Understanding and extracting of information from large documents, such as business opportunities, academic articles, medical documents and technical reports, poses challenges not present in short documents. Such large documents may be multi-themed, complex, noisy and cover diverse topics. We describe a framework that can analyze large documents and help people and computer systems locate desired information in them. We aim to automatically identify and classify different sections of documents and understand their purpose within the document. A key contribution of our research is modeling and extracting the logical and semantic structure of electronic documents using deep learning techniques. We evaluate the effectiveness and robustness of our framework through extensive experiments on two collections: more than one million scholarly articles from arXiv and a collection of requests for proposal documents from government sources.
Experimental protocols at synchrotron light sources typically process and validate data only after an experiment has completed, which can lead to undetected errors and cannot enable online steering. Real-time data analysis can enable both detection of, and recovery from, errors, and optimization of data acquisition. However, modern scientific instruments, such as detectors at synchrotron light sources, can generate data at GBs/sec rates. Data processing methods such as the widely used computational tomography usually require considerable computational resources, and yield poor quality reconstructions in the early stages of data acquisition when available views are sparse. We describe here how a deep convolutional neural network can be integrated into the real-time streaming tomography pipeline to enable better-quality images in the early stages of data acquisition. Compared with conventional streaming tomography processing, our method can significantly improve tomography image quality, deliver comparable images using only 32% of the data needed for conventional streaming processing, and save 68% experiment time for data acquisition.
Architecture, size, and shape of glands are most important patterns used by pathologists for assessment of cancer malignancy in prostate histopathological tissue slides. Varying structures of glands along with cumbersome manual observations may result in subjective and inconsistent assessment. Cribriform gland with irregular border is an important feature in Gleason pattern 4. We propose using deep neural networks for cribriform pattern classification in prostate histopathological images. $163708$ Hematoxylin and Eosin (H\&E) stained images were extracted from histopathologic tissue slides of $19$ patients with prostate cancer and annotated for cribriform patterns. Our automated image classification system analyses the H\&E images to classify them as either `Cribriform' or `Non-cribriform'. Our system uses various deep learning approaches and hand-crafted image pixel intensity-based features. We present our results for cribriform pattern detection across various parameters and configuration allowed by our system. The combination of fine-tuned deep learning models outperformed the state-of-art nuclei feature based methods. Our image classification system achieved the testing accuracy of $85.93~\pm~7.54$ (cross-validated) and $88.04~\pm~5.63$ ( additional unseen test set) across three folds. In this paper, we present an annotated cribriform dataset along with analysis of deep learning models and hand-crafted features for cribriform pattern detection in prostate histopathological images.
Traditional neuroimage analysis pipelines involve computationally intensive, time-consuming optimization steps, and thus, do not scale well to large cohort studies with thousands or tens of thousands of individuals. In this work we propose a fast and accurate deep learning based neuroimaging pipeline for the automated processing of structural human brain MRI scans, replicating FreeSurfer's anatomical segmentation including surface reconstruction and cortical parcellation. To this end, we introduce an advanced deep learning architecture capable of whole brain segmentation into 95 classes. The network architecture incorporates local and global competition via competitive dense blocks and competitive skip pathways, as well as multi-slice information aggregation that specifically tailor network performance towards accurate segmentation of both cortical and sub-cortical structures. Further, we perform fast cortical surface reconstruction and thickness analysis by introducing a spectral spherical embedding and by directly mapping the cortical labels from the image to the surface. This approach provides a full FreeSurfer alternative for volumetric analysis (in under 1 minute) and surface-based thickness analysis (within only around 1h runtime). For sustainability of this approach we perform extensive validation: we assert high segmentation accuracy on several unseen datasets, measure generalizability and demonstrate increased test-retest reliability, and high sensitivity to group differences in dementia.
While many deep learning (DL)-based networking systems have demonstrated superior performance, the underlying Deep Neural Networks (DNNs) remain blackboxes and stay uninterpretable for network operators. The lack of interpretability makes DL-based networking systems prohibitive to deploy in practice. In this paper, we propose Metis, a framework that provides interpretability for two general categories of networking problems spanning local and global control. Accordingly, Metis introduces two different interpretation methods based on decision tree and hypergraph, where it converts DNN policies to interpretable rule-based controllers and highlight critical components based on analysis over hypergraph. We evaluate Metis over several state-of-the-art DL-based networking systems and show that Metis provides human-readable interpretations while preserving nearly no degradation in performance. We further present four concrete use cases of Metis, showcasing how Metis helps network operators to design, debug, deploy, and ad-hoc adjust DL-based networking systems.
