Modern weather forecast models perform uncertainty quantification using ensemble prediction systems, which collect nonparametric statistics based on multiple perturbed simulations. To provide accurate estimation, dozens of such computationally intensive simulations must be run. We show that deep neural networks can be used on a small set of numerical weather simulations to estimate the spread of a weather forecast, significantly reducing computational cost. To train the system, we both modify the 3D U-Net architecture and explore models that incorporate temporal data. Our models serve as a starting point to improve uncertainty quantification in current real-time weather forecasting systems, which is vital for predicting extreme events.

The current paradigm for using machine learning models on a device is to train a model in the cloud and perform inference using the trained model on the device. However, with the increasing number of smart devices and improved hardware, there is interest in performing model training on the device. Given this surge in interest, a comprehensive survey of the field from a device-agnostic perspective sets the stage for both understanding the state-of-the-art and for identifying open challenges and future avenues of research. Since on-device learning is an expansive field with connections to a large number of related topics in AI and machine learning (including online learning, model adaptation, one/few-shot learning, etc), covering such a large number of topics in a single survey is impractical. Instead, this survey finds a middle ground by reformulating the problem of on-device learning as resource constrained learning where the resources are compute and memory. This reformulation allows tools, techniques, and algorithms from a wide variety of research areas to be compared equitably. In addition to summarizing the state of the art, the survey also identifies a number of challenges and next steps for both the algorithmic and theoretical aspects of on-device learning.

In this paper we offer a method and algorithm, which make possible fully autonomous (unsupervised) detection of new classes, and learning following a very parsimonious training priming (few labeled data samples only). Moreover, new unknown classes may appear at a later stage and the proposed xClass method and algorithm are able to successfully discover this and learn from the data autonomously. Furthermore, the features (inputs to the classifier) are automatically sub-selected by the algorithm based on the accumulated data density per feature per class. As a result, a highly efficient, lean, human-understandable, autonomously self-learning model (which only needs an extremely parsimonious priming) emerges from the data. To validate our proposal we tested it on two challenging problems, including imbalanced Caltech-101 data set and iRoads dataset. Not only we achieved higher precision, but, more significantly, we only used a single class beforehand, while other methods used all the available classes) and we generated interpretable models with smaller number of features used, through extremely weak and weak supervision.

An accurate road surface friction prediction algorithm can enable intelligent transportation systems to share timely road surface condition to the public for increasing the safety of the road users. Previously, scholars developed multiple prediction models for forecasting road surface conditions using historical data. However, road surface condition data cannot be perfectly collected at every timestamp, e.g. the data collected by on-vehicle sensors may be influenced when vehicles cannot travel due to economic cost issue or weather issues. Such resulted missing values in the collected data can damage the effectiveness and accuracy of the existing prediction methods since they are assumed to have the input data with a fixed temporal resolution. This study proposed a road surface friction prediction model employing a Gated Recurrent Unit network-based decay mechanism (GRU-D) to handle the missing values. The evaluation results present that the proposed GRU-D networks outperform all baseline models. The impact of missing rate on predictive accuracy, learning efficiency and learned decay rate are analyzed as well. The findings can help improve the prediction accuracy and efficiency of forecasting road surface friction using historical data sets with missing values, therefore mitigating the impact of wet or icy road conditions on traffic safety.

