Deep unfolding networks have recently gained popularity in the context of solving imaging inverse problems. However, the computational and memory complexity of data-consistency layers within traditional deep unfolding networks scales with the number of measurements, limiting their applicability to large-scale imaging inverse problems. We propose SGD-Net as a new methodology for improving the efficiency of deep unfolding through stochastic approximations of the data-consistency layers. Our theoretical analysis shows that SGD-Net can be trained to approximate batch deep unfolding networks to an arbitrary precision. Our numerical results on intensity diffraction tomography and sparse-view computed tomography show that SGD-Net can match the performance of the batch network at a fraction of training and testing complexity.
Predicting the 3D structure of proteins is one of the most important problems in current biochemical research. In this article, we explain how to combine recent deep learning advances with the well known technique of quantum walks applied to a Metropolis algorithm. The result, QFold, is a fully scalable hybrid quantum algorithm that, in contrast to previous quantum approaches, does not require a lattice model simplification and instead relies on the much more realistic assumption of parameterization in terms of torsion angles of the amino acids. We compare it with its classical analog for different annealing schedules and find a polynomial quantum advantage, and implement a minimal realization of the quantum Metropolis in the IBMQ Casablanca quantum system.
Activation functions shape the outputs of artificial neurons and, therefore, are integral parts of neural networks in general and deep learning in particular. Some activation functions, such as logistic and relu, have been used for many decades. But with deep learning becoming a mainstream research topic, new activation functions have mushroomed, leading to confusion in both theory and practice. This paper provides an analytic yet up-to-date overview of popular activation functions and their properties, which makes it a timely resource for anyone who studies or applies neural networks.
Which samples should be labelled in a large data set is one of the most important problems for trainingof deep learning. So far, a variety of active sample selection strategies related to deep learning havebeen proposed in many literatures. We defined them as Active Deep Learning (ADL) only if theirpredictor is deep model, where the basic learner is called as predictor and the labeling schemes iscalled selector. In this survey, three fundamental factors in selector designation were summarized. Wecategory ADL into model-driven ADL and data-driven ADL, by whether its selector is model-drivenor data-driven. The different characteristics of the two major type of ADL were addressed in indetail respectively. Furthermore, different sub-classes of data-driven and model-driven ADL are alsosummarized and discussed emphatically. The advantages and disadvantages between data-driven ADLand model-driven ADL are thoroughly analyzed. We pointed out that, with the development of deeplearning, the selector in ADL also is experiencing the stage from model-driven to data-driven. Finally,we make discussion on ADL about its uncertainty, explanatory, foundations of cognitive science etc.and survey on the trend of ADL from model-driven to data-driven.
Visual surveillance aims to stably detect a foreground object using a continuous image acquired from a fixed camera. Recent deep learning methods based on supervised learning show superior performance compared to classical background subtraction algorithms. However, there is still a room for improvement in static foreground, dynamic background, hard shadow, illumination changes, camouflage, etc. In addition, most of the deep learning-based methods operates well on environments similar to training. If the testing environments are different from training ones, their performance degrades. As a result, additional training on those operating environments is required to ensure a good performance. Our previous work which uses spatio-temporal input data consisted of a number of past images, background images and current image showed promising results in different environments from training, although it uses a simple U-NET structure. In this paper, we propose a data augmentation technique suitable for visual surveillance for additional performance improvement using the same network used in our previous work. In deep learning, most data augmentation techniques deal with spatial-level data augmentation techniques for use in image classification and object detection. In this paper, we propose a new method of data augmentation in the spatio-temporal dimension suitable for our previous work. Two data augmentation methods of adjusting background model images and past images are proposed. Through this, it is shown that performance can be improved in difficult areas such as static foreground and ghost objects, compared to previous studies. Through quantitative and qualitative evaluation using SBI, LASIESTA, and our own dataset, we show that it gives superior performance compared to deep learning-based algorithms and background subtraction algorithms.
The use of fundus images for the early screening of eye diseases is of great clinical importance. Due to its powerful performance, deep learning is becoming more and more popular in related applications, such as lesion segmentation, biomarkers segmentation, disease diagnosis and image synthesis. Therefore, it is very necessary to summarize the recent developments in deep learning for fundus images with a review paper. In this review, we introduce 143 application papers with a carefully designed hierarchy. Moreover, 33 publicly available datasets are presented. Summaries and analyses are provided for each task. Finally, limitations common to all tasks are revealed and possible solutions are given. We will also release and regularly update the state-of-the-art results and newly-released datasets at https://github.com/nkicsl/Fundus Review to adapt to the rapid development of this field.
