From computer vision and speech recognition to forecasting trajectories in autonomous vehicles, deep learning approaches are at the forefront of so many domains. Deep learning models are developed using plethora of high-level, generic frameworks and libraries. Running those models on the mobile devices require hardware-aware optimizations and in most cases converting the models to other formats or using a third-party framework. In reality, most of the developed models need to undergo a process of conversion, adaptation, and, in some cases, full retraining to match the requirements and features of the framework that is deploying the model on the target platform. Variety of hardware platforms with heterogeneous computing elements, from wearable devices to high-performance GPU clusters are used to run deep learning models. In this paper, we present the existing challenges, obstacles, and practical solutions towards deploying deep learning models on mobile devices.
Detecting security vulnerabilities in software before they are exploited has been a challenging problem for decades. Traditional code analysis methods have been proposed, but are often ineffective and inefficient. In this work, we model software vulnerability detection as a natural language processing (NLP) problem with source code treated as texts, and address the automated software venerability detection with recent advanced deep learning NLP models assisted by transfer learning on written English. For training and testing, we have preprocessed the NIST NVD/SARD databases and built a dataset of over 100,000 files in $C$ programming language with 123 types of vulnerabilities. The extensive experiments generate the best performance of over 93\% accuracy in detecting security vulnerabilities.
Most current work in NLP utilizes deep learning, which requires a lot of training data and computational power. This paper investigates the strengths of Genetic Algorithms (GAs) for extractive summarization, as we hypothesized that GAs could construct more efficient solutions for the summarization task due to their relative customizability relative to deep learning models. This is done by building a vocabulary set, the words of which are represented as an array of weights, and optimizing those set of weights with the GA. These weights can be used to build an overall weighting of a sentence, which can then be passed to some threshold for extraction. Our results showed that the GA was able to learn a weight representation that could filter out excessive vocabulary and thus dictate sentence importance based on common English words.
Despite over two decades of progress, imbalanced data is still considered a significant challenge for contemporary machine learning models. Modern advances in deep learning have magnified the importance of the imbalanced data problem. The two main approaches to address this issue are based on loss function modifications and instance resampling. Instance sampling is typically based on Generative Adversarial Networks (GANs), which may suffer from mode collapse. Therefore, there is a need for an oversampling method that is specifically tailored to deep learning models, can work on raw images while preserving their properties, and is capable of generating high quality, artificial images that can enhance minority classes and balance the training set. We propose DeepSMOTE - a novel oversampling algorithm for deep learning models. It is simple, yet effective in its design. It consists of three major components: (i) an encoder/decoder framework; (ii) SMOTE-based oversampling; and (iii) a dedicated loss function that is enhanced with a penalty term. An important advantage of DeepSMOTE over GAN-based oversampling is that DeepSMOTE does not require a discriminator, and it generates high-quality artificial images that are both information-rich and suitable for visual inspection. DeepSMOTE code is publicly available at: https://github.com/dd1github/DeepSMOTE
Algorithms have been fundamental to recent global technological advances and, in particular, they have been the cornerstone of technical advances in one field rapidly being applied to another. We argue that algorithms possess fundamentally different qualities to deep learning methods, and this strongly suggests that, were deep learning methods better able to mimic algorithms, generalisation of the sort seen with algorithms would become possible with deep learning -- something far out of the reach of current machine learning methods. Furthermore, by representing elements in a continuous space of learnt algorithms, neural networks are able to adapt known algorithms more closely to real-world problems, potentially finding more efficient and pragmatic solutions than those proposed by human computer scientists. Here we present neural algorithmic reasoning -- the art of building neural networks that are able to execute algorithmic computation -- and provide our opinion on its transformative potential for running classical algorithms on inputs previously considered inaccessible to them.