Deep learning (DL) methods have in recent years yielded impressive results in medical imaging, with the potential to function as clinical aid to radiologists. However, DL models in medical imaging are often trained on public research cohorts with images acquired with a single scanner or with strict protocol harmonization, which is not representative of a clinical setting. The aim of this study was to investigate how well a DL model performs in unseen clinical data sets---collected with different scanners, protocols and disease populations---and whether more heterogeneous training data improves generalization. In total, 3117 MRI scans of brains from multiple dementia research cohorts and memory clinics, that had been visually rated by a neuroradiologist according to Scheltens' scale of medial temporal atrophy (MTA), were included in this study. By training multiple versions of a convolutional neural network on different subsets of this data to predict MTA ratings, we assessed the impact of including images from a wider distribution during training had on performance in external memory clinic data. Our results showed that our model generalized well to data sets acquired with similar protocols as the training data, but substantially worse in clinical cohorts with visibly different tissue contrasts in the images. This implies that future DL studies investigating performance in out-of-distribution (OOD) MRI data need to assess multiple external cohorts for reliable results. Further, by including data from a wider range of scanners and protocols the performance improved in OOD data, which suggests that more heterogeneous training data makes the model generalize better. To conclude, this is the most comprehensive study to date investigating the domain shift in deep learning on MRI data, and we advocate rigorous evaluation of DL models on clinical data prior to being certified for deployment.

Wide accessibility of imaging and profile sensors in modern industrial systems created an abundance of high-dimensional sensing variables. This led to a a growing interest in the research of high-dimensional process monitoring. However, most of the approaches in the literature assume the in-control population to lie on a linear manifold with a given basis (i.e., spline, wavelet, kernel, etc) or an unknown basis (i.e., principal component analysis and its variants), which cannot be used to efficiently model profiles with a nonlinear manifold which is common in many real-life cases. We propose deep probabilistic autoencoders as a viable unsupervised learning approach to model such manifolds. To do so, we formulate nonlinear and probabilistic extensions of the monitoring statistics from classical approaches as the expected reconstruction error (ERE) and the KL-divergence (KLD) based monitoring statistics. Through extensive simulation study, we provide insights on why latent-space based statistics are unreliable and why residual-space based ones typically perform much better for deep learning based approaches. Finally, we demonstrate the superiority of deep probabilistic models via both simulation study and a real-life case study involving images of defects from a hot steel rolling process.

Deep learning training accesses vast amounts of data at high velocity, posing challenges for datasets retrieved over commodity networks and storage devices. We introduce a way to dynamically reduce the overhead of fetching and transporting training data with a method we term Progressive Compressed Records (PCRs). PCRs deviate from previous formats by using progressive compression to convert a single dataset into multiple datasets of increasing fidelity--all without adding to the total dataset size. Empirically, we implement PCRs and evaluate them on a wide range of datasets: ImageNet, HAM10000, Stanford Cars, and CelebA-HQ. Our results show that different tasks can tolerate different levels of compression. PCRs use an on-disk layout that enables applications to efficiently and dynamically access appropriate levels of compression at runtime. In turn, we demonstrate that PCRs can seamlessly enable a 2 speedup in training time on average over baseline formats. Distributed deep learning exploits parallelism to reduce training time, and consists of three key components: the data pipeline (storage), the forward/backward computation (compute), and the variable synchronization (network). However, little attention has been paid toward scaling the storage layer, where training starts and training data is sourced. Unfortunately, hardware trends point to an increasing divide between compute and networking or storage bandwidth (Li et al., 2016; Lim et al., 2019; Kurth et al., 2018). For example, the transportation of data for machine learning is a key factor in the design of modern data centers (Hazelwood et al., 2018), which are expected to be serviced by slow, yet high capacity, storage media for the foreseeable future (David Reinsel, 2018; Cheng et al., 2015; Rosenthal et al., 2012). This, combined with the memory wall--a lack of bandwidth between compute and memory--suggests that, while computation may be sufficient moving forward, the mechanisms for moving data to the compute may not (Wulf & McKee, 1995; Kwon & Rhu, 2018; Hsieh et al., 2017; Zinkevich et al., 2010). The storage pipeline is therefore a natural area to seek improvements in overall training times, which manifest from the storage medium, through the network, and into the compute nodes.