We study the generalization of deep learning models in relation to the convex hull of their training sets. A trained image classifier basically partitions its domain via decision boundaries and assigns a class to each of those partitions. The location of decision boundaries inside the convex hull of training set can be investigated in relation to the training samples. However, our analysis shows that in standard image classification datasets, all testing images are considerably outside that convex hull, in the pixel space, in the wavelet space, and in the internal representations learned by deep networks. Therefore, the performance of a trained model partially depends on how its decision boundaries are extended outside the convex hull of its training data. From this perspective which is not studied before, over-parameterization of deep learning models may be considered a necessity for shaping the extension of decision boundaries. At the same time, over-parameterization should be accompanied by a specific training regime, in order to yield a model that not only fits the training set, but also its decision boundaries extend desirably outside the convex hull. To illustrate this, we investigate the decision boundaries of a neural network, with various degrees of parameters, inside and outside the convex hull of its training set. Moreover, we use a polynomial decision boundary to study the necessity of over-parameterization and the influence of training regime in shaping its extensions outside the convex hull of training set.
Computer vision and image processing address many challenging applications. While the last decade has seen deep neural network architectures revolutionizing those fields, early methods relied on 'classic', i.e., non-learned approaches. In this study, we explore the differences between classic and deep learning (DL) algorithms to gain new insight regarding which is more suitable for a given application. The focus is on two challenging ill-posed problems, namely faint edge detection and multispectral image registration, studying recent state-of-the-art DL and classic solutions. While those DL algorithms outperform classic methods in terms of accuracy and development time, they tend to have higher resource requirements and are unable to perform outside their training space. Moreover, classic algorithms are more transparent, which facilitates their adoption for real-life applications. As both classes of approaches have unique strengths and limitations, the choice of a solution is clearly application dependent.
Ground filtering has remained a widely studied but incompletely resolved bottleneck for decades in the automatic generation of high-precision digital elevation model, due to the dramatic changes of topography and the complex structures of objects. The recent breakthrough of supervised deep learning algorithms in 3D scene understanding brings new solutions for better solving such problems. However, there are few large-scale and scene-rich public datasets dedicated to ground extraction, which considerably limits the development of effective deep-learning-based ground filtering methods. To this end, we present OpenGF, first Ultra-Large-Scale Ground Filtering dataset covering over 47 $km^2$ of 9 different typical terrain scenes built upon open ALS point clouds of 4 different countries around the world. OpenGF contains more than half a billion finely labeled ground and non-ground points, thousands of times the number of labeled points than the de facto standard ISPRS filtertest dataset. We extensively evaluate the performance of state-of-the-art rule-based algorithms and 3D semantic segmentation networks on our dataset and provide a comprehensive analysis. The results have confirmed the capability of OpenGF to train deep learning models effectively. This dataset is released at https://github.com/Nathan-UW/OpenGF to promote more advancing research for ground filtering and large-scale 3D geographic environment understanding.
Increasing numbers of patients with disabilities or elderly people with mobility issues often suffer from a pressure ulcer. The affected areas need regular checks, but they have a difficulty in accessing a hospital. Some remote diagnosis systems are being used for them, but there are limitations in checking a patient's status regularly. In this paper, we present a remote medical assistant that can help pressure ulcer management with image processing techniques. The proposed system includes a mobile application with a deep learning model for wound segmentation and analysis. As there are not enough data to train the deep learning model, we make use of a pretrained model from a relevant domain and data augmentation that is appropriate for this task. First of all, an image preprocessing method using bilinear interpolation is used to resize images and normalize the images. Second, for data augmentation, we use rotation, reflection, and a watershed algorithm. Third, we use a pretrained deep learning model generated from skin wound images similar to pressure ulcer images. Finally, we added an attention module that can provide hints on the pressure ulcer image features. The resulting model provides an accuracy of 99.0%, an intersection over union (IoU) of 99.99%, and a dice similarity coefficient (DSC) of 93.4% for pressure ulcer segmentation, which is better than existing results.
