Despite the empirical success of using Adversarial Training to defend deep learning models against adversarial perturbations, so far, it still remains rather unclear what the principles are behind the existence of adversarial perturbations, and what adversarial training does to the neural network to remove them. In this paper, we present a principle that we call Feature Purification, where we show one of the causes of the existence of adversarial examples is the accumulation of certain small dense mixtures in the hidden weights during the training process of a neural network; and more importantly, one of the goals of adversarial training is to remove such mixtures to purify hidden weights. We present both experiments on the CIFAR-10 dataset to illustrate this principle, and a theoretical result proving that for certain natural classification tasks, training a two-layer neural network with ReLU activation using randomly initialized gradient descent indeed satisfies this principle. Technically, we give, to the best of our knowledge, the first result proving that the following two can hold simultaneously for training a neural network with ReLU activation. (1) Training over the original data is indeed non-robust to small adversarial perturbations of some radius. (2) Adversarial training, even with an empirical perturbation algorithm such as FGM, can in fact be provably robust against ANY perturbations of the same radius. Finally, we also prove a complexity lower bound, showing that low complexity models such as linear classifiers, low-degree polynomials, or even the neural tangent kernel for this network, CANNOT defend against perturbations of this same radius, no matter what algorithms are used to train them.
Image reconstruction from insufficient data is common in computed tomography (CT), e.g., image reconstruction from truncated data, limited-angle data and sparse-view data. Deep learning has achieved impressive results in this field. However, the robustness of deep learning methods is still a concern for clinical applications due to the following two challenges: a) With limited access to sufficient training data, a learned deep learning model may not generalize well to unseen data; b) Deep learning models are sensitive to noise. Therefore, the quality of images processed by neural networks only may be inadequate. In this work, we investigate the robustness of deep learning in CT image reconstruction by showing false negative and false positive lesion cases. Since learning-based images with incorrect structures are likely not consistent with measured projection data, we propose a data consistent reconstruction (DCR) method to improve their image quality, which combines the advantages of compressed sensing and deep learning: First, a prior image is generated by deep learning. Afterwards, unmeasured projection data are inpainted by forward projection of the prior image. Finally, iterative reconstruction with reweighted total variation regularization is applied, integrating data consistency for measured data and learned prior information for missing data. The efficacy of the proposed method is demonstrated in cone-beam CT with truncated data, limited-angle data and sparse-view data, respectively. For example, for truncated data, DCR achieves a mean root-mean-square error of 24 HU and a mean structure similarity index of 0.999 inside the field-of-view for different patients in the noisy case, while the state-of-the-art U-Net method achieves 55 HU and 0.995 respectively for these two metrics.
Estimating the forces acting between instruments and tissue is a challenging problem for robot-assisted minimally-invasive surgery. Recently, numerous vision-based methods have been proposed to replace electro-mechanical approaches. Moreover, optical coherence tomography (OCT) and deep learning have been used for estimating forces based on deformation observed in volumetric image data. The method demonstrated the advantage of deep learning with 3D volumetric data over 2D depth images for force estimation. In this work, we extend the problem of deep learning-based force estimation to 4D spatio-temporal data with streams of 3D OCT volumes. For this purpose, we design and evaluate several methods extending spatio-temporal deep learning to 4D which is largely unexplored so far. Furthermore, we provide an in-depth analysis of multi-dimensional image data representations for force estimation, comparing our 4D approach to previous, lower-dimensional methods. Also, we analyze the effect of temporal information and we study the prediction of short-term future force values, which could facilitate safety features. For our 4D force estimation architectures, we find that efficient decoupling of spatial and temporal processing is advantageous. We show that using 4D spatio-temporal data outperforms all previously used data representations with a mean absolute error of 10.7mN. We find that temporal information is valuable for force estimation and we demonstrate the feasibility of force prediction.
