Solving partial differential equations with deep learning makes it possible to reduce simulation times by multiple orders of magnitude and unlock scientific methods that typically rely on large numbers of sequential simulations, such as optimization and uncertainty quantification. Two of the largest challenges of adopting scientific AI for industrial problem settings is that training datasets must be simulated in advance and that neural networks for solving large-scale PDEs exceed the memory capabilities of current GPUs. We introduce a distributed programming API in the Julia language for simulating training data in parallel on the cloud and without requiring users to manage the underlying HPC infrastructure. In addition, we show that model-parallel deep learning based on domain decomposition allows us to scale neural networks for solving PDEs to commercial-scale problem settings and achieve above 90% parallel efficiency. Combining our cloud API for training data generation and model-parallel deep learning, we train large-scale neural networks for solving the 3D Navier-Stokes equation and simulating 3D CO2 flow in porous media. For the CO2 example, we simulate a training dataset based on a commercial carbon capture and storage (CCS) project and train a neural network for CO2 flow simulation on a 3D grid with over 2 million cells that is 5 orders of magnitudes faster than a conventional numerical simulator and 3,200 times cheaper.
Multi-spectral imagery is invaluable for remote sensing due to different spectral signatures exhibited by materials that often appear identical in greyscale and RGB imagery. Paired with modern deep learning methods, this modality has great potential utility in a variety of remote sensing applications, such as humanitarian assistance and disaster recovery efforts. State-of-the-art deep learning methods have greatly benefited from large-scale annotations like in ImageNet, but existing MSI image datasets lack annotations at a similar scale. As an alternative to transfer learning on such data with few annotations, we apply complex-valued co-domain symmetric models to classify real-valued MSI images. Our experiments on 8-band xView data show that our ultra-lean model trained on xView from scratch without data augmentations can outperform ResNet with data augmentation and modified transfer learning on xView. Our work is the first to demonstrate the value of complex-valued deep learning on real-valued MSI data.
Recent advances in deep learning have led to interest in training deep learning models on longitudinal healthcare records to predict a range of medical events, with models demonstrating high predictive performance. Predictive performance is necessary but insufficient, however, with explanations and reasoning from models required to convince clinicians for sustained use. Rigorous evaluation of explainability is often missing, as comparisons between models (traditional versus deep) and various explainability methods have not been well-studied. Furthermore, ground truths needed to evaluate explainability can be highly subjective depending on the clinician's perspective. Our work is one of the first to evaluate explainability performance between and within traditional (XGBoost) and deep learning (LSTM with Attention) models on both a global and individual per-prediction level on longitudinal healthcare data. We compared explainability using three popular methods: 1) SHapley Additive exPlanations (SHAP), 2) Layer-Wise Relevance Propagation (LRP), and 3) Attention. These implementations were applied on synthetically generated datasets with designed ground-truths and a real-world medicare claims dataset. We showed that overall, LSTMs with SHAP or LRP provides superior explainability compared to XGBoost on both the global and local level, while LSTM with dot-product attention failed to produce reasonable ones. With the explosion of the volume of healthcare data and deep learning progress, the need to evaluate explainability will be pivotal towards successful adoption of deep learning models in healthcare settings.
Predicting material properties base on micro structure of materials has long been a challenging problem. Recently many deep learning methods have been developed for material property prediction. In this study, we propose a crystal representation learning framework, Orbital CrystalNet, OCrystalNet, which consists of two parts: atomic descriptor generation and graph representation learning. In OCrystalNet, we first incorporate orbital field matrix (OFM) and atomic features to construct OFM-feature atomic descriptor, and then the atomic descriptor is used as atom embedding in the atom-bond message passing module which takes advantage of the topological structure of crystal graphs to learn crystal representation. To demonstrate the capabilities of OCrystalNet we performed a number of prediction tasks on Material Project dataset and JARVIS dataset and compared our model with other baselines and state of art methods. To further present the effectiveness of OCrystalNet, we conducted ablation study and case study of our model. The results show that our model have various advantages over other state of art models.
