Bayesian inference was once a gold standard for learning with neural networks, providing accurate full predictive distributions and well calibrated uncertainty. However, scaling Bayesian inference techniques to deep neural networks is challenging due to the high dimensionality of the parameter space. In this paper, we construct low-dimensional subspaces of parameter space, such as the first principal components of the stochastic gradient descent (SGD) trajectory, which contain diverse sets of high performing models. In these subspaces, we are able to apply elliptical slice sampling and variational inference, which struggle in the full parameter space. We show that Bayesian model averaging over the induced posterior in these subspaces produces accurate predictions and well calibrated predictive uncertainty for both regression and image classification.
Neural networks predictions are unreliable when the input sample is out of the training distribution or corrupted by noise. Being able to detect such failures automatically is fundamental to integrate deep learning algorithms into robotics. Current approaches for uncertainty estimation of neural networks require changes to the network and optimization process, typically ignore prior knowledge about the data, and tend to make over-simplifying assumptions which underestimate uncertainty. To address these limitations, we propose a novel framework for uncertainty estimation. Based on Bayesian belief networks and Monte-Carlo sampling, our framework not only fully models the different sources of prediction uncertainty, but also incorporates prior data information, e.g. sensor noise. We show theoretically that this gives us the ability to capture uncertainty better than existing methods. In addition, our framework has several desirable properties: (i) it is agnostic to the network architecture and task; (ii) it does not require changes in the optimization process; (iii) it can be applied to already trained architectures. We thoroughly validate the proposed framework through extensive experiments on both computer vision and control tasks, where we outperform previous methods by up to 23% in accuracy.
In this paper, we present the Metamorphic Testing of an in-use deep learning based forecasting application. The application looks at the past data of system characteristics (e.g. `memory allocation') to predict outages in the future. We focus on two statistical / machine learning based components - a) detection of co-relation between system characteristics and b) estimating the future value of a system characteristic using an LSTM (a deep learning architecture). In total, 19 Metamorphic Relations have been developed and we provide proofs & algorithms where applicable. We evaluated our method through two settings. In the first, we executed the relations on the actual application and uncovered 8 issues not known before. Second, we generated hypothetical bugs, through Mutation Testing, on a reference implementation of the LSTM based forecaster and found that 65.9% of the bugs were caught through the relations.
Deep learning has achieved remarkable success in diverse applications; however, its use in solving partial differential equations (PDEs) has emerged only recently. Here, we present an overview of physics-informed neural networks (PINNs), which embed a PDE into the loss of the neural network using automatic differentiation. The PINN algorithm is simple, and it can be applied to different types of PDEs, including integro-differential equations, fractional PDEs, and stochastic PDEs. Moreover, from the implementation point of view, PINNs solve inverse problems as easily as forward problems. We propose a new residual-based adaptive refinement (RAR) method to improve the training efficiency of PINNs. For pedagogical reasons, we compare the PINN algorithm to a standard finite element method. We also present a Python library for PINNs, DeepXDE, which is designed to serve both as an education tool to be used in the classroom as well as a research tool for solving problems in computational science and engineering. Specifically, DeepXDE can solve forward problems given initial and boundary conditions, as well as inverse problems given some extra measurements. DeepXDE supports complex-geometry domains based on the technique of constructive solid geometry, and enables the user code to be compact, resembling closely the mathematical formulation. We introduce the usage of DeepXDE and its customizability, and we also demonstrate the capability of PINNs and the user-friendliness of DeepXDE for five different examples. More broadly, DeepXDE contributes to the more rapid development of the emerging Scientific Machine Learning field.
Multivariate time series with missing values are common in areas such as healthcare and finance, and have grown in number and complexity over the years. This raises the question whether deep learning methodologies can outperform classical data imputation methods in this domain. However, naive applications of deep learning fall short in giving reliable confidence estimates and lack interpretability. We propose a new deep sequential latent variable model for dimensionality reduction and data imputation. Our modeling assumption is simple and interpretable: the high dimensional time series has a lower-dimensional representation which evolves smoothly in time according to a Gaussian process. The non-linear dimensionality reduction in the presence of missing data is achieved using a VAE approach with a novel structured variational approximation. We demonstrate that our approach outperforms several classical and deep learning-based data imputation methods on high-dimensional data from the domains of computer vision and healthcare, while additionally improving the smoothness of the imputations and providing interpretable uncertainty estimates.
