To break the GPU memory wall for scaling deep learning workloads, a variety of architecture and system techniques have been proposed recently. Their typical approaches include memory extension with flash memory and direct storage access. However, these techniques still suffer from suboptimal performance and introduce complexity to the GPU memory management, making them hard to meet the scalability requirement of deep learning workloads today. In this paper, we present a unified GPU memory and storage architecture named G10 driven by the fact that the tensor behaviors of deep learning workloads are highly predictable. G10 integrates the host memory, GPU memory, and flash memory into a unified memory space, to scale the GPU memory capacity while enabling transparent data migrations. Based on this unified GPU memory and storage architecture, G10 utilizes compiler techniques to characterize the tensor behaviors in deep learning workloads. Therefore, it can schedule data migrations in advance by considering the available bandwidth of flash memory and host memory. The cooperative mechanism between deep learning compilers and the unified memory architecture enables G10 to hide data transfer overheads in a transparent manner. We implement G10 based on an open-source GPU simulator. Our experiments demonstrate that G10 outperforms state-of-the-art GPU memory solutions by up to 1.75$\times$, without code modifications to deep learning workloads. With the smart data migration mechanism, G10 can reach 90.3\% of the performance of the ideal case assuming unlimited GPU memory.
In recent years, deep learning has revolutionized natural language processing (NLP) by enabling the development of models that can learn complex representations of language data, leading to significant improvements in performance across a wide range of NLP tasks. Deep learning models for NLP typically use large amounts of data to train deep neural networks, allowing them to learn the patterns and relationships in language data. This is in contrast to traditional NLP approaches, which rely on hand-engineered features and rules to perform NLP tasks. The ability of deep neural networks to learn hierarchical representations of language data, handle variable-length input sequences, and perform well on large datasets makes them well-suited for NLP applications. Driven by the exponential growth of textual data and the increasing demand for condensed, coherent, and informative summaries, text summarization has been a critical research area in the field of NLP. Applying deep learning to text summarization refers to the use of deep neural networks to perform text summarization tasks. In this survey, we begin with a review of fashionable text summarization tasks in recent years, including extractive, abstractive, multi-document, and so on. Next, we discuss most deep learning-based models and their experimental results on these tasks. The paper also covers datasets and data representation for summarization tasks. Finally, we delve into the opportunities and challenges associated with summarization tasks and their corresponding methodologies, aiming to inspire future research efforts to advance the field further. A goal of our survey is to explain how these methods differ in their requirements as understanding them is essential for choosing a technique suited for a specific setting.
The dynamic and unpredictable nature of road traffic necessitates effective accident detection methods for enhancing safety and streamlining traffic management in smart cities. This paper offers a comprehensive exploration study of prevailing accident detection techniques, shedding light on the nuances of other state-of-the-art methodologies while providing a detailed overview of distinct traffic accident types like rear-end collisions, T-bone collisions, and frontal impact accidents. Our novel approach introduces the I3D-CONVLSTM2D model architecture, a lightweight solution tailored explicitly for accident detection in smart city traffic surveillance systems by integrating RGB frames with optical flow information. Our experimental study's empirical analysis underscores our approach's efficacy, with the I3D-CONVLSTM2D RGB + Optical-Flow (Trainable) model outperforming its counterparts, achieving an impressive 87\% Mean Average Precision (MAP). Our findings further elaborate on the challenges posed by data imbalances, particularly when working with a limited number of datasets, road structures, and traffic scenarios. Ultimately, our research illuminates the path towards a sophisticated vision-based accident detection system primed for real-time integration into edge IoT devices within smart urban infrastructures.
The ability of Deep Learning to process and extract relevant information in complex brain dynamics from raw EEG data has been demonstrated in various recent works. Deep learning models, however, have also been shown to perform best on large corpora of data. When processing EEG, a natural approach is to combine EEG datasets from different experiments to train large deep-learning models. However, most EEG experiments use custom channel montages, requiring the data to be transformed into a common space. Previous methods have used the raw EEG signal to extract features of interest and focused on using a common feature space across EEG datasets. While this is a sensible approach, it underexploits the potential richness of EEG raw data. Here, we explore using spatial attention applied to EEG electrode coordinates to perform channel harmonization of raw EEG data, allowing us to train deep learning on EEG data using different montages. We test this model on a gender classification task. We first show that spatial attention increases model performance. Then, we show that a deep learning model trained on data using different channel montages performs significantly better than deep learning models trained on fixed 23- and 128-channel data montages.
