Abstract
Due to convenience and usability, many deep learning (DL) jobs resort to the available shared parallel file system (PFS) for storing and accessing training data when running in HPC environments. Under such a scenario, however, where multiple I/O-intensive applications operate concurrently, the PFS can quickly get saturated with simultaneous storage requests and become a critical performance bottleneck, leading to throughput variability and performance loss. We present MONARCH, a framework-agnostic middleware for hierarchical storage management. This solution leverages the existing storage tiers present at modern supercomputers (e.g., compute node's local storage, PFS) to improve DL training performance and alleviate the current I/O pressure of the shared PFS. We validate the applicability of our approach by developing and integrating an early prototype with the TensorFlow DL framework. Results show that MONARCH can reduce I/O operations submitted to the shared PFS by up to 45%, decreasing training time by 24% and 12%, for I/O-intensive models, namely LeNet and AlexNet.

Abstract
Transformer-based language models have taken many fields in NLP by storm. BERT and its derivatives dominate most of the existing evaluation benchmarks, including those for Word Sense Disambiguation (WSD), thanks to their ability in capturing context-sensitive semantic nuances. However, there is still little knowledge about their capabilities and potential limitations in encoding and recovering word senses. In this article, we provide an in-depth quantitative and qualitative analysis of the celebrated BERT model with respect to lexical ambiguity. One of the main conclusions of our analysis is that BERT can accurately capture high-level sense distinctions, even when a limited number of examples is available for each word sense. Our analysis also reveals that in some cases language models come close to solving coarse grained noun disambiguation under ideal conditions in terms of availability of training data and computing resources. However, this scenario rarely occurs in real-world settings and, hence, many practical challenges remain even in the coarse-grained setting. We also perform an in-depth comparison of the two main language model-based WSD strategies, namely, fine-tuning and feature extraction, finding that the latter approach is more robust with respect to sense bias and it can better exploit limited available training data. In fact, the simple feature extraction strategy of averaging contextualized embeddings proves robust even using only three training sentences per word sense, with minimal improvements obtained by increasing the size of this training data.

Abstract
IntroductionPeople with schizophrenia experience higher levels of stigma compared with other diseases. The analysis of social media content is a tool of great importance to understand the public opinion toward a particular topic.ObjectivesThe aim of this study is to analyse the content of social media on schizophrenia and the most prevalent sentiments towards this disorder.MethodsTweets were retrieved using Twitter’s Application Programming Interface and the keyword “schizophrenia”. Parameters were set to allow the retrieval of recent and popular tweets on the topic and no restrictions were made in terms of geolocation. Analysis of 8 basic emotions (anger, anticipation, disgust, fear, joy, sadness, surprise, and trust) was conducted automatically using a lexicon-based approach and the NRC Word-Emotion Association Lexicon.ResultsTweets on schizophrenia were heterogeneous. The most prevalent sentiments on the topic were mainly negative, namely anger, fear, sadness and disgust. Qualitative analyses of the most retweeted posts added insight into the nature of the public dialogue on schizophrenia.ConclusionsAnalyses of social media content can add value to the research on stigma toward psychiatric disorders. This tool is of growing importance in many fields and further research in mental health can help the development of public health strategies in order to decrease the stigma towards psychiatric disorders.

Abstract
This work presents a framework of dynamic structural models with covariates for short-term forecasting of time series with complex seasonal patterns. The framework is based on the multiple sources of randomness formulation. A noise model is formulated to allow the incorporation of randomness into the seasonal component and to propagate this same randomness in the coefficients of the variant trigonometric terms over time. A unique, recursive and systematic computational procedure based on the maximum likelihood estimation under the hypothesis of Gaussian errors is introduced. The referred procedure combines the Kalman filter with recursive adjustment of the covariance matrices and the selection method of harmonics number in the trigonometric terms. A key feature of this method is that it allows estimating not only the states of the system but also allows obtaining the standard errors of the estimated parameters and the prediction intervals. In addition, this work also presents a non-parametric bootstrap approach to improve the forecasting method based on Kalman filter recursions. The proposed framework is empirically explored with two real time series.

Abstract
Artificial Intelligence (AI) is continuously improving several aspects of our daily lives. There has been a great use of gadgets & monitoring devices for health and physical activity monitoring. Thus, by analyzing large amounts of data and applying Machine Learning (ML) techniques, we have been able to infer fruitful conclusions in various contexts. Activity Recognition is one of them, in which it is possible to recognize and monitor our daily actions. The main focus of the traditional systems is only to detect pre-established activities according to the previously configured parameters, and not to detect novel ones. However, when applying activity recognizers in real-world applications, it is necessary to detect new activities that were not considered during the training of the model. We propose a method for Novelty Detection in the context of physical activity. Our solution is based on the establishment of a threshold confidence value, which determines whether an activity is novel or not. We built and train our models by experimenting with three different algorithms and four threshold values. The best results were obtained by using the Random Forest algorithm with a threshold value of 0.8, resulting in 90.9% of accuracy and 85.1% for precision.

Abstract
Thin graphene oxide (GO) film layers are being widely used as sensing layers in different types of electrical and optical sensor devices. GO layers are particularly popular because of their tuned interface reflectivity. The stability of GO layers is fundamental for sensor device reliability, particularly in complex aqueous environments such as wastewater. In this work, the stability of GO layers in layer-by-layer (LbL) films of polyethyleneimine (PEI) and GO was investigated. The results led to the following conclusions: PEI/GO films grow linearly with the number of bilayers as long as the adsorption time is kept constant; the adsorption kinetics of a GO layer follow the behavior of the adsorption of polyelectrolytes; and the interaction associated with the growth of these films is of the ionic type since the desorption activation energy has a value of 119 +/- 17 kJ/mol. Therefore, it is possible to conclude that PEI/GO films are suitable for application in optical fiber sensor devices; most importantly, an optical fiber-based interrogation setup can easily be adapted to investigate in situ desorption via a thermally stimulated process. In addition, it is possible to draw inferences about film stability in solution in a fast, reliable way when compared with the traditional ones.