We present Point-Voxel CNN (PVCNN) for efficient, fast 3D deep learning. Previous work processes 3D data using either voxel-based or point-based NN models. However, both approaches are computationally inefficient. The computation cost and memory footprints of the voxel-based models grow cubically with the input resolution, making it memory-prohibitive to scale up the resolution. As for point-based networks, up to 80% of the time is wasted on structuring the sparse data which have rather poor memory locality, not on the actual feature extraction. In this paper, we propose PVCNN that represents the 3D input data in points to reduce the memory consumption, while performing the convolutions in voxels to reduce the irregular, sparse data access and improve the locality. Our PVCNN model is both memory and computation efficient. Evaluated on semantic and part segmentation datasets, it achieves much higher accuracy than the voxel-based baseline with 10x GPU memory reduction; it also outperforms the state-of-the-art point-based models with 7x measured speedup on average. Remarkably, the narrower version of PVCNN achieves 2x speedup over PointNet (an extremely efficient model) on part and scene segmentation benchmarks with much higher accuracy. We validate the general effectiveness of PVCNN on 3D object detection: by replacing the primitives in Frustrum PointNet with PVConv, it outperforms Frustrum PointNet++ by 2.4% mAP on average with 1.5x measured speedup and GPU memory reduction.
Generative neural network is a new category of neural networks and it has been widely utilized in applications such as content generation, unsupervised learning, segmentation and pose estimation. It typically involves massive computing-intensive deconvolution operations that cannot be fitted to conventional neural network processors directly. However, prior works mainly investigated specialized hardware architectures through intensive hardware modifications to the existing deep learning processors to accelerate deconvolution together with the convolution. In contrast, this work proposes a novel deconvolution implementation with a software approach and enables fast and efficient deconvolution execution on the legacy deep learning processors. Our proposed method reorganizes the computation of deconvolution and allows the deep learning processors to treat it as the standard convolution by splitting the original deconvolution filters into multiple small filters. Compared to prior acceleration schemes, the implemented acceleration scheme achieves 2.41x - 4.34x performance speedup and reduces the energy consumption by 27.7% - 54.5% on a set of realistic benchmarks. In addition, we also applied the deconvolution computing approach to the off-the-shelf commodity deep learning processors. The performance of deconvolution also exhibits significant performance speedup over prior deconvolution implementations.
Data selection methods, such as active learning and core-set selection, are useful tools for machine learning on large datasets. However, they can be prohibitively expensive to apply in deep learning because they depend on feature representations that need to be learned. In this work, we show that we can greatly improve the computational efficiency by using a small proxy model to perform data selection (e.g., selecting data points to label for active learning). By removing hidden layers from the target model, using smaller architectures, and training for fewer epochs, we create proxies that are an order of magnitude faster to train. Although these small proxy models have higher error rates, we find that they empirically provide useful signals for data selection. We evaluate this "selection via proxy" (SVP) approach on several data selection tasks across five datasets: CIFAR10, CIFAR100, ImageNet, Amazon Review Polarity, and Amazon Review Full. For active learning, applying SVP can give an order of magnitude improvement in data selection runtime (i.e., the time it takes to repeatedly train and select points) without significantly increasing the final error (often within 0.1%). For core-set selection on CIFAR10, proxies that are over 10x faster to train than their larger, more accurate targets can remove up to 50% of the data without harming the final accuracy of the target, leading to a 1.6x end-to-end training time improvement.
Large neural networks are very successful in various tasks. However, with limited data, the generalization capabilities of deep neural networks are also very limited. In this paper, we empirically start showing that intrinsically sparse neural networks with adaptive sparse connectivity, which by design have a strict parameter budget during the training phase, have better generalization capabilities than their fully-connected counterparts. Besides this, we propose a new technique to train these sparse models by combining the Sparse Evolutionary Training (SET) procedure with neurons pruning. Operated on MultiLayer Perceptron (MLP) and tested on 15 datasets, our proposed technique zeros out around 50% of the hidden neurons during training, while having a linear number of parameters to optimize with respect to the number of neurons. The results show a competitive classification and generalization performance.
Water supplies are crucial for the development of living beings. However, change in the hydrological process i.e. climate and land usage are the key issues. Sustaining water level and accurate estimating for dynamic conditions is a critical job for hydrologists, but predicting hydrological extremes is an open issue. In this paper, we proposed two deep learning techniques and three machine learning algorithms to predict stream flow, given the present climate conditions. The results showed that the Recurrent Neural Network (RNN) or Long Short-term Memory (LSTM), an artificial neural network based method, outperform other conventional and machine-learning algorithms for predicting stream flow. Furthermore, we analyzed that stream flow is directly affected by precipitation, land usage, and temperature. These indexes are critical, which can be used by hydrologists to identify the potential for stream flow. We make the dataset publicly available (https://github.com/sadaqat007/Dataset) so that others should be able to replicate and build upon the results published.
Machine learning algorithms generally suffer from a problem of explainability. Given a classification result from a model, it is typically hard to determine what caused the decision to be made, and to give an informative explanation. We explore a new method of generating counterfactual explanations, which instead of explaining why a particular classification was made explain how a different outcome can be achieved. This gives the recipients of the explanation a better way to understand the outcome, and provides an actionable suggestion. We show that the introduced method of Constrained Adversarial Examples (CADEX) can be used in real world applications, and yields explanations which incorporate business or domain constraints such as handling categorical attributes and range constraints.