While multiple studies have explored the relation between inter-rater variability and deep learning model uncertainty in medical segmentation tasks, little is known about the impact of individual rater style. This study quantifies rater style in the form of bias and consistency and explores their impacts when used to train deep learning models. Two multi-rater public datasets were used, consisting of brain multiple sclerosis lesion and spinal cord grey matter segmentation. On both datasets, results show a correlation ($R^2 = 0.60$ and $0.93$) between rater bias and deep learning uncertainty. The impact of label fusion between raters' annotations on this relationship is also explored, and we show that multi-center consensuses are more effective than single-center consensuses to reduce uncertainty, since rater style is mostly center-specific.
We propose a method to train deep networks to decompose videos into 3D geometry (camera and depth), moving objects, and their motions, with no supervision. We build on the idea of view synthesis, which uses classical camera geometry to re-render a source image from a different point-of-view, specified by a predicted relative pose and depth map. By minimizing the error between the synthetic image and the corresponding real image in a video, the deep network that predicts pose and depth can be trained completely unsupervised. However, the view synthesis equations rely on a strong assumption: that objects do not move. This rigid-world assumption limits the predictive power, and rules out learning about objects automatically. We propose a simple solution: minimize the error on small regions of the image instead. While the scene as a whole may be non-rigid, it is always possible to find small regions that are approximately rigid, such as inside a moving object. Our network can then predict different poses for each region, in a sliding window from a learned dense pose map. This represents a significantly richer model, including 6D object motions, with little additional complexity. We achieve very competitive performance on unsupervised odometry and depth prediction on KITTI. We also demonstrate new capabilities on EPIC-Kitchens, a challenging dataset of indoor videos, where there is no ground truth information for depth, odometry, object segmentation or motion. Yet all are recovered automatically by our method.
Cannulation for hemodialysis is the act of inserting a needle into a surgically created vascular access (e.g., an arteriovenous fistula) for the purpose of dialysis. The main risk associated with cannulation is infiltration, the puncture of the wall of the vascular access after entry, which can cause medical complications. Simulator-based training allows clinicians to gain cannulation experience without putting patients at risk. In this paper, we propose to use deep-learning-based techniques for detecting, based on video, whether the needle tip is in or has infiltrated the simulated fistula. Three categories of deep neural networks are investigated in this work: modified pre-trained models based on VGG-16 and ResNet-50, light convolutional neural networks (light CNNs), and convolutional recurrent neural networks (CRNNs). CRNNs consist of convolutional layers and a long short-term memory (LSTM) layer. A data set of cannulation experiments was collected and analyzed. The results show that both the light CNN and the CRNN achieve better performance than the pre-trained baseline models. The CRNN was implemented in real time on commodity hardware for use in the cannulation simulator, and the performance was verified. Deep-learning video analysis is a viable method for detecting needle state in a low cost cannulation simulator. Our data sets and code are released at https://github.com/axin233/DL_for_Needle_Detection_Cannulation
During recent years, the renaissance of neural networks as the major machine learning paradigm and more specifically, the confirmation that deep learning techniques provide state-of-the-art results for most of computer vision tasks has been shaking up traditional research in image processing. The same can be said for research in communities working on applied harmonic analysis, information geometry, variational methods, etc. For many researchers, this is viewed as an existential threat. On the one hand, research funding agencies privilege mainstream approaches especially when these are unquestionably suitable for solving real problems and for making progress on artificial intelligence. On the other hand, successful publishing of research in our communities is becoming almost exclusively based on a quantitative improvement of the accuracy of any benchmark task. As most of my colleagues sharing this research field, I am confronted with the dilemma of continuing to invest my time and intellectual effort on mathematical morphology as my driving force for research, or simply focussing on how to use deep learning and contributing to it. The solution is not obvious to any of us since our research is not fundamental, it is just oriented to solve challenging problems, which can be more or less theoretical. Certainly, it would be foolish for anyone to claim that deep learning is insignificant or to think that one's favourite image processing domain is productive enough to ignore the state-of-the-art. I fully understand that the labs and leading people in image processing communities have been shifting their research to almost exclusively focus on deep learning techniques. My own position is different: I do think there is room for progress on mathematically grounded image processing branches, under the condition that these are rethought in a broader sense from the deep learning paradigm. Indeed, I firmly believe that the convergence between mathematical morphology and the computation methods which gravitate around deep learning (fully connected networks, convolutional neural networks, residual neural networks, recurrent neural networks, etc.) is worthwhile. The goal of this talk is to discuss my personal vision regarding these potential interactions. Without any pretension of being exhaustive, I want to address it with a series of open questions, covering a wide range of specificities of morphological operators and representations, which could be tackled and revisited under the paradigm of deep learning. An expected benefit of such convergence between morphology and deep learning is a cross-fertilization of concepts and techniques between both fields. In addition, I think the future answer to some of these questions can provide some insight on understanding, interpreting and simplifying deep learning networks.