We propose a deep learning based method for classification of commonly occurring pathologies in chest X-ray images. The vast number of publicly available chest X-ray images provides the data necessary for successfully employing deep learning methodologies to reduce the misdiagnosis of thoracic diseases. We applied our method to the classification of two example pathologies, pulmonary nodules and cardiomegaly, and we compared the performance of our method to three existing methods. The results show an improvement in AUC for detection of nodules and cardiomegaly compared to the existing methods.
Extracting valuable information from large sets of diverse meteorological data is a time-intensive process. Machine learning methods can help improve both speed and accuracy of this process. Specifically, deep learning image segmentation models using the U-Net structure perform faster and can identify areas missed by more restrictive approaches, such as expert hand-labeling and a priori heuristic methods. This paper discusses four different state-of-the-art U-Net models designed for detection of tropical and extratropical cyclone Regions Of Interest (ROI) from two separate input sources: total precipitable water output from the Global Forecasting System (GFS) model and water vapor radiance images from the Geostationary Operational Environmental Satellite (GOES). These models are referred to as IBTrACS-GFS, Heuristic-GFS, IBTrACS-GOES, and Heuristic-GOES. All four U-Nets are fast information extraction tools and perform with a ROI detection accuracy ranging from 80% to 99%. These are additionally evaluated with the Dice and Tversky Intersection over Union (IoU) metrics, having Dice coefficient scores ranging from 0.51 to 0.76 and Tversky coefficients ranging from 0.56 to 0.74. The extratropical cyclone U-Net model performed 3 times faster than the comparable heuristic model used to detect the same ROI. The U-Nets were specifically selected for their capabilities in detecting cyclone ROI beyond the scope of the training labels. These machine learning models identified more ambiguous and active ROI missed by the heuristic model and hand-labeling methods commonly used in generating real-time weather alerts, having a potentially direct impact on public safety.
Tensor networks (TN) have found a wide use in machine learning, and in particular, TN and deep learning bear striking similarities. In this work, we propose the quantum-classical hybrid tensor networks (HTN) which combine tensor networks with classical neural networks in a uniform deep learning framework to overcome the limitations of regular tensor networks in machine learning. We first analyze the limitations of regular tensor networks in the applications of machine learning involving the representation power and architecture scalability. We conclude that in fact the regular tensor networks are not competent to be the basic building blocks of deep learning. Then, we discuss the performance of HTN which overcome all the deficiency of regular tensor networks for machine learning. In this sense, we are able to train HTN in the deep learning way which is the standard combination of algorithms such as Back Propagation and Stochastic Gradient Descent. We finally provide two applicable cases to show the potential applications of HTN, including quantum states classification and quantum-classical autoencoder. These cases also demonstrate the great potentiality to design various HTN in deep learning way.
Accurate mortality prediction allows Intensive Care Units (ICUs) to adequately benchmark clinical practice and identify patients with unexpected outcomes. Traditionally, simple statistical models have been used to assess patient death risk, many times with sub-optimal performance. On the other hand deep learning holds promise to positively impact clinical practice by leveraging medical data to assist diagnosis and prediction, including mortality prediction. However, as the question of whether powerful Deep Learning models attend correlations backed by sound medical knowledge when generating predictions remains open, additional interpretability tools are needed to foster trust and encourage the use of AI by clinicians. In this work we show a Deep Learning model trained on MIMIC-III to predict mortality using raw nursing notes, together with visual explanations for word importance. Our model reaches a ROC of 0.8629 (+/-0.0058), outperforming the traditional SAPS-II score and providing enhanced interpretability when compared with similar Deep Learning approaches.
The optimal design of compounds through manipulating properties at the molecular level is often the key to considerable scientific advances and improved process systems performance. This paper highlights key trends, challenges, and opportunities underpinning the Computer-Aided Molecular Design (CAMD) problems. A brief review of knowledge-driven property estimation methods and solution techniques, as well as corresponding CAMD tools and applications, are first presented. In view of the computational challenges plaguing knowledge-based methods and techniques, we survey the current state-of-the-art applications of deep learning to molecular design as a fertile approach towards overcoming computational limitations and navigating uncharted territories of the chemical space. The main focus of the survey is given to deep generative modeling of molecules under various deep learning architectures and different molecular representations. Further, the importance of benchmarking and empirical rigor in building deep learning models is spotlighted. The review article also presents a detailed discussion of the current perspectives and challenges of knowledge-based and data-driven CAMD and identifies key areas for future research directions. Special emphasis is on the fertile avenue of hybrid modeling paradigm, in which deep learning approaches are exploited while leveraging the accumulated wealth of knowledge-driven CAMD methods and tools.
