With the recent increase in the number of underwater activities, having effective underwater communication systems has become increasingly important. Underwater acoustic communication has been widely used but greatly impaired due to the complicated nature of the underwater environment. In a bid to better understand the underwater acoustic channel so as to help in the design and improvement of underwater communication systems, attempts have been made to model the underwater acoustic channel using mathematical equations and approximations under some assumptions. In this paper, we explore the capability of machine learning and deep learning methods to learn and accurately model the underwater acoustic channel using real underwater data collected from a water tank with disturbance and from lake Tahoe. Specifically, Deep Neural Network (DNN) and Long Short Term Memory (LSTM) are applied to model the underwater acoustic channel. Experimental results show that these models are able to model the underwater acoustic communication channel well and that deep learning models, especially LSTM are better models in terms of mean absolute percentage error.
Many applications in image-guided surgery and therapy require fast and reliable non-linear, multi-modal image registration. Recently proposed unsupervised deep learning-based registration methods have demonstrated superior performance compared to iterative methods in just a fraction of the time. Most of the learning-based methods have focused on mono-modal image registration. The extension to multi-modal registration depends on the use of an appropriate similarity function, such as the mutual information (MI). We propose guiding the training of a deep learning-based registration method with MI estimation between an image-pair in an end-to-end trainable network. Our results show that a small, 2-layer network produces competitive results in both mono- and multi-modal registration, with sub-second run-times. Comparisons to both iterative and deep learning-based methods show that our MI-based method produces topologically and qualitatively superior results with an extremely low rate of non-diffeomorphic transformations. Real-time clinical application will benefit from a better visual matching of anatomical structures and less registration failures/outliers.
Recent years have witnessed enormous progress of online learning. However, a major challenge on the road to artificial agents is concept drift, that is, the data probability distribution would change where the data instance arrives sequentially in a stream fashion, which would lead to catastrophic forgetting and degrade the performance of the model. In this paper, we proposed a new Bilevel Online Deep Learning (BODL) framework, which combine bilevel optimization strategy and online ensemble classifier. In BODL algorithm, we use an ensemble classifier, which use the output of different hidden layers in deep neural network to build multiple base classifiers, the important weights of the base classifiers are updated according to exponential gradient descent method in an online manner. Besides, we apply the similar constraint to overcome the convergence problem of online ensemble framework. Then an effective concept drift detection mechanism utilizing the error rate of classifier is designed to monitor the change of the data probability distribution. When the concept drift is detected, our BODL algorithm can adaptively update the model parameters via bilevel optimization and then circumvent the large drift and encourage positive transfer. Finally, the extensive experiments and ablation studies are conducted on various datasets and the competitive numerical results illustrate that our BODL algorithm is a promising approach.
Snapshot compressive imaging (SCI) can record the 3D information by a 2D measurement and from this 2D measurement to reconstruct the original 3D information by reconstruction algorithm. As we can see, the reconstruction algorithm plays a vital role in SCI. Recently, deep learning algorithm show its outstanding ability, outperforming the traditional algorithm. Therefore, to improve deep learning algorithm reconstruction accuracy is an inevitable topic for SCI. Besides, deep learning algorithms are usually limited by scalability, and a well trained model in general can not be applied to new systems if lacking the new training process. To address these problems, we develop the ensemble learning priors to further improve the reconstruction accuracy and propose the scalable learning to empower deep learning the scalability just like the traditional algorithm. What's more, our algorithm has achieved the state-of-the-art results, outperforming existing algorithms. Extensive results on both simulation and real datasets demonstrate the superiority of our proposed algorithm. The code and models will be released to the public.
Imperative programming allows users to implement their deep neural networks (DNNs) easily and has become an essential part of recent deep learning (DL) frameworks. Recently, several systems have been proposed to combine the usability of imperative programming with the optimized performance of symbolic graph execution. Such systems convert imperative Python DL programs to optimized symbolic graphs and execute them. However, they cannot fully support the usability of imperative programming. For example, if an imperative DL program contains a Python feature with no corresponding symbolic representation (e.g., third-party library calls or unsupported dynamic control flows) they fail to execute the program. To overcome this limitation, we propose Terra, an imperative-symbolic co-execution system that can handle any imperative DL programs while achieving the optimized performance of symbolic graph execution. To achieve this, Terra builds a symbolic graph by decoupling DL operations from Python features. Then, Terra conducts the imperative execution to support all Python features, while delegating the decoupled operations to the symbolic execution. We evaluated the performance improvement and coverage of Terra with ten imperative DL programs for several DNN architectures. The results show that Terra can speed up the execution of all ten imperative DL programs, whereas AutoGraph, one of the state-of-the-art systems, fails to execute five of them.