Total lung volume is an important quantitative biomarker and is used for the assessment of restrictive lung diseases. In this study, we investigate the performance of several deep-learning approaches for automated measurement of total lung volume from chest radiographs. 7621 posteroanterior and lateral view chest radiographs (CXR) were collected from patients with chest CT available. Similarly, 928 CXR studies were chosen from patients with pulmonary function test (PFT) results. The reference total lung volume was calculated from lung segmentation on CT or PFT data, respectively. This dataset was used to train deep-learning architectures to predict total lung volume from chest radiographs. The experiments were constructed in a step-wise fashion with increasing complexity to demonstrate the effect of training with CT-derived labels only and the sources of error. The optimal models were tested on 291 CXR studies with reference lung volume obtained from PFT. The optimal deep-learning regression model showed an MAE of 408 ml and a MAPE of 8.1\% and Pearson's r = 0.92 using both frontal and lateral chest radiographs as input. CT-derived labels were useful for pre-training but the optimal performance was obtained by fine-tuning the network with PFT-derived labels. We demonstrate, for the first time, that state-of-the-art deep learning solutions can accurately measure total lung volume from plain chest radiographs. The proposed model can be used to obtain total lung volume from routinely acquired chest radiographs at no additional cost and could be a useful tool to identify trends over time in patients referred regularly for chest x-rays.
Visual change detection, aiming at segmentation of video frames into foreground and background regions, is one of the elementary tasks in computer vision and video analytics. The applications of change detection include anomaly detection, object tracking, traffic monitoring, human machine interaction, behavior analysis, action recognition, and visual surveillance. Some of the challenges in change detection include background fluctuations, illumination variation, weather changes, intermittent object motion, shadow, fast/slow object motion, camera motion, heterogeneous object shapes and real-time processing. Traditionally, this problem has been solved using hand-crafted features and background modelling techniques. In recent years, deep learning frameworks have been successfully adopted for robust change detection. This article aims to provide an empirical review of the state-of-the-art deep learning methods for change detection. More specifically, we present a detailed analysis of the technical characteristics of different model designs and experimental frameworks. We provide model design based categorization of the existing approaches, including the 2D-CNN, 3D-CNN, ConvLSTM, multi-scale features, residual connections, autoencoders and GAN based methods. Moreover, an empirical analysis of the evaluation settings adopted by the existing deep learning methods is presented. To the best of our knowledge, this is a first attempt to comparatively analyze the different evaluation frameworks used in the existing deep change detection methods. Finally, we point out the research needs, future directions and draw our own conclusions.
Intracranial hemorrhage is a life-threatening disease, which requires fast medical intervention. Owing to the duration of data annotation, head CT images are usually available only with slice-level labeling. This paper presents a weakly supervised method of precise hemorrhage localization in axial slices using only position-free labels, which is based on multiple instance learning. An algorithm is introduced that generates hemorrhage likelihood maps and finds the coordinates of bleeding. The Dice coefficient of 58.08 % is achieved on data from a publicly available dataset.