Deep learning has led to state-of-the-art results for many medical imaging tasks, such as segmentation of different anatomical structures. With the increased numbers of deep learning publications and openly available code, the approach to choosing a model for a new task becomes more complicated, while time and (computational) resources are limited. A possible solution to choosing a model efficiently is meta-learning, a learning method in which prior performance of a model is used to predict the performance for new tasks. We investigate meta-learning for segmentation across ten datasets of different organs and modalities. We propose four ways to represent each dataset by meta-features: one based on statistical features of the images and three are based on deep learning features. We use support vector regression and deep neural networks to learn the relationship between the meta-features and prior model performance. On three external test datasets these methods give Dice scores within 0.10 of the true performance. These results demonstrate the potential of meta-learning in medical imaging.
In this paper, we introduce SciANN, a Python package for scientific computing and physics-informed deep learning using artificial neural networks. SciANN uses the widely used deep-learning packages Tensorflow and Keras to build deep neural networks and optimization models, thus inheriting many of Keras's functionalities, such as batch optimization and model reuse for transfer learning. SciANN is designed to abstract neural network construction for scientific computations and solution and discovery of partial differential equations (PDE) using the physics-informed neural networks (PINN) architecture, therefore providing the flexibility to set up complex functional forms. We illustrate, in a series of examples, how the framework can be used for curve fitting on discrete data, and for solution and discovery of PDEs in strong and weak forms. We summarize the features currently available in SciANN, and also outline ongoing and future developments.
Packet-loss is a common problem in data transmission, using Voice over IP. The problem is an old problem, and there has been a variety of classical approaches that were developed to overcome this problem. However, with the rise of deep learning and generative models like Generative Adversarial Networks and Autoencoders, a new avenue has emerged for attempting to solve packet-loss using deep learning, by generating replacements for lost packets. In this mini-survey, we review all the literature we found to date, that attempt to solve the packet-loss in speech using deep learning methods. Additionally, we briefly review how the problem of packet-loss in a realistic setting is modelled, and how to evaluate Packet Loss Concealment techniques. Moreover, we review a few modern deep learning techniques in related domains that have shown promising results. These techniques shed light on future potentially better solutions for PLC and additional challenges that need to be considered simultaneously with packet-loss.
We investigate the use of deep learning in the context of X-ray polarization detection from astrophysical sources as will be observed by the Imaging X-ray Polarimetry Explorer (IXPE), a future NASA selected space-based mission expected to be operative in 2021. In particular, we propose two models that can be used to estimate the impact point as well as the polarization direction of the incoming radiation. The results obtained show that data-driven approaches depict a promising alternative to the existing analytical approaches. We also discuss problems and challenges to be addressed in the near future.
Brain gender differences have been known for a long time and are the possible reason for many psychological, psychiatric and behavioral differences between males and females. Predicting genders from brain functional connectivity (FC) can build the relationship between brain activities and gender, and extracting important gender related FC features from the prediction model offers a way to investigate the brain gender difference. Current predictive models applied to gender prediction demonstrate good accuracies, but usually extract individual functional connections instead of connectivity patterns in the whole connectivity matrix as features. In addition, current models often omit the effect of the input brain FC scale on prediction and cannot give any model uncertainty information. Hence, in this study we propose to predict gender from multiple scales of brain FC with deep learning, which can extract full FC patterns as features. We further develop the understanding of the feature extraction mechanism in deep neural network (DNN) and propose a DNN feature ranking method to extract the highly important features based on their contributions to the prediction. Moreover, we apply Bayesian deep learning to the brain FC gender prediction, which as a probabilistic model can not only make accurate predictions but also generate model uncertainty for each prediction. Experiments were done on the high-quality Human Connectome Project S1200 release dataset comprising the resting state functional MRI data of 1003 healthy adults. First, DNN reaches 83.0%, 87.6%, 92.0%, 93.5% and 94.1% accuracies respectively with the FC input derived from 25, 50, 100, 200, 300 independent component analysis (ICA) components. DNN outperforms the conventional machine learning methods on the 25-ICA-component scale FC, but the linear machine learning method catches up as the number of ICA components increases...