Fiber nonlinearity effects cap achievable rates and ranges in long-haul optical fiber communication links. Conventional nonlinearity compensation methods, such as perturbation theory-based nonlinearity compensation (PB-NLC), attempt to compensate for the nonlinearity by approximating analytical solutions to the signal propagation over optical fibers. However, their practical usability is limited by model mismatch and the immense computational complexity associated with the analytical computation of perturbation triplets and the nonlinearity distortion field. Recently, machine learning techniques have been used to optimise parameters of PB-based approaches, which traditionally have been determined analytically from physical models. It has been claimed in the literature that the learned PB-NLC approaches have improved performance and/or reduced computational complexity over their non-learned counterparts. In this paper, we first revisit the acclaimed benefits of the learned PB-NLC approaches by carefully carrying out a comprehensive performance-complexity analysis utilizing state-of-the-art complexity reduction methods. Interestingly, our results show that least squares-based PB-NLC with clustering quantization has the best performance-complexity trade-off among the learned PB-NLC approaches. Second, we advance the state-of-the-art of learned PB-NLC by proposing and designing a fully learned structure. We apply a bi-directional recurrent neural network for learning perturbation triplets that are alike those obtained from the analytical computation and are used as input features for the neural network to estimate the nonlinearity distortion field. Finally, we demonstrate through numerical simulations that our proposed fully learned approach achieves an improved performance-complexity trade-off compared to the existing learned and non-learned PB-NLC techniques.
Cancer claims millions of lives yearly worldwide. While many therapies have been made available in recent years, by in large cancer remains unsolved. Exploiting computational predictive models to study and treat cancer holds great promise in improving drug development and personalized design of treatment plans, ultimately suppressing tumors, alleviating suffering, and prolonging lives of patients. A wave of recent papers demonstrates promising results in predicting cancer response to drug treatments while utilizing deep learning methods. These papers investigate diverse data representations, neural network architectures, learning methodologies, and evaluations schemes. However, deciphering promising predominant and emerging trends is difficult due to the variety of explored methods and lack of standardized framework for comparing drug response prediction models. To obtain a comprehensive landscape of deep learning methods, we conducted an extensive search and analysis of deep learning models that predict the response to single drug treatments. A total of 60 deep learning-based models have been curated and summary plots were generated. Based on the analysis, observable patterns and prevalence of methods have been revealed. This review allows to better understand the current state of the field and identify major challenges and promising solution paths.
Deep learning has been widely used in various applications from different fields such as computer vision, natural language processing, etc. However, the training models are often manually developed via many costly experiments. This manual work usually requires substantial computing resources, time, and experience. To simplify the use of deep learning and alleviate human effort, automated deep learning has emerged as a potential tool that releases the burden for both users and researchers. Generally, an automatic approach should support the diversity of model selection and the evaluation should allow users to decide upon their demands. To that end, we propose a multi-objective end-to-end self-optimized approach for constructing deep learning models automatically. Experimental results on well-known datasets such as MNIST, Fashion, and Cifar10 show that our algorithm can discover various competitive models compared with the state-of-the-art approach. In addition, our approach also introduces multi-objective trade-off solutions for both accuracy and uncertainty metrics for users to make better decisions.
We present a Deep-Learning (DL) pipeline developed for the detection and characterization of astronomical sources within simulated Atacama Large Millimeter/submillimeter Array (ALMA) data cubes. The pipeline is composed of six DL models: a Convolutional Autoencoder for source detection within the spatial domain of the integrated data cubes, a Recurrent Neural Network (RNN) for denoising and peak detection within the frequency domain, and four Residual Neural Networks (ResNets) for source characterization. The combination of spatial and frequency information improves completeness while decreasing spurious signal detection. To train and test the pipeline, we developed a simulation algorithm able to generate realistic ALMA observations, i.e. both sky model and dirty cubes. The algorithm simulates always a central source surrounded by fainter ones scattered within the cube. Some sources were spatially superimposed in order to test the pipeline deblending capabilities. The detection performances of the pipeline were compared to those of other methods and significant improvements in performances were achieved. Source morphologies are detected with subpixel accuracies obtaining mean residual errors of $10^{-3}$ pixel ($0.1$ mas) and $10^{-1}$ mJy/beam on positions and flux estimations, respectively. Projection angles and flux densities are also recovered within $10\%$ of the true values for $80\%$ and $73\%$ of all sources in the test set, respectively. While our pipeline is fine-tuned for ALMA data, the technique is applicable to other interferometric observatories, as SKA, LOFAR, VLBI, and VLTI.
Advanced deep learning architectures consist of tens of fully connected and convolutional hidden layers, currently extended to hundreds, are far from their biological realization. Their implausible biological dynamics relies on changing a weight in a non-local manner, as the number of routes between an output unit and a weight is typically large, using the backpropagation technique. Here, a 3-layer tree architecture inspired by experimental-based dendritic tree adaptations is developed and applied to the offline and online learning of the CIFAR-10 database. The proposed architecture outperforms the achievable success rates of the 5-layer convolutional LeNet. Moreover, the highly pruned tree backpropagation approach of the proposed architecture, where a single route connects an output unit and a weight, represents an efficient dendritic deep learning.