High-Frequency (HF) signals are ubiquitous in the industrial world and are of great use for monitoring of industrial assets. Most deep learning tools are designed for inputs of fixed and/or very limited size and many successful applications of deep learning to the industrial context use as inputs extracted features, which is a manually and often arduously obtained compact representation of the original signal. In this paper, we propose a fully unsupervised deep learning framework that is able to extract a meaningful and sparse representation of raw HF signals. We embed in our architecture important properties of the fast discrete wavelet transformation (FDWT) such as (1) the cascade algorithm, (2) the conjugate quadrature filter property that links together the wavelet, the scaling and transposed filter functions, and (3) the coefficient denoising. Using deep learning, we make this architecture fully learnable: both the wavelet bases and the wavelet coefficient denoising are learnable. To achieve this objective, we propose a new activation function that performs a learnable hard-thresholding of the wavelet coefficients. With our framework, the denoising FDWT becomes a fully learnable unsupervised tool that does neither require any type of pre- nor post-processing, nor any prior knowledge on wavelet transform. We demonstrate the benefits of embedding all these properties on three machine-learning tasks performed on open source sound datasets. We perform an ablation study of the impact of each property on the performance of the architecture, achieve results well above baseline and outperform other state-of-the-art methods.
In this letter, we use deep-learning convolution neural networks (CNNs) to assess the landslide mapping and classification performances on optical images (from Sentinel-2) and SAR images (from Sentinel-1). The training and test zones used to independently evaluate the performance of the CNNs on different datasets are located in the eastern Iburi subprefecture in Hokkaido, where, at 03.08 local time (JST) on September 6, 2018, an Mw 6.6 earthquake triggered about 8000 coseismic landslides. We analyzed the conditions before and after the earthquake exploiting multi-polarization SAR as well as optical data by means of a CNN implemented in TensorFlow that points out the locations where the Landslide class is predicted as more likely. As expected, the CNN run on optical images proved itself excellent for the landslide detection task, achieving an overall accuracy of 99.20% while CNNs based on the combination of ground range detected (GRD) SAR data reached overall accuracies beyond 94%. Our findings show that the integrated use of SAR data may also allow for rapid mapping even during storms and under dense cloud cover and seems to provide comparable accuracy to classical optical change detection in landslide recognition and mapping.
The problem of measuring sentence similarity is an essential issue in the natural language processing (NLP) area. It is necessary to measure the similarity between sentences accurately. There are many approaches to measuring sentence similarity. Deep learning methodology shows a state-of-the-art performance in many natural language processing fields and is used a lot in sentence similarity measurement methods. However, in the natural language processing field, considering the structure of the sentence or the word structure that makes up the sentence is also important. In this study, we propose a methodology combined with both deep learning methodology and a method considering lexical relationships. Our evaluation metric is the Pearson correlation coefficient and Spearman correlation coefficient. As a result, the proposed method outperforms the current approaches on a KorSTS standard benchmark Korean dataset. Moreover, it performs a maximum of 65% increase than only using deep learning methodology. Experiments show that our proposed method generally results in better performance than those with only a deep learning model.