Location based services, already popular with end users, are now inevitably becoming part of new wireless infrastructures and emerging business processes. The increasingly popular Deep Learning (DL) artificial intelligence methods perform very well in wireless fingerprinting localization based on extensive indoor radio measurement data. However, with the increasing complexity these methods become computationally very intensive and energy hungry, both for their training and subsequent operation. Considering only mobile users, estimated to exceed 7.4billion by the end of 2025, and assuming that the networks serving these users will need to perform only one localization per user per hour on average, the machine learning models used for the calculation would need to perform 65*10^12 predictions per year. Add to this equation tens of billions of other connected devices and applications that rely heavily on more frequent location updates, and it becomes apparent that localization will contribute significantly to carbon emissions unless more energy-efficient models are developed and used. This motivated our work on a new DL-based architecture for indoor localization that is more energy efficient compared to related state-of-the-art approaches while showing only marginal performance degradation. A detailed performance evaluation shows that the proposed model producesonly 58 % of the carbon footprint while maintaining 98.7 % of the overall performance compared to state of the art model external to our group. Additionally, we elaborate on a methodology to calculate the complexity of the DL model and thus the CO2 footprint during its training and operation.
Automatic classification of diabetic retinopathy from retinal images has been widely studied using deep neural networks with impressive results. However, there is a clinical need for estimation of the uncertainty in the classifications, a shortcoming of modern neural networks. Recently, approximate Bayesian deep learning methods have been proposed for the task but the studies have only considered the binary referable/non-referable diabetic retinopathy classification applied to benchmark datasets. We present novel results by systematically investigating a clinical dataset and a clinically relevant 5-class classification scheme, in addition to benchmark datasets and the binary classification scheme. Moreover, we derive a connection between uncertainty measures and classifier risk, from which we develop a new uncertainty measure. We observe that the previously proposed entropy-based uncertainty measure generalizes to the clinical dataset on the binary classification scheme but not on the 5-class scheme, whereas our new uncertainty measure generalizes to the latter case.
Deep learning has been the answer to many machine learning problems during the past two decades. However, it comes with two major constraints: dependency on extensive labeled data and training costs. Transfer learning in deep learning, known as Deep Transfer Learning (DTL), attempts to reduce such dependency and costs by reusing an obtained knowledge from a source data/task in training on a target data/task. Most applied DTL techniques are network/model-based approaches. These methods reduce the dependency of deep learning models on extensive training data and drastically decrease training costs. As a result, researchers detected Covid-19 infection on chest X-Rays with high accuracy at the beginning of the pandemic with minimal data using DTL techniques. Also, the training cost reduction makes DTL viable on edge devices with limited resources. Like any new advancement, DTL methods have their own limitations, and a successful transfer depends on some adjustments for different scenarios. In this paper, we review the definition and taxonomy of deep transfer learning and well-known methods. Then we investigate the DTL approaches by reviewing recent applied DTL techniques in the past five years. Further, we review some experimental analyses of DTLs to learn the best practice for applying DTL in different scenarios. Moreover, the limitations of DTLs (catastrophic forgetting dilemma and overly biased pre-trained models) are discussed, along with possible solutions and research trends.
Lung cancer is the leading cause of cancer death worldwide and a good prognosis depends on early diagnosis. Unfortunately, screening programs for the early diagnosis of lung cancer are uncommon. This is in-part due to the at-risk groups being located in rural areas far from medical facilities. Reaching these populations would require a scaled approach that combines mobility, low cost, speed, accuracy, and privacy. We can resolve these issues by combining the chest X-ray imaging mode with a federated deep-learning approach, provided that the federated model is trained on homogenous data to ensure that no single data source can adversely bias the model at any point in time. In this study we show that an image pre-processing pipeline that homogenizes and debiases chest X-ray images can improve both internal classification and external generalization, paving the way for a low-cost and accessible deep learning-based clinical system for lung cancer screening. An evolutionary pruning mechanism is used to train a nodule detection deep learning model on the most informative images from a publicly available lung nodule X-ray dataset. Histogram equalization is used to remove systematic differences in image brightness and contrast. Model training is performed using all combinations of lung field segmentation, close cropping, and rib suppression operators. We show that this pre-processing pipeline results in deep learning models that successfully generalize an independent lung nodule dataset using ablation studies to assess the contribution of each operator in this pipeline. In stripping chest X-ray images of known confounding variables by lung field segmentation, along with suppression of signal noise from the bone structure we can train a highly accurate deep learning lung nodule detection algorithm with outstanding generalization accuracy of 89% to nodule samples in unseen data.