The realization of complex classification tasks requires training of deep learning (DL) architectures consisting of tens or even hundreds of convolutional and fully connected hidden layers, which is far from the reality of the human brain. According to the DL rationale, the first convolutional layer reveals localized patterns in the input and large-scale patterns in the following layers, until it reliably characterizes a class of inputs. Here, we demonstrate that with a fixed ratio between the depths of the first and second convolutional layers, the error rates of the generalized shallow LeNet architecture, consisting of only five layers, decay as a power law with the number of filters in the first convolutional layer. The extrapolation of this power law indicates that the generalized LeNet can achieve small error rates that were previously obtained for the CIFAR-10 database using DL architectures. A power law with a similar exponent also characterizes the generalized VGG-16 architecture. However, this results in a significantly increased number of operations required to achieve a given error rate with respect to LeNet. This power law phenomenon governs various generalized LeNet and VGG-16 architectures, hinting at its universal behavior and suggesting a quantitative hierarchical time-space complexity among machine learning architectures. Additionally, the conservation law along the convolutional layers, which is the square-root of their size times their depth, is found to asymptotically minimize error rates. The efficient shallow learning that is demonstrated in this study calls for further quantitative examination using various databases and architectures and its accelerated implementation using future dedicated hardware developments.
While it has been well known in the ML community that deep learning models suffer from instability, the consequences for healthcare deployments are under characterised. We study the stability of different model architectures trained on electronic health records, using a set of outpatient prediction tasks as a case study. We show that repeated training runs of the same deep learning model on the same training data can result in significantly different outcomes at a patient level even though global performance metrics remain stable. We propose two stability metrics for measuring the effect of randomness of model training, as well as mitigation strategies for improving model stability.
Video analysis tasks such as action recognition have received increasing research interest with growing applications in fields such as smart healthcare, thanks to the introduction of large-scale datasets and deep learning-based representations. However, video models trained on existing datasets suffer from significant performance degradation when deployed directly to real-world applications due to domain shifts between the training public video datasets (source video domains) and real-world videos (target video domains). Further, with the high cost of video annotation, it is more practical to use unlabeled videos for training. To tackle performance degradation and address concerns in high video annotation cost uniformly, the video unsupervised domain adaptation (VUDA) is introduced to adapt video models from the labeled source domain to the unlabeled target domain by alleviating video domain shift, improving the generalizability and portability of video models. This paper surveys recent progress in VUDA with deep learning. We begin with the motivation of VUDA, followed by its definition, and recent progress of methods for both closed-set VUDA and VUDA under different scenarios, and current benchmark datasets for VUDA research. Eventually, future directions are provided to promote further VUDA research. The repository of this survey is provided at https://github.com/xuyu0010/awesome-video-domain-adaptation.
Sparse-view computed tomography (CT) can be used to reduce radiation dose greatly but is suffers from severe image artifacts. Recently, the deep learning based method for sparse-view CT reconstruction has attracted a major attention. However, neural networks often have a limited ability to remove the artifacts when they only work in the image domain. Deep learning-based sinogram processing can achieve a better anti-artifact performance, but it inevitably requires feature maps of the whole image in a video memory, which makes handling large-scale or three-dimensional (3D) images rather challenging. In this paper, we propose a patch-based denoising diffusion probabilistic model (DDPM) for sparse-view CT reconstruction. A DDPM network based on patches extracted from fully sampled projection data is trained and then used to inpaint down-sampled projection data. The network does not require paired full-sampled and down-sampled data, enabling unsupervised learning. Since the data processing is patch-based, the deep learning workflow can be distributed in parallel, overcoming the memory problem of large-scale data. Our experiments show that the proposed method can effectively suppress few-view artifacts while faithfully preserving textural details.
Deep learning has emerged as a promising paradigm to give access to highly accurate predictions of molecular and materials properties. A common short-coming shared by current approaches, however, is that neural networks only give point estimates of their predictions and do not come with predictive uncertainties associated with these estimates. Existing uncertainty quantification efforts have primarily leveraged the standard deviation of predictions across an ensemble of independently trained neural networks. This incurs a large computational overhead in both training and prediction that often results in order-of-magnitude more expensive predictions. Here, we propose a method to estimate the predictive uncertainty based on a single neural network without the need for an ensemble. This allows us to obtain uncertainty estimates with virtually no additional computational overhead over standard training and inference. We demonstrate that the quality of the uncertainty estimates matches those obtained from deep ensembles. We further examine the uncertainty estimates of our methods and deep ensembles across the configuration space of our test system and compare the uncertainties to the potential energy surface. Finally, we study the efficacy of the method in an active learning setting and find the results to match an ensemble-based strategy at order-of-magnitude reduced computational cost.