Deep learning algorithms have made many breakthroughs and have various applications in real life. Computational resources become a bottleneck as the data and complexity of the deep learning pipeline increases. In this paper, we propose optimized deep learning pipelines in multiple aspects of training including time and memory. OpTorch is a machine learning library designed to overcome weaknesses in existing implementations of neural network training. OpTorch provides features to train complex neural networks with limited computational resources. OpTorch achieved the same accuracy as existing libraries on Cifar-10 and Cifar-100 datasets while reducing memory usage to approximately 50%. We also explore the effect of weights on total memory usage in deep learning pipelines. In our experiments, parallel encoding-decoding along with sequential checkpoints results in much improved memory and time usage while keeping the accuracy similar to existing pipelines. OpTorch python package is available at available at https://github.com/cbrl-nuces/optorch
Deep learning-based video manipulation methods have become widely accessible to the masses. With little to no effort, people can quickly learn how to generate deepfake (DF) videos. While deep learning-based detection methods have been proposed to identify specific types of DFs, their performance suffers for other types of deepfake methods, including real-world deepfakes, on which they are not sufficiently trained. In other words, most of the proposed deep learning-based detection methods lack transferability and generalizability. Beyond detecting a single type of DF from benchmark deepfake datasets, we focus on developing a generalized approach to detect multiple types of DFs, including deepfakes from unknown generation methods such as DeepFake-in-the-Wild (DFW) videos. To better cope with unknown and unseen deepfakes, we introduce a Convolutional LSTM-based Residual Network (CLRNet), which adopts a unique model training strategy and explores spatial as well as the temporal information in deepfakes. Through extensive experiments, we show that existing defense methods are not ready for real-world deployment. Whereas our defense method (CLRNet) achieves far better generalization when detecting various benchmark deepfake methods (97.57% on average). Furthermore, we evaluate our approach with a high-quality DeepFake-in-the-Wild dataset, collected from the Internet containing numerous videos and having more than 150,000 frames. Our CLRNet model demonstrated that it generalizes well against high-quality DFW videos by achieving 93.86% detection accuracy, outperforming existing state-of-the-art defense methods by a considerable margin.
Humans can learn a variety of concepts and skills incrementally over the course of their lives while exhibiting many desirable properties, such as continual learning without forgetting, forward transfer and backward transfer of knowledge, and learning a new concept or task with only a few examples. Several lines of machine learning research, such as lifelong machine learning, few-shot learning, and transfer learning attempt to capture these properties. However, most previous approaches can only demonstrate subsets of these properties, often by different complex mechanisms. In this work, we propose a simple yet powerful unified deep learning framework that supports almost all of these properties and approaches through one central mechanism. Experiments on toy examples support our claims. We also draw connections between many peculiarities of human learning (such as memory loss and "rain man") and our framework. As academics, we often lack resources required to build and train, deep neural networks with billions of parameters on hundreds of TPUs. Thus, while our framework is still conceptual, and our experiment results are surely not SOTA, we hope that this unified lifelong learning framework inspires new work towards large-scale experiments and understanding human learning in general. This paper is summarized in two short YouTube videos: https://youtu.be/gCuUyGETbTU (part 1) and https://youtu.be/XsaGI01b-1o (part 2).
We describe a novel application of the end-to-end deep learning technique to the task of discriminating top quark-initiated jets from those originating from the hadronization of a light quark or a gluon. The end-to-end deep learning technique combines deep learning algorithms and low-level detector representation of the high-energy collision event. In this study, we use low-level detector information from the simulated CMS Open Data samples to construct the top jet classifiers. To optimize classifier performance we progressively add low-level information from the CMS tracking detector, including pixel detector reconstructed hits and impact parameters, and demonstrate the value of additional tracking information even when no new spatial structures are added. Relying only on calorimeter energy deposits and reconstructed pixel detector hits, the end-to-end classifier achieves an AUC score of 0.975$\pm$0.002 for the task of classifying boosted top quark jets. After adding derived track quantities, the classifier AUC score increases to 0.9824$\pm$0.0013, serving as the first performance benchmark for these CMS Open Data samples. We additionally provide a timing performance comparison of different processor unit architectures for training the network.
This study identifies and proposes techniques to alleviate two key bottlenecks to executing deep neural networks in trusted execution environments (TEEs): page thrashing during the execution of convolutional layers and the decryption of large weight matrices in fully-connected layers. For the former, we propose a novel partitioning scheme, y-plane partitioning, designed to (i) provide consistent execution time when the layer output is large compared to the TEE secure memory; and (ii) significantly reduce the memory footprint of convolutional layers. For the latter, we leverage quantization and compression. In our evaluation, the proposed optimizations incurred latency overheads ranging from 1.09X to 2X baseline for a wide range of TEE sizes; in contrast, an unmodified implementation incurred latencies of up to 26X when running inside of the TEE.