The evolution of the information and communication technologies has dramatically increased the number of people with access to the Internet, which has changed the way the information is consumed. As a consequence of the above, fake news have become one of the major concerns because its potential to destabilize governments, which makes them a potential danger to modern society. An example of this can be found in the US. electoral campaign, where the term "fake news" gained great notoriety due to the influence of the hoaxes in the final result of these. In this work the feasibility of applying deep learning techniques to discriminate fake news on the Internet using only their text is studied. In order to accomplish that, three different neural network architectures are proposed, one of them based on BERT, a modern language model created by Google which achieves state-of-the-art results.
In many real-world applications of Machine Learning it is of paramount importance not only to provide accurate predictions, but also to ensure certain levels of robustness. Adversarial Training is a training procedure aiming at providing models that are robust to worst-case perturbations around predefined points. Unfortunately, one of the main issues in adversarial training is that robustness w.r.t. gradient-based attackers is always achieved at the cost of prediction accuracy. In this paper, a new algorithm, called Wasserstein Projected Gradient Descent (WPGD), for adversarial training is proposed. WPGD provides a simple way to obtain cost-sensitive robustness, resulting in a finer control of the robustness-accuracy trade-off. Moreover, WPGD solves an optimal transport problem on the output space of the network and it can efficiently discover directions where robustness is required, allowing to control the directional trade-off between accuracy and robustness. The proposed WPGD is validated in this work on image recognition tasks with different benchmark datasets and architectures. Moreover, real world-like datasets are often unbalanced: this paper shows that when dealing with such type of datasets, the performance of adversarial training are mainly affected in term of standard accuracy.
Bayesian deep learning is recently regarded as an intrinsic way to characterize the weight uncertainty of deep neural networks~(DNNs). Stochastic Gradient Langevin Dynamics~(SGLD) is an effective method to enable Bayesian deep learning on large-scale datasets. Previous theoretical studies have shown various appealing properties of SGLD, ranging from the convergence properties to the generalization bounds. In this paper, we study the properties of SGLD from a novel perspective of membership privacy protection (i.e., preventing the membership attack). The membership attack, which aims to determine whether a specific sample is used for training a given DNN model, has emerged as a common threat against deep learning algorithms. To this end, we build a theoretical framework to analyze the information leakage (w.r.t. the training dataset) of a model trained using SGLD. Based on this framework, we demonstrate that SGLD can prevent the information leakage of the training dataset to a certain extent. Moreover, our theoretical analysis can be naturally extended to other types of Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC) methods. Empirical results on different datasets and models verify our theoretical findings and suggest that the SGLD algorithm can not only reduce the information leakage but also improve the generalization ability of the DNN models in real-world applications.
Witnessed the development of deep learning, increasing number of studies try to build computer aided diagnosis systems for 3D volumetric medical data. However, as the annotations of 3D medical data are difficult to acquire, the number of annotated 3D medical images is often not enough to well train the deep learning networks. The self-supervised learning deeply exploiting the information of raw data is one of the potential solutions to loose the requirement of training data. In this paper, we propose a self-supervised learning framework for the volumetric medical images. A novel proxy task, i.e., Rubik's cube recovery, is formulated to pre-train 3D neural networks. The proxy task involves two operations, i.e., cube rearrangement and cube rotation, which enforce networks to learn translational and rotational invariant features from raw 3D data. Compared to the train-from-scratch strategy, fine-tuning from the pre-trained network leads to a better accuracy on various tasks, e.g., brain hemorrhage classification and brain tumor segmentation. We show that our self-supervised learning approach can substantially boost the accuracies of 3D deep learning networks on the volumetric medical datasets without using extra data. To our best knowledge, this is the first work focusing on the self-supervised learning of 3D neural networks.