Narrow bit-width data formats are key to reducing the computational and storage costs of modern deep learning applications. This paper evaluates Microscaling (MX) data formats that combine a per-block scaling factor with narrow floating-point and integer types for individual elements. MX formats balance the competing needs of hardware efficiency, model accuracy, and user friction. Empirical results on over two dozen benchmarks demonstrate practicality of MX data formats as a drop-in replacement for baseline FP32 for AI inference and training with low user friction. We also show the first instance of training generative language models at sub-8-bit weights, activations, and gradients with minimal accuracy loss and no modifications to the training recipe.
Many analyses in particle and nuclear physics use simulations to infer fundamental, effective, or phenomenological parameters of the underlying physics models. When the inference is performed with unfolded cross sections, the observables are designed using physics intuition and heuristics. We propose to design targeted observables with machine learning. Unfolded, differential cross sections in a neural network output contain the most information about parameters of interest and can be well-measured by construction. The networks are trained using a custom loss function that rewards outputs that are sensitive to the parameter(s) of interest while simultaneously penalizing outputs that are different between particle-level and detector-level (to minimize detector distortions). We demonstrate this idea in simulation using two physics models for inclusive measurements in deep inelastic scattering. We find that the new approach is more sensitive than classical observables at distinguishing the two models and also has a reduced unfolding uncertainty due to the reduced detector distortions.
The work proposes a novel deep-learning framework for the synthesis of three-dimensional MRI volumes from corresponding 3D ultrasound images of the brain, leveraging a modified iteration of the Pix2Pix Generative Adversarial Network (GAN) model. Addressing the formidable challenge of bridging the modality disparity between ultrasound and MRI, this research holds promise for transformative applications in medical diagnostics and treatment planning within the neuroimaging domain. While the findings reveal a discernible degree of similarity between the synthesized MRI volumes and anticipated outcomes, they fall short of practical deployment standards, primarily due to constraints associated with dataset scale and computational resources. The methodology yields MRI volumes with a satisfactory similarity score, establishing a foundational benchmark for subsequent investigations.
Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. It is associated with an increased risk of stroke, heart failure, and other cardiovascular complications, but can be clinically silent. Passive AF monitoring with wearables may help reduce adverse clinical outcomes related to AF. Detecting AF in noisy wearable data poses a significant challenge, leading to the emergence of various deep learning techniques. Previous deep learning models learn from a single modality, either electrocardiogram (ECG) or photoplethysmography (PPG) signals. However, deep learning models often struggle to learn generalizable features and rely on features that are more susceptible to corruption from noise, leading to sub-optimal performances in certain scenarios, especially with low-quality signals. Given the increasing availability of ECG and PPG signal pairs from wearables and bedside monitors, we propose a new approach, SiamAF, leveraging a novel Siamese network architecture and joint learning loss function to learn shared information from both ECG and PPG signals. At inference time, the proposed model is able to predict AF from either PPG or ECG and outperforms baseline methods on three external test sets. It learns medically relevant features as a result of our novel architecture design. The proposed model also achieves comparable performance to traditional learning regimes while requiring much fewer training labels, providing a potential approach to reduce future reliance on manual labeling.
In this paper, we investigate the convergence properties of a wide class of Adam-family methods for minimizing quadratically regularized nonsmooth nonconvex optimization problems, especially in the context of training nonsmooth neural networks with weight decay. Motivated by the AdamW method, we propose a novel framework for Adam-family methods with decoupled weight decay. Within our framework, the estimators for the first-order and second-order moments of stochastic subgradients are updated independently of the weight decay term. Under mild assumptions and with non-diminishing stepsizes for updating the primary optimization variables, we establish the convergence properties of our proposed framework. In addition, we show that our proposed framework encompasses a wide variety of well-known Adam-family methods, hence offering convergence guarantees for these methods in the training of nonsmooth neural networks. More importantly, we show that our proposed framework asymptotically approximates the SGD method, thereby providing an explanation for the empirical observation that decoupled weight decay enhances generalization performance for Adam-family methods. As a practical application of our proposed framework, we propose a novel Adam-family method named Adam with Decoupled Weight Decay (AdamD), and establish its convergence properties under mild conditions. Numerical experiments demonstrate that AdamD outperforms Adam and is comparable to AdamW, in the aspects of both generalization performance and efficiency.
Classical wisdom in machine learning holds that the generalization error can be decomposed into bias and variance, and these two terms exhibit a \emph{trade-off}. However, in this paper, we show that for an ensemble of deep learning based classification models, bias and variance are \emph{aligned} at a sample level, where squared bias is approximately \emph{equal} to variance for correctly classified sample points. We present empirical evidence confirming this phenomenon in a variety of deep learning models and datasets. Moreover, we study this phenomenon from two theoretical perspectives: calibration and neural collapse. We first show theoretically that under the assumption that the models are well calibrated, we can observe the bias-variance alignment. Second, starting from the picture provided by the neural collapse theory, we show an approximate correlation between bias and variance.