As communities represent similar opinions, similar functions, similar purposes, etc., community detection is an important and extremely useful tool in both scientific inquiry and data analytics. However, the classic methods of community detection, such as spectral clustering and statistical inference, are falling by the wayside as deep learning techniques demonstrate an increasing capacity to handle high-dimensional graph data with impressive performance. Thus, a survey of current progress in community detection through deep learning is timely. Structured into three broad research streams in this domain - deep neural networks, deep graph embedding, and graph neural networks, this article summarizes the contributions of the various frameworks, models, and algorithms in each stream along with the current challenges that remain unsolved and the future research opportunities yet to be explored.
Recently, recommender systems that aim to suggest personalized lists of items for users to interact with online have drawn a lot of attention. In fact, many of these state-of-the-art techniques have been deep learning based. Recent studies have shown that these deep learning models (in particular for recommendation systems) are vulnerable to attacks, such as data poisoning, which generates users to promote a selected set of items. However, more recently, defense strategies have been developed to detect these generated users with fake profiles. Thus, advanced injection attacks of creating more `realistic' user profiles to promote a set of items is still a key challenge in the domain of deep learning based recommender systems. In this work, we present our framework CopyAttack, which is a reinforcement learning based black-box attack method that harnesses real users from a source domain by copying their profiles into the target domain with the goal of promoting a subset of items. CopyAttack is constructed to both efficiently and effectively learn policy gradient networks that first select, and then further refine/craft, user profiles from the source domain to ultimately copy into the target domain. CopyAttack's goal is to maximize the hit ratio of the targeted items in the Top-$k$ recommendation list of the users in the target domain. We have conducted experiments on two real-world datasets and have empirically verified the effectiveness of our proposed framework and furthermore performed a thorough model analysis.
Quantifying uncertainty in weather forecasts is critical, especially for predicting extreme weather events. This is typically accomplished with ensemble prediction systems, which consist of many perturbed numerical weather simulations, or trajectories, run in parallel. These systems are associated with a high computational cost and often involve statistical post-processing steps to inexpensively improve their raw prediction qualities. We propose a mixed model that uses only a subset of the original weather trajectories combined with a post-processing step using deep neural networks. These enable the model to account for non-linear relationships that are not captured by current numerical models or post-processing methods. Applied to global data, our mixed models achieve a relative improvement in ensemble forecast skill (CRPS) of over 14%. Furthermore, we demonstrate that the improvement is larger for extreme weather events on select case studies. We also show that our post-processing can use fewer trajectories to achieve comparable results to the full ensemble. By using fewer trajectories, the computational costs of an ensemble prediction system can be reduced, allowing it to run at higher resolution and produce more accurate forecasts.
Objectives: With the technological advancements in the field of tele-health monitoring, it is now possible to gather huge amounts of electro-physiological signals such as electrocardiogram (ECG). It is therefore necessary to develop models/algorithms that are capable of analysing these massive amounts of data in real-time. This paper proposes a deep learning model for real-time segmentation of heartbeats. Methods: The proposed algorithm, named as the DENS-ECG algorithm, combines convolutional neural network (CNN) and long short-term memory (LSTM) model to detect onset, peak, and offset of different heartbeat waveforms such as the P-wave, QRS complex, T-wave, and No wave (NW). Using ECG as the inputs, the model learns to extract high level features through the training process, which, unlike other classical machine learning based methods, eliminates the feature engineering step. Results: The proposed DENS-ECG model was trained and validated on a dataset with 105 ECGs of length 15 minutes each and achieved an average sensitivity and precision of 97.95% and 95.68%, respectively, using a 5-fold cross validation. Additionally, the model was evaluated on an unseen dataset to examine its robustness in QRS detection, which resulted in a sensitivity of 99.61% and precision of 99.52%. Conclusion: The empirical results show the flexibility and accuracy of the combined CNN-LSTM model for ECG signal delineation. Significance: This paper proposes an efficient and easy to use approach using deep learning for heartbeat segmentation, which could potentially be used in real-time tele-health monitoring systems.