Deep learning techniques for 3D brain vessel image segmentation have not been as successful as in the segmentation of other organs and tissues. This can be explained by two factors. First, deep learning techniques tend to show poor performances at the segmentation of relatively small objects compared to the size of the full image. Second, due to the complexity of vascular trees and the small size of vessels, it is challenging to obtain the amount of annotated training data typically needed by deep learning methods. To address these problems, we propose a novel annotation-efficient deep learning vessel segmentation framework. The framework avoids pixel-wise annotations, only requiring weak patch-level labels to discriminate between vessel and non-vessel 2D patches in the training set, in a setup similar to the CAPTCHAs used to differentiate humans from bots in web applications. The user-provided weak annotations are used for two tasks: 1) to synthesize pixel-wise pseudo-labels for vessels and background in each patch, which are used to train a segmentation network, and 2) to train a classifier network. The classifier network allows to generate additional weak patch labels, further reducing the annotation burden, and it acts as a noise filter for poor quality images. We use this framework for the segmentation of the cerebrovascular tree in Time-of-Flight angiography (TOF) and Susceptibility-Weighted Images (SWI). The results show that the framework achieves state-of-the-art accuracy, while reducing the annotation time by ~77% w.r.t. learning-based segmentation methods using pixel-wise labels for training.
Irregular sampling occurs in many time series modeling applications where it presents a significant challenge to standard deep learning models. This work is motivated by the analysis of physiological time series data in electronic health records, which are sparse, irregularly sampled, and multivariate. In this paper, we propose a new deep learning framework for this setting that we call Multi-Time Attention Networks. Multi-Time Attention Networks learn an embedding of continuous-time values and use an attention mechanism to produce a fixed-length representation of a time series containing a variable number of observations. We investigate the performance of this framework on interpolation and classification tasks using multiple datasets. Our results show that the proposed approach performs as well or better than a range of baseline and recently proposed models while offering significantly faster training times than current state-of-the-art methods.
{\em Computability logic} (CoL) is a powerful, mathematically rigorous computational model. In this paper, we show that CoL-web, a web extension to CoL, naturally supports web programming where database updates are involved. To be specific, we discuss an implementation of the AI ATM based on CoL (CL9 to be exact). More importantly, we argue that CoL-web supports a general AI and, therefore, is a good alternative to neural nets and deep learning. We also discuss how to integrate neural nets into CoL-web.
We introduce Gravity, another algorithm for gradient-based optimization. In this paper, we explain how our novel idea change parameters to reduce the deep learning model's loss. It has three intuitive hyper-parameters that the best values for them are proposed. Also, we propose an alternative to moving average. To compare the performance of the Gravity optimizer with two common optimizers, Adam and RMSProp, five standard datasets were trained on two VGGNet models with a batch size of 128 for 100 epochs. Gravity hyper-parameters did not need to be tuned for different models. As will be explained more in the paper, to investigate the direct impact of the optimizer itself on loss reduction no overfitting prevention technique was used. The obtained results show that the Gravity optimizer has more stable performance than Adam and RMSProp and gives greater values of validation accuracy for datasets with more output classes like CIFAR-100 (Fine).
Deep learning approaches may help radiologists in the early diagnosis and timely treatment of cerebrovascular diseases. Accurate cerebral vessel segmentation of Time-of-Flight Magnetic Resonance Angiographs (TOF-MRAs) is an essential step in this process. This study investigates deep learning approaches for automatic, fast and accurate cerebrovascular segmentation for TOF-MRAs. The performance of several data augmentation and selection methods for training a 2D and 3D U-Net for vessel segmentation was investigated in five experiments: a) without augmentation, b) Gaussian blur, c) rotation and flipping, d) Gaussian blur, rotation and flipping and e) different input patch sizes. All experiments were performed by patch-training both a 2D and 3D U-Net and predicted on a test set of MRAs. Ground truth was manually defined using an interactive threshold and region growing method. The performance was evaluated using the Dice Similarity Coefficient (DSC), Modified Hausdorff Distance and Volumetric Similarity, between the predicted images and the interactively defined ground truth. The segmentation performance of all trained networks on the test set was found to be good, with DSC scores ranging from 0.72 to 0.83. Both the 2D and 3D U-Net had the best segmentation performance with Gaussian blur, rotation and flipping compared to other experiments without augmentation or only one of those augmentation techniques. Additionally, training on larger patches or slices gave optimal segmentation results. In conclusion, vessel segmentation can be optimally performed on TOF-MRAs using a trained 3D U-Net on larger patches, where data augmentation including Gaussian blur, rotation and flipping was performed on the training data.