We address Lagrangian drift simulation in geophysical dynamics and explore deep learning approaches to overcome known limitations of state-of-the-art model-based and Markovian approaches in terms of computational complexity and error propagation. We introduce a novel architecture, referred to as DriftNet, inspired from the Eulerian Fokker-Planck representation of Lagrangian dynamics. Numerical experiments for Lagrangian drift simulation at the sea surface demonstrates the relevance of DriftNet w.r.t. state-of-the-art schemes. Benefiting from the fully-convolutional nature of Drift-Net, we explore through a neural inversion how to diagnose modelderived velocities w.r.t. real drifter trajectories.
Deep Learning and big data have shown tremendous success in bioinformatics and computational biology in recent years; artificial intelligence methods have also significantly contributed in the task of protein function classification. This review paper analyzes the recent developments in approaches for the task of predicting protein function using deep learning. We explain the importance of determining the protein function and why automating the following task is crucial. Then, after reviewing the widely used deep learning techniques for this task, we continue our review and highlight the emergence of the modern State of The Art (SOTA) deep learning models which have achieved groundbreaking results in the field of computer vision, natural language processing and multi-modal learning in the last few years. We hope that this review will provide a broad view of the current role and advances of deep learning in biological sciences, especially in predicting protein function tasks and encourage new researchers to contribute to this area.
Variable selection problem for the nonlinear Cox regression model is considered. In survival analysis, one main objective is to identify the covariates that are associated with the risk of experiencing the event of interest. The Cox proportional hazard model is being used extensively in survival analysis in studying the relationship between survival times and covariates, where the model assumes that the covariate has a log-linear effect on the hazard function. However, this linearity assumption may not be satisfied in practice. In order to extract a representative subset of features, various variable selection approaches have been proposed for survival data under the linear Cox model. However, there exists little literature on variable selection for the nonlinear Cox model. To break this gap, we extend the recently developed deep learning-based variable selection model LassoNet to survival data. Simulations are provided to demonstrate the validity and effectiveness of the proposed method. Finally, we apply the proposed methodology to analyze a real data set on diffuse large B-cell lymphoma.
While deep learning models have replaced hand-designed features across many domains, these models are still trained with hand-designed optimizers. In this work, we leverage the same scaling approach behind the success of deep learning to learn versatile optimizers. We train an optimizer for deep learning which is itself a small neural network that ingests gradients and outputs parameter updates. Meta-trained with approximately four thousand TPU-months of compute on a wide variety of optimization tasks, our optimizer not only exhibits compelling performance, but optimizes in interesting and unexpected ways. It requires no hyperparameter tuning, instead automatically adapting to the specifics of the problem being optimized. We open source our learned optimizer, meta-training code, the associated train and test data, and an extensive optimizer benchmark suite with baselines at velo-code.github.io.
Validation metrics are a key prerequisite for the reliable tracking of scientific progress and for deciding on the potential clinical translation of methods. While recent initiatives aim to develop comprehensive theoretical frameworks for understanding metric-related pitfalls in image analysis problems, there is a lack of experimental evidence on the concrete effects of common and rare pitfalls on specific applications. We address this gap in the literature in the context of colon cancer screening. Our contribution is twofold. Firstly, we present the winning solution of the Endoscopy computer vision challenge (EndoCV) on colon cancer detection, conducted in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2022. Secondly, we demonstrate the sensitivity of commonly used metrics to a range of hyperparameters as well as the consequences of poor metric choices. Based on comprehensive validation studies performed with patient data from six clinical centers, we found all commonly applied object detection metrics to be subject to high inter-center variability. Furthermore, our results clearly demonstrate that the adaptation of standard hyperparameters used in the computer vision community does not generally lead to the clinically most plausible results. Finally, we present localization criteria that correspond well to clinical relevance. Our work could be a first step towards reconsidering common validation strategies in automatic colon cancer screening applications.
Very large deep learning models trained using gradient descent are remarkably resistant to memorization given their huge capacity, but are at the same time capable of fitting large datasets of pure noise. Here methods are introduced by which models may be trained to memorize datasets that normally are generalized. We find that memorization is difficult relative to generalization, but that adding noise makes memorization easier. Increasing the dataset size exaggerates the characteristics of that dataset: model access to more training samples makes overfitting easier for random data, but somewhat harder for natural images. The bias of deep learning towards generalization is explored theoretically, and we show that generalization results from a model's parameters being attracted to points of maximal stability with respect to that model's inputs during gradient descent.