Deep learning-based methods have pushed the limits of the state-of-the-art in face analysis. However, despite their success, these models have raised concerns regarding their bias towards certain demographics. This bias is inflicted both by limited diversity across demographics in the training set, as well as the design of the algorithms. In this work, we investigate the demographic bias of deep learning models in face recognition, age estimation, gender recognition and kinship verification. To this end, we introduce the most comprehensive, large-scale dataset of facial images and videos to date. It consists of 40K still images and 44K sequences (14.5M video frames in total) captured in unconstrained, real-world conditions from 1,045 subjects. The data are manually annotated in terms of identity, exact age, gender and kinship. The performance of state-of-the-art models is scrutinized and demographic bias is exposed by conducting a series of experiments. Lastly, a method to debias network embeddings is introduced and tested on the proposed benchmarks.
A wide range of applications in marine ecology extensively uses underwater cameras. Still, to efficiently process the vast amount of data generated, we need to develop tools that can automatically detect and recognize species captured on film. Classifying fish species from videos and images in natural environments can be challenging because of noise and variation in illumination and the surrounding habitat. In this paper, we propose a two-step deep learning approach for the detection and classification of temperate fishes without pre-filtering. The first step is to detect each single fish in an image, independent of species and sex. For this purpose, we employ the You Only Look Once (YOLO) object detection technique. In the second step, we adopt a Convolutional Neural Network (CNN) with the Squeeze-and-Excitation (SE) architecture for classifying each fish in the image without pre-filtering. We apply transfer learning to overcome the limited training samples of temperate fishes and to improve the accuracy of the classification. This is done by training the object detection model with ImageNet and the fish classifier via a public dataset (Fish4Knowledge), whereupon both the object detection and classifier are updated with temperate fishes of interest. The weights obtained from pre-training are applied to post-training as a priori. Our solution achieves the state-of-the-art accuracy of 99.27\% on the pre-training. The percentage values for accuracy on the post-training are good; 83.68\% and 87.74\% with and without image augmentation, respectively, indicating that the solution is viable with a more extensive dataset.
The ability to detect anomalies in time series is considered highly valuable in numerous application domains. The sequential nature of time series objects is responsible for an additional feature complexity, ultimately requiring specialized approaches in order to solve the task. Essential characteristics of time series, situated outside the time domain, are often difficult to capture with state-of-the-art anomaly detection methods when no transformations have been applied to the time series. Inspired by the success of deep learning methods in computer vision, several studies have proposed transforming time series into image-like representations, used as inputs for deep learning models, and have led to very promising results in classification tasks. In this paper, we first review the signal to image encoding approaches found in the literature. Second, we propose modifications to some of their original formulations to make them more robust to the variability in large datasets. Third, we compare them on the basis of a common unsupervised task to demonstrate how the choice of the encoding can impact the results when used in the same deep learning architecture. We thus provide a comparison between six encoding algorithms with and without the proposed modifications. The selected encoding methods are Gramian Angular Field, Markov Transition Field, recurrence plot, grey scale encoding, spectrogram, and scalogram. We also compare the results achieved with the raw signal used as input for another deep learning model. We demonstrate that some encodings have a competitive advantage and might be worth considering within a deep learning framework. The comparison is performed on a dataset collected and released by Airbus SAS, containing highly complex vibration measurements from real helicopter flight tests. The different encodings provide competitive results for anomaly detection.