Fully automatic deep learning has become the state-of-the-art technique for many tasks including image acquisition, analysis and interpretation, and for the extraction of clinically useful information for computer-aided detection, diagnosis, treatment planning, intervention and therapy. However, the unique challenges posed by medical image analysis suggest that retaining a human end user in any deep learning enabled system will be beneficial. In this review we investigate the role that humans might play in the development and deployment of deep learning enabled diagnostic applications and focus on techniques that will retain a significant input from a human end user. Human-in-the-Loop computing is an area that we see as increasingly important in future research due to the safety-critical nature of working in the medical domain. We evaluate four key areas that we consider vital for deep learning in the clinical practice: (1) Active Learning to choose the best data to annotate for optimal model performance; (2) Interaction with model outputs - using iterative feedback to steer models to optima for a given prediction and offering meaningful ways to interpret and respond to predictions; (3) Practical considerations - developing full scale applications and the key considerations that need to be made before deployment; (4) Future Prospective and Unanswered Questions - knowledge gaps and related research fields that will benefit human-in-the-loop computing as they evolve. We offer our opinions on the most promising directions of research and how various aspects of each area might be unified towards common goals.
We investigate a deep learning approach to efficiently perform Bayesian inference in partial differential equation (PDE) and integral equation models over potentially high-dimensional parameter spaces. The contributions of this paper are two-fold; the first is the introduction of a neural network approach to approximating the solutions of Fredholm and Volterra integral equations of the first and second kind. The second is the development of a new, efficient deep learning-based method for Bayesian inversion applied to problems that can be described by PDEs or integral equations. To achieve this we introduce a surrogate model, and demonstrate how this allows efficient sampling from a Bayesian posterior distribution in which the likelihood depends on the solutions of PDEs or integral equations. Our method relies on the direct approximation of parametric solutions by neural networks, without need of traditional numerical solves. This deep learning approach allows the accurate and efficient approximation of parametric solutions in significantly higher dimensions than is possible using classical discretisation schemes. Since the approximated solutions can be cheaply evaluated, the solutions of Bayesian inverse problems over large parameter spaces are efficient using Markov chain Monte Carlo. We demonstrate the performance of our method using two real-world examples; these include Bayesian inference in the PDE and integral equation case for an example from electrochemistry, and Bayesian inference of a function-valued heat-transfer parameter with applications in aviation.
This article is an overview of supervised machine learning problems for regression and classification. Topics include: kernel methods, training by stochastic gradient descent, deep learning architecture, losses for classification, statistical learning theory, and dimension independent generalization bounds. Implicit regularization in deep learning examples are presented, including data augmentation, adversarial training, and additive noise. These methods are reframed as explicit gradient regularization.
In the last few years, deep learning has solved seemingly intractable problems, boosting the hope to find approximate solutions to problems that now are considered unsolvable. Earthquake prediction, the Grail of Seismology, is, in this context of continuous exciting discoveries, an obvious choice for deep learning exploration. We review the entire literature of artificial neural network (ANN) applications for earthquake prediction (77 articles, 1994-2019 period) and find two emerging trends: an increasing interest in this domain, and a complexification of ANN models over time, towards deep learning. Despite apparent positive results observed in this corpus, we demonstrate that simpler models seem to offer similar predictive powers, if not better ones. Due to the structured, tabulated nature of earthquake catalogues, and the limited number of features so far considered, simpler and more transparent machine learning models seem preferable at the present stage of research. Those baseline models follow first physical principles and are consistent with the known empirical laws of Statistical Seismology, which have minimal abilities to predict large earthquakes.
Magnetic resonance imaging has been widely applied in clinical diagnosis, however, is limited by its long data acquisition time. Although imaging can be accelerated by sparse sampling and parallel imaging, achieving promising reconstruction images with a fast reconstruction speed remains a challenge. Recently, deep learning approaches have attracted a lot of attention for its encouraging reconstruction results but without a proper interpretability. In this letter, to enable high-quality image reconstruction for the parallel magnetic resonance imaging, we design the network structure from the perspective of sparse iterative reconstruction and enhance it with the residual structure. The experimental results of a public knee dataset show that compared with the optimization-based method and the latest deep learning parallel imaging methods, the proposed network has less error in reconstruction and is more stable under different acceleration factors.