Recent research in feature learning has been extended to sequence data, where each instance consists of a sequence of heterogeneous items with a variable length. However, in many real-world applications, the data exists in the form of attributed sequences, which is composed of a set of fixed-size attributes and variable-length sequences with dependencies between them. In the attributed sequence context, feature learning remains challenging due to the dependencies between sequences and their associated attributes. In this dissertation, we focus on analyzing and building deep learning models for four new problems on attributed sequences. Our extensive experiments on real-world datasets demonstrate that the proposed solutions significantly improve the performance of each task over the state-of-the-art methods on attributed sequences.
Supervised deep learning approaches to underdetermined audio source separation achieve state-of-the-art performance but require a dataset of mixtures along with their corresponding isolated source signals. Such datasets can be extremely costly to obtain for musical mixtures. This raises a need for unsupervised methods. We propose a novel unsupervised model-based deep learning approach to musical source separation. Each source is modelled with a differentiable parametric source-filter model. A neural network is trained to reconstruct the observed mixture as a sum of the sources by estimating the source models' parameters given their fundamental frequencies. At test time, soft masks are obtained from the synthesized source signals. The experimental evaluation on a vocal ensemble separation task shows that the proposed method outperforms learning-free methods based on nonnegative matrix factorization and a supervised deep learning baseline. Integrating domain knowledge in the form of source models into a data-driven method leads to high data efficiency: the proposed approach achieves good separation quality even when trained on less than three minutes of audio. This work makes powerful deep learning based separation usable in scenarios where training data with ground truth is expensive or nonexistent.
Deep learning is applied to many complex tasks in the field of wireless communication, such as modulation recognition of spectrum waveforms, because of its convenience and efficiency. This leads to the problem of a malicious third party using a deep learning model to easily recognize the modulation format of the transmitted waveform. Some existing works address this problem directly using the concept of adversarial examples in the image domain without fully considering the characteristics of the waveform transmission in the physical world. Therefore, we propose a low-intercept waveform~(LIW) generation method that can reduce the probability of the modulation being recognized by a third party without affecting the reliable communication of the friendly party. Our LIW exhibits significant low-interception performance even in the physical hardware experiment, decreasing the accuracy of the state of the art model to approximately $15\%$ with small perturbations.
The potential energy formulation and deep learning are merged to solve partial differential equations governing the deformation in hyperelastic and viscoelastic materials. The presented deep energy method (DEM) is self-contained and meshfree. It can accurately capture the three-dimensional (3D) mechanical response without requiring any time-consuming training data generation by classical numerical methods such as the finite element method. Once the model is appropriately trained, the response can be attained almost instantly at any point in the physical domain, given its spatial coordinates. Therefore, the deep energy method is potentially a promising standalone method for solving partial differential equations describing the mechanical deformation of materials or structural systems and other physical phenomena.
Deep learning has shown great promise in the domain of medical image analysis. Medical professionals and healthcare providers have been adopting the technology to speed up and enhance their work. These systems use deep neural networks (DNN) which are vulnerable to adversarial samples; images with imperceivable changes that can alter the model's prediction. Researchers have proposed defences which either make a DNN more robust or detect the adversarial samples before they do harm. However, none of these works consider an informed attacker which can adapt to the defence mechanism. We show that an informed attacker can evade five of the current state of the art defences while successfully fooling the victim's deep learning model, rendering these defences useless. We then suggest better alternatives for securing healthcare DNNs from such attacks: (1) harden the system's security and (2) use digital signatures.
The second edition of the "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" challenges featured five data-impaired challenges, where models are trained from scratch on a reduced number of training samples for various key computer vision tasks. To encourage new and creative ideas on incorporating relevant inductive biases to improve the data efficiency of deep learning models, we prohibited the use of pre-trained checkpoints and other transfer learning techniques. The provided baselines are outperformed by a large margin in all five challenges, mainly thanks to extensive data augmentation policies, model ensembling, and data efficient network architectures.