Starck 1 1 Laboratoire AIM, CEA, CNRS, Université Paris-Saclay, Université Paris Diderot, Sorbonne Paris Cité, F-91191 Gif-sur-Yvette, France 2 ONERA - The French Aerospace Lab, 6 chemin de la V auve aux Granges, BP 80100, FR-91123 P ALAISEAU cedex, France November 4, 2019 ABSTRACT Deconvolution of large survey images with millions of galaxies requires to develop a new generation of methods which can take into account a space variant Point Spread Function (PSF) and have to be at the same time accurate and fast. We investigate in this paper how Deep Learning (DL) could be used to perform this task. We employ a U-NET Deep Neural Network (DNN) architecture to learn in a supervised setting parameters adapted for galaxy image processing and study two strategies for deconvolution. The first approach is a post-processing of a mere Tikhonov deconvolution with closed form solution and the second one is an iterative deconvolution framework based on the Alternating Direction Method of Multipliers (ADMM). Our numerical results based on GREA T3 simulations with realistic galaxy images and PSFs show that our two approaches outperforms standard techniques based on convex optimization, whether assessed in galaxy image reconstruction or shape recovery. The approach based on Tikhonov deconvolution leads to the most accurate results except for ellipticity errors at high signal to noise ratio where the ADMM approach performs slightly better, is also more computation-time e fficient to process a large number of galaxies, and is therefore recommended in this scenario. Methods:statistical, Methods:data analysis, Methods:numerical 1. Introduction Deconvolution of large galaxy survey images requires to take into account spatial-variation of the Point Spread Function (PSF) across the field of view. The PSF field is usually estimated beforehand, via parametric models and simulations as in Krist et al. (2011) or directly estimated from the (noisy) observations of stars in the field of view (Bertin 2011; Kuijken et al. 2015; Zuntz et al. 2018; Mboula et al. 2016; Schmitz et al. 2019). Even with the "perfect" knowledge of the PSF, this ill-posed deconvolution problem is challenging, in particular due to the size of the image to process. Starck et al. (2000) proposed an Object-Oriented Deconvolution, consisting in first detecting galaxies and then deconvolving each object independently. Following this idea, Farrens et al. (2017) introduced a space-variant deconvolution approach for galaxy images, based on two regularization strategies: using either a sparse prior in a transformed domain (Starck et al. 2015a) or trying to learn unsupervisedly a low-dimensional subspace for galaxy representation using a low-rank prior on the recovered galaxy images.

We present a novel framework to deal with relation extraction tasks in cases where there is complete lack of supervision, either in the form of gold annotations, or relations from a knowledge base. Our approach leverages syntactic parsing and pre-trained word embeddings to extract few but precise relations,which are then used to annotate a larger cor-pus, in a manner identical to distant supervision. The resulting data set is employed to fine tune a pre-trained BERT model in order to perform relation extraction. Empirical evaluation on four data sets from the biomedical domain shows that our method significantly outperforms two simple baselines for unsupervised relation extraction and, even if not using any supervision at all, achieves slightly worse results than the state-of-the-art in three out of four data sets. Importantly, we show that it is possible to successfully fine tune a large pre-trained language model with noisy data, as op-posed to previous works that rely on gold data for fine tuning.

Deep learning, in general, and natural language processing methods, in particular, rely heavily on annotated samples to achieve good performance. However, manually annotating data is expensive and time consuming. Active Learning (AL) strategies reduce the need for huge volumes of labelled data by iteratively selecting a small number of examples for manual annotation based on their estimated utility in training the given model. In this paper, we argue that since AL strategies choose examples independently, they may potentially select similar examples, all of which do not aid in the learning process. We propose a method, referred to as Active$\mathbf{^2}$ Learning (A$\mathbf{^2}$L), that actively adapts to the sequence tagging model being trained, to further eliminate such redundant examples chosen by an AL strategy. We empirically demonstrate that A$\mathbf{^2}$L improves the performance of state-of-the-art AL strategies on different sequence tagging tasks. Furthermore, we show that A$\mathbf{^2}$L is widely applicable by using it in conjunction with different AL strategies and sequence tagging models. We demonstrate that the proposed A$\mathbf{^2}$L able to reach full data F-score with $\approx\mathbf{2-16 \%}$ less data compared to state-of-art AL strategies on different sequence tagging datasets.