High spatiotemporal resolution dynamic magnetic resonance imaging (MRI) is a powerful clinical tool for imaging moving structures as well as to reveal and quantify other physical and physiological dynamics. The low speed of MRI necessitates acceleration methods such as deep learning reconstruction from under-sampled data. However, the massive size of many dynamic MRI problems prevents deep learning networks from directly exploiting global temporal relationships. In this work, we show that by applying deep neural networks inside a priori calculated temporal feature spaces, we enable deep learning reconstruction with global temporal modeling even for image sequences with >40,000 frames. One proposed variation of our approach using dilated multi-level Densely Connected Network (mDCN) speeds up feature space coordinate calculation by 3000x compared to conventional iterative methods, from 20 minutes to 0.39 seconds. Thus, the combination of low-rank tensor and deep learning models not only makes large-scale dynamic MRI feasible but also practical for routine clinical application.
Modeling of turbulent flows is still challenging. One way to deal with the large scale separation due to turbulence is to simulate only the large scales and model the unresolved contributions as done in large-eddy simulation (LES). This paper focuses on two deep learning (DL) strategies, regression and reconstruction, which are data-driven and promising alternatives to classical modeling concepts. Using three-dimensional (3-D) forced turbulence direct numerical simulation (DNS) data, subgrid models are evaluated, which predict the unresolved part of quantities based on the resolved solution. For regression, it is shown that feedforward artificial neural networks (ANNs) are able to predict the fully-resolved scalar dissipation rate using filtered input data. It was found that a combination of a large-scale quantity, such as the filtered passive scalar itself, and a small-scale quantity, such as the filtered energy dissipation rate, gives the best agreement with the actual DNS data. Furthermore, a DL network motivated by enhanced super-resolution generative adversarial networks (ESRGANs) was used to reconstruct fully-resolved 3-D velocity fields from filtered velocity fields. The energy spectrum shows very good agreement. As size of scientific data is often in the order of terabytes or more, DL needs to be combined with high performance computing (HPC). Necessary code improvements for HPC-DL are discussed with respect to the supercomputer JURECA. After optimizing the training code, 396.2 TFLOPS were achieved.
This paper introduces Selective-Backprop, a technique that accelerates the training of deep neural networks (DNNs) by prioritizing examples with high loss at each iteration. Selective-Backprop uses the output of a training example's forward pass to decide whether to use that example to compute gradients and update parameters, or to skip immediately to the next example. By reducing the number of computationally-expensive backpropagation steps performed, Selective-Backprop accelerates training. Evaluation on CIFAR10, CIFAR100, and SVHN, across a variety of modern image models, shows that Selective-Backprop converges to target error rates up to 3.5x faster than with standard SGD and between 1.02--1.8x faster than a state-of-the-art importance sampling approach. Further acceleration of 26% can be achieved by using stale forward pass results for selection, thus also skipping forward passes of low priority examples.
It is crucial today that economies harness renewable energies and integrate them into the existing grid. Conventionally, energy has been generated based on forecasts of peak and low demands. Renewable energy can neither be produced on demand nor stored efficiently. Thus, the aim of this paper is to evaluate Deep Learning-based forecasts of energy consumption to align energy consumption with renewable energy production. Using a dataset from a use-case related to landfill leachate management, multiple prediction models were used to forecast energy demand.The results were validated based on the same dataset from the recycling industry. Shallow models showed the lowest Mean Absolute Percentage Error (MAPE), significantly outperforming a persistence baseline for both, long-term (30 days), mid-term (7 days) and short-term (1 day) forecasts. A potential decrease of up to 23% in peak energy demand was found that could lead to a reduction of 3,091 kg in CO2-emissions per year. Our approach requires low finanacial investments for energy-management hardware, making it suitable for usage in Small and Medium sized Enterprises (SMEs).
We propose a novel method to train deep convolutional neural networks which learn from multiple data sets of varying input sizes through weight sharing. This is an advantage in chemometrics where individual measurements represent exact chemical compounds and thus signals cannot be translated or resized without disturbing their interpretation. Our approach show superior performance compared to transfer learning when a medium sized and a small data set are trained together. While we observe a small improvement compared to individual training when two medium sized data sets are trained together, in particular through a reduction in the variance.
We empirically evaluate common assumptions about neural networks that are widely held by practitioners and theorists alike. In this work, we: (1) prove the widespread existence of suboptimal local minima in the loss landscape of neural networks, and we use our theory to find examples; (2) show that small-norm parameters are not optimal for generalization; (3) demonstrate that ResNets do not conform to wide-network theories, such as the neural tangent kernel, and that the interaction between skip connections and batch normalization plays a role; (4) find that rank does not correlate with generalization or robustness in a practical setting.