Deep learning (DL) has proven to be a highly effective approach for developing models in diverse contexts, including visual perception, speech recognition, and machine translation. However, the end-to-end process for applying DL is not trivial. It requires grappling with problem formulation and context understanding, data engineering, model development, deployment, continuous monitoring and maintenance, and so on. Moreover, each of these steps typically relies heavily on humans, in terms of both knowledge and interactions, which impedes the further advancement and democratization of DL. Consequently, in response to these issues, a new field has emerged over the last few years: automated deep learning (AutoDL). This endeavor seeks to minimize the need for human involvement and is best known for its achievements in neural architecture search (NAS), a topic that has been the focus of several surveys. That stated, NAS is not the be-all and end-all of AutoDL. Accordingly, this review adopts an overarching perspective, examining research efforts into automation across the entirety of an archetypal DL workflow. In so doing, this work also proposes a comprehensive set of ten criteria by which to assess existing work in both individual publications and broader research areas. These criteria are: novelty, solution quality, efficiency, stability, interpretability, reproducibility, engineering quality, scalability, generalizability, and eco-friendliness. Thus, ultimately, this review provides an evaluative overview of AutoDL in the early 2020s, identifying where future opportunities for progress may exist.
Deep learning methods have shown strong performance in solving tasks for historical document image analysis. However, despite current libraries and frameworks, programming an experiment or a set of experiments and executing them can be time-consuming. This is why we propose an open-source deep learning framework, DIVA-DAF, which is based on PyTorch Lightning and specifically designed for historical document analysis. Pre-implemented tasks such as segmentation and classification can be easily used or customized. It is also easy to create one's own tasks with the benefit of powerful modules for loading data, even large data sets, and different forms of ground truth. The applications conducted have demonstrated time savings for the programming of a document analysis task, as well as for different scenarios such as pre-training or changing the architecture. Thanks to its data module, the framework also allows to reduce the time of model training significantly.
We propose a new deep learning-based method for estimating the occupancy of vegetation strata from airborne 3D LiDAR point clouds. Our model predicts rasterized occupancy maps for three vegetation strata corresponding to lower, medium, and higher cover. Our weakly-supervised training scheme allows our network to only be supervised with vegetation occupancy values aggregated over cylindrical plots containing thousands of points. Such ground truth is easier to produce than pixel-wise or point-wise annotations. Our method outperforms handcrafted and deep learning baselines in terms of precision by up to 30%, while simultaneously providing visual and interpretable predictions. We provide an open-source implementation along with a dataset of 199 agricultural plots to train and evaluate weakly supervised occupancy regression algorithms.
Communication overhead is one of the major obstacles to train large deep learning models at scale. Gradient sparsification is a promising technique to reduce the communication volume. However, it is very challenging to obtain real performance improvement because of (1) the difficulty of achieving an scalable and efficient sparse allreduce algorithm and (2) the sparsification overhead. This paper proposes O$k$-Top$k$, a scheme for distributed training with sparse gradients. O$k$-Top$k$ integrates a novel sparse allreduce algorithm (less than 6$k$ communication volume which is asymptotically optimal) with the decentralized parallel Stochastic Gradient Descent (SGD) optimizer, and its convergence is proved. To reduce the sparsification overhead, O$k$-Top$k$ efficiently selects the top-$k$ gradient values according to an estimated threshold. Evaluations are conducted on the Piz Daint supercomputer with neural network models from different deep learning domains. Empirical results show that O$k$-Top$k$ achieves similar model accuracy to dense allreduce. Compared with the optimized dense and the state-of-the-art sparse allreduces, O$k$-Top$k$ is more scalable and significantly improves training throughput (e.g., 3.29x-12.95x improvement for BERT on 256 GPUs).
Understanding deep learning is increasingly emergent as it penetrates more and more into industry and science. In recent years, a research line from Fourier analysis sheds lights on this magical "black box" by showing a Frequency Principle (F-Principle or spectral bias) of the training behavior of deep neural networks (DNNs) -- DNNs often fit functions from low to high frequency during the training. The F-Principle is first demonstrated by onedimensional synthetic data followed by the verification in high-dimensional real datasets. A series of works subsequently enhance the validity of the F-Principle. This low-frequency implicit bias reveals the strength of neural network in learning low-frequency functions as well as its deficiency in learning high-frequency functions. Such understanding inspires the design of DNN-based algorithms in practical problems, explains experimental phenomena emerging in various scenarios, and further advances the study of deep learning from the frequency perspective. Although incomplete, we provide an overview of F-Principle and propose some open problems for future research.