Craters are one of the most prominent features on planetary surfaces, used in applications such as age estimation, hazard detection, and spacecraft navigation. Crater detection is a challenging problem due to various aspects, including complex crater characteristics such as varying sizes and shapes, data resolution, and planetary data types. Similar to other computer vision tasks, deep learning-based approaches have significantly impacted research on crater detection in recent years. This survey aims to assist researchers in this field by examining the development of deep learning-based crater detection algorithms (CDAs). The review includes over 140 research works covering diverse crater detection approaches, including planetary data, craters database, and evaluation metrics. To be specific, we discuss the challenges in crater detection due to the complex properties of the craters and survey the DL-based CDAs by categorizing them into three parts: (a) semantic segmentation-based, (b) object detection-based, and (c) classification-based. Additionally, we have conducted training and testing of all the semantic segmentation-based CDAs on a common dataset to evaluate the effectiveness of each architecture for crater detection and its potential applications. Finally, we have provided recommendations for potential future works.
Ultrasound Localization Microscopy can resolve the microvascular bed down to a few micrometers. To achieve such performance microbubble contrast agents must perfuse the entire microvascular network. Microbubbles are then located individually and tracked over time to sample individual vessels, typically over hundreds of thousands of images. To overcome the fundamental limit of diffraction and achieve a dense reconstruction of the network, low microbubble concentrations must be used, which lead to acquisitions lasting several minutes. Conventional processing pipelines are currently unable to deal with interference from multiple nearby microbubbles, further reducing achievable concentrations. This work overcomes this problem by proposing a Deep Learning approach to recover dense vascular networks from ultrasound acquisitions with high microbubble concentrations. A realistic mouse brain microvascular network, segmented from 2-photon microscopy, was used to train a three-dimensional convolutional neural network based on a V-net architecture. Ultrasound data sets from multiple microbubbles flowing through the microvascular network were simulated and used as ground truth to train the 3D CNN to track microbubbles. The 3D-CNN approach was validated in silico using a subset of the data and in vivo on a rat brain acquisition. In silico, the CNN reconstructed vascular networks with higher precision (81%) than a conventional ULM framework (70%). In vivo, the CNN could resolve micro vessels as small as 10 $\mu$m with an increase in resolution when compared against a conventional approach.
Deep learning vulnerability detection has shown promising results in recent years. However, an important challenge that still blocks it from being very useful in practice is that the model is not robust under perturbation and it cannot generalize well over the out-of-distribution (OOD) data, e.g., applying a trained model to unseen projects in real world. We hypothesize that this is because the model learned non-robust features, e.g., variable names, that have spurious correlations with labels. When the perturbed and OOD datasets no longer have the same spurious features, the model prediction fails. To address the challenge, in this paper, we introduced causality into deep learning vulnerability detection. Our approach CausalVul consists of two phases. First, we designed novel perturbations to discover spurious features that the model may use to make predictions. Second, we applied the causal learning algorithms, specifically, do-calculus, on top of existing deep learning models to systematically remove the use of spurious features and thus promote causal based prediction. Our results show that CausalVul consistently improved the model accuracy, robustness and OOD performance for all the state-of-the-art models and datasets we experimented. To the best of our knowledge, this is the first work that introduces do calculus based causal learning to software engineering models and shows it's indeed useful for improving the model accuracy, robustness and generalization. Our replication package is located at https://figshare.com/s/0ffda320dcb96c249ef2.
The Type-II codebook in Release 17 (R17) exploits the angular-delay-domain partial reciprocity between uplink and downlink channels to select part of angular-delay-domain ports for measuring and feeding back the downlink channel state information (CSI), where the performance of existing deep learning enhanced CSI feedback methods is limited due to the deficiency of sparse structures. To address this issue, we propose two new perspectives of adopting deep learning to improve the R17 Type-II codebook. Firstly, considering the low signal-to-noise ratio of uplink channels, deep learning is utilized to accurately select the dominant angular-delay-domain ports, where the focal loss is harnessed to solve the class imbalance problem. Secondly, we propose to adopt deep learning to reconstruct the downlink CSI based on the feedback of the R17 Type-II codebook at the base station, where the information of sparse structures can be effectively leveraged. Besides, a weighted shortcut module is designed to facilitate the accurate reconstruction. Simulation results demonstrate that our proposed methods could improve the sum rate performance compared with its traditional R17 Type-II codebook and deep learning benchmarks.