Ultrasound images are widespread in medical diagnosis for muscle-skeletal, cardiac, and obstetrical diseases, due to the efficiency and non-invasiveness of the acquisition methodology. However, ultrasound acquisition introduces noise in the signal, which corrupts the resulting image and affects further processing steps, e.g., segmentation and quantitative analysis. We define a novel deep learning framework for the real-time denoising of ultrasound images. Firstly, we compare state-of-the-art methods for denoising (e.g., spectral, low-rank methods) and select WNNM (Weighted Nuclear Norm Minimisation) as the best denoising in terms of accuracy, preservation of anatomical features, and edge enhancement. Then, we propose a tuned version of WNNM (tuned-WNNM) that improves the quality of the denoised images and extends its applicability to ultrasound images. Through a deep learning framework, the tuned-WNNM qualitatively and quantitatively replicates WNNM results in real-time. Finally, our approach is general in terms of its building blocks and parameters of the deep learning and high-performance computing framework; in fact, we can select different denoising algorithms and deep learning architectures.
It is shown that for deep neural networks, a single wide layer of width $N+1$ ($N$ being the number of training samples) suffices to prove the connectivity of sublevel sets of the training loss function. In the two-layer setting, the same property may not hold even if one has just one neuron less (i.e. width $N$ can lead to disconnected sublevel sets).
Deep learning has emerged as a technique of choice for rapid feature extraction across imaging disciplines, allowing rapid conversion of the data streams to spatial or spatiotemporal arrays of features of interest. However, applications of deep learning in experimental domains are often limited by the out-of-distribution drift between the experiments, where the network trained for one set of imaging conditions becomes sub-optimal for different ones. This limitation is particularly stringent in the quest to have an automated experiment setting, where retraining or transfer learning becomes impractical due to the need for human intervention and associated latencies. Here we explore the reproducibility of deep learning for feature extraction in atom-resolved electron microscopy and introduce workflows based on ensemble learning and iterative training to greatly improve feature detection. This approach both allows incorporating uncertainty quantification into the deep learning analysis and also enables rapid automated experimental workflows where retraining of the network to compensate for out-of-distribution drift due to subtle change in imaging conditions is substituted for a human operator or programmatic selection of networks from the ensemble. This methodology can be further applied to machine learning workflows in other imaging areas including optical and chemical imaging.
Quantitative susceptibility mapping (QSM) has demonstrated great potential in quantifying tissue susceptibility in various brain diseases. However, the intrinsic ill-posed inverse problem relating the tissue phase to the underlying susceptibility distribution affects the accuracy for quantifying tissue susceptibility. Recently, deep learning has shown promising results to improve accuracy by reducing the streaking artifacts. However, there exists a mismatch between the observed phase and the theoretical forward phase estimated by the susceptibility label. In this study, we proposed a model-based deep learning architecture that followed the STI (susceptibility tensor imaging) physical model, referred to as MoDL-QSM. Specifically, MoDL-QSM accounts for the relationship between STI-derived phase contrast induced by the susceptibility tensor terms (ki13,ki23,ki33) and the acquired single-orientation phase. The convolution neural networks are embedded into the physical model to learn a regularization term containing prior information. ki33 and phase induced by ki13 and ki23 terms were used as the labels for network training. Quantitative evaluation metrics (RSME, SSIM, and HFEN) were compared with recently developed deep learning QSM methods. The results showed that MoDL-QSM achieved superior performance, demonstrating its potential for future applications.
Due to the dominant position of deep learning (mostly deep neural networks) in various artificial intelligence applications, recently, ensemble learning based on deep neural networks (ensemble deep learning) has shown significant performances in improving the generalization of learning system. However, since modern deep neural networks usually have millions to billions of parameters, the time and space overheads for training multiple base deep learners and testing with the ensemble deep learner are far greater than that of traditional ensemble learning. Though several algorithms of fast ensemble deep learning have been proposed to promote the deployment of ensemble deep learning in some applications, further advances still need to be made for many applications in specific fields, where the developing time and computing resources are usually restricted or the data to be processed is of large dimensionality. An urgent problem needs to be solved is how to take the significant advantages of ensemble deep learning while reduce the required expenses so that many more applications in specific fields can benefit from it. For the alleviation of this problem, it is essential to know about how ensemble learning has developed under the era of deep learning. Thus, in this article, we present fundamental discussions focusing on data analyses of published works, methodologies, recent advances and unattainability of traditional ensemble learning and ensemble deep learning. We hope this article will be helpful to realize the intrinsic problems and technical challenges faced by future developments of ensemble learning under the era of deep learning.