochs to discover a low-loss simplex, starting from a pre-trained solution. Code is available at https://github.com/g-benton/loss-surface-simplexes.

Commonly, Deep Neural Networks (DNNs) generalize well on samples drawn from a distribution similar to that of the training set. However, DNNs' predictions are brittle and unreliable when the test samples are drawn from a dissimilar distribution. This presents a major concern for deployment in real-world applications, where such behavior may come at a great cost -- as in the case of autonomous vehicles or healthcare applications. This paper frames the Out Of Distribution (OOD) detection problem in DNN as a statistical hypothesis testing problem. Unlike previous OOD detection heuristics, our framework is guaranteed to maintain the false positive rate (detecting OOD as in-distribution) for test data. We build on this framework to suggest a novel OOD procedure based on low-order statistics. Our method achieves comparable or better than state-of-the-art results on well-accepted OOD benchmarks without retraining the network parameters -- and at a fraction of the computational cost.

Recently, some researches are devoted to the topic of end-to-end learning a physical layer secure communication system based on autoencoder under Gaussian wiretap channel. However, in those works, the reliability and security of the encoder model were learned through necessary decoding outputs of not only legitimate receiver but also the eavesdropper. In fact, the assumption of known eavesdropper's decoder or its output is not practical. To address this issue, in this paper we propose a dual mutual information neural estimation (MINE) based neural secure communications model. The security constraints of this method is constructed only with the input and output signal samples of the legal and eavesdropper channels and benefit that training the encoder is completely independent of the decoder. Moreover, since the design of secure coding does not rely on the eavesdropper's decoding results, the security performance would not be affected by the eavesdropper's decoding means. Numerical results show that the performance of our model is guaranteed whether the eavesdropper learns the decoder himself or uses the legal decoder.

Objective: This paper presents an Alzheimer's disease (AD) detection method based on learning structural similarity between Magnetic Resonance Images (MRIs) and representing this similarity as a graph. Methods: We construct the similarity graph using embedded features of the input image (i.e., Non-Demented (ND), Very Mild Demented (VMD), Mild Demented (MD), and Moderated Demented (MDTD)). We experiment and compare different dimension-reduction and clustering algorithms to construct the best similarity graph to capture the similarity between the same class images using the cosine distance as a similarity measure. We utilize the similarity graph to present (sample) the training data to a convolutional neural network (CNN). We use the similarity graph as a regularizer in the loss function of a CNN model to minimize the distance between the input images and their k-nearest neighbours in the similarity graph while minimizing the categorical cross-entropy loss between the training image predictions and the actual image class labels. Results: We conduct extensive experiments with several pre-trained CNN models and compare the results to other recent methods. Conclusion: Our method achieves superior performance on the testing dataset (accuracy = 0.986, area under receiver operating characteristics curve = 0.998, F1 measure = 0.987). Significance: The classification results show an improvement in the prediction accuracy compared to the other methods. We release all the code used in our experiments to encourage reproducible research in this area