With changing climatic conditions, we are already seeing an increase in extreme weather events and their secondary consequences, including landslides. Landslides threaten infrastructure, including roads, railways, buildings, and human life. Hazard-based spatial planning and early warning systems are cost-effective strategies to reduce the risk to society from landslides. However, these both rely on data from previous landslide events, which is often scarce. Many deep learning (DL) models have recently been applied for landside mapping using medium- to high-resolution satellite images as input. However, they often suffer from sensitivity problems, overfitting, and low mapping accuracy. This study addresses some of these limitations by using a diverse global landslide dataset, using different segmentation models, such as Unet, Linknet, PSP-Net, PAN, and DeepLab and based on their performances, building an ensemble model. The ensemble model achieved the highest F1-score (0.69) when combining both Sentinel-1 and Sentinel-2 bands, with the highest average improvement of 6.87 % when the ensemble size was 20. On the other hand, Sentinel-2 bands only performed very well, with an F1 score of 0.61 when the ensemble size is 20 with an improvement of 14.59 % when the ensemble size is 20. This result shows considerable potential in building a robust and reliable monitoring system based on changes in vegetation index dNDVI only.
The existence of adversarial images has seriously affected the task of image recognition and practical application of deep learning, it is also a key scientific problem that deep learning urgently needs to solve. By far the most effective approach is to train the neural network with a large number of adversarial examples. However, this adversarial training method requires a huge amount of computing resources when applied to ImageNet, and has not yet achieved satisfactory results for high-intensity adversarial attacks. In this paper, we construct an interference neural network by applying additional background images and corresponding labels, and use pre-trained ResNet-152 to efficiently complete the training. Compared with the state-of-the-art results under the PGD attack, it has a better defense effect with much smaller computing resources. This work provides new ideas for academic research and practical applications of effective defense against adversarial attacks.
We introduce a fully 3D, deep learning-based approach for the joint inversion of time-lapse surface gravity and seismic data for reconstructing subsurface density and velocity models. The target application of this proposed inversion approach is the prediction of subsurface CO2 plumes as a complementary tool for monitoring CO2 sequestration deployments. Our joint inversion technique outperforms deep learning-based gravity-only and seismic-only inversion models, achieving improved density and velocity reconstruction, accurate segmentation, and higher R-squared coefficients. These results indicate that deep learning-based joint inversion is an effective tool for CO$_2$ storage monitoring. Future work will focus on validating our approach with larger datasets, simulations with other geological storage sites, and ultimately field data.
Our goal is to provide a review of deep learning methods which provide insight into structured high-dimensional data. Rather than using shallow additive architectures common to most statistical models, deep learning uses layers of semi-affine input transformations to provide a predictive rule. Applying these layers of transformations leads to a set of attributes (or, features) to which probabilistic statistical methods can be applied. Thus, the best of both worlds can be achieved: scalable prediction rules fortified with uncertainty quantification, where sparse regularization finds the features.
Numerous cellular functions rely on protein$\unicode{x2013}$protein interactions. Efforts to comprehensively characterize them remain challenged however by the diversity of molecular recognition mechanisms employed within the proteome. Deep learning has emerged as a promising approach for tackling this problem by exploiting both experimental data and basic biophysical knowledge about protein interactions. Here, we review the growing ecosystem of deep learning methods for modeling protein interactions, highlighting the diversity of these biophysically-informed models and their respective trade-offs. We discuss recent successes in using representation learning to capture complex features pertinent to predicting protein interactions and interaction sites, geometric deep learning to reason over protein structures and predict complex structures, and generative modeling to design de novo protein assemblies. We also outline some of the outstanding challenges and promising new directions. Opportunities abound to discover novel interactions, elucidate their physical mechanisms, and engineer binders to modulate their functions using deep learning and, ultimately, unravel how protein interactions orchestrate complex cellular behaviors.
The rapid evolution of deep learning has significantly advanced the field of medical image analysis. However, despite these achievements, the further enhancement of deep learning models for medical image analysis faces a significant challenge due to the scarcity of large, well-annotated datasets. To address this issue, recent years have witnessed a growing emphasis on the development of data-efficient deep learning methods. This paper conducts a thorough review of data-efficient deep learning methods for medical image analysis. To this end, we categorize these methods based on the level of supervision they rely on, encompassing categories such as no supervision, inexact supervision, incomplete supervision, inaccurate supervision, and only limited supervision. We further divide these categories into finer subcategories. For example, we categorize inexact supervision into multiple instance learning and learning with weak annotations. Similarly, we categorize incomplete supervision into semi-supervised learning, active learning, and domain-adaptive learning and so on. Furthermore, we systematically summarize commonly used datasets for data efficient deep learning in medical image analysis and investigate future research directions to conclude this survey.