While deep learning technologies for computer vision have developed rapidly since 2012, modeling of remote sensing systems has remained focused around human vision. In particular, remote sensing systems are usually constructed to optimize sensing cost-quality trade-offs with respect to human image interpretability. While some recent studies have explored remote sensing system design as a function of simple computer vision algorithm performance, there has been little work relating this design to the state-of-the-art in computer vision: deep learning with convolutional neural networks. We develop experimental systems to conduct this analysis, showing results with modern deep learning algorithms and recent overhead image data. Our results are compared to standard image quality measurements based on human visual perception, and we conclude not only that machine and human interpretability differ significantly, but that computer vision performance is largely self-consistent across a range of disparate conditions. This research is presented as a cornerstone for a new generation of sensor design systems which focus on computer algorithm performance instead of human visual perception.
Crater ellipticity determination is a complex and time consuming task that so far has evaded successful automation. We train a state of the art computer vision algorithm to identify craters in Lunar digital elevation maps and retrieve their sizes and 2D shapes. The computational backbone of the model is MaskRCNN, an "instance segmentation" general framework that detects craters in an image while simultaneously producing a mask for each crater that traces its outer rim. Our post-processing pipeline then finds the closest fitting ellipse to these masks, allowing us to retrieve the crater ellipticities. Our model is able to correctly identify 87\% of known craters in the longitude range we hid from the network during training and validation (test set), while predicting thousands of additional craters not present in our training data. Manual validation of a subset of these "new" craters indicates that a majority of them are real, which we take as an indicator of the strength of our model in learning to identify craters, despite incomplete training data. The crater size, ellipticity, and depth distributions predicted by our model are consistent with human-generated results. The model allows us to perform a large scale search for differences in crater diameter and shape distributions between the lunar highlands and maria, and we exclude any such differences with a high statistical significance. The predicted test set catalogue and trained model are available here: https://github.com/malidib/Craters_MaskRCNN/.
In the integrative analyses of omics data, it is often of interest to extract data representation from one data type that best reflect its relations with another data type. This task is traditionally fulfilled by linear methods such as canonical correlation analysis (CCA) and partial least squares (PLS). However, information contained in one data type pertaining to the other data type may be complex and in nonlinear form. Deep learning provides a convenient alternative to extract low-dimensional nonlinear data embedding. In addition, the deep learning setup can naturally incorporate the effects of clinical confounding factors into the integrative analysis. Here we report a deep learning setup, named Autoencoder-based Integrative Multi-omics data Embedding (AIME), to extract data representation for omics data integrative analysis. The method can adjust for confounder variables, achieve informative data embedding, rank features in terms of their contributions, and find pairs of features from the two data types that are related to each other through the data embedding. In simulation studies, the method was highly effective in the extraction of major contributing features between data types. Using a real microRNA-gene expression dataset, and a real DNA methylation-gene expression dataset, we show that AIME excluded the influence of confounders including batch effects, and extracted biologically plausible novel information. The R package based on Keras and the TensorFlow backend is available at https://github.com/tianwei-yu/AIME.
Rendering an accurate image of an isosurface in a volumetric field typically requires large numbers of data samples. Reducing the number of required samples lies at the core of research in volume rendering. With the advent of deep learning networks, a number of architectures have been proposed recently to infer missing samples in multi-dimensional fields, for applications such as image super-resolution and scan completion. In this paper, we investigate the use of such architectures for learning the upscaling of a low-resolution sampling of an isosurface to a higher resolution, with high fidelity reconstruction of spatial detail and shading. We introduce a fully convolutional neural network, to learn a latent representation generating a smooth, edge-aware normal field and ambient occlusions from a low-resolution normal and depth field. By adding a frame-to-frame motion loss into the learning stage, the upscaling can consider temporal variations and achieves improved frame-to-frame coherence. We demonstrate the quality of the network for isosurfaces which were never seen during training, and discuss remote and in-situ visualization as well as focus+context visualization as potential applications
Virtual reality (VR) offers immersive visualization and intuitive interaction. We leverage VR to enable any biomedical professional to deploy a deep learning (DL) model for image classification. While DL models can be powerful tools for data analysis, they are also challenging to understand and develop. To make deep learning more accessible and intuitive, we have built a virtual reality-based DL development environment. Within our environment, the user can move tangible objects to construct a neural network only using their hands. Our software automatically translates these configurations into a trainable model and then reports its resulting accuracy on a test dataset in real-time. Furthermore, we have enriched the virtual objects with visualizations of the model's components such that users can achieve insight about the DL models that they are developing. With this approach, we bridge the gap between professionals in different fields of expertise while offering a novel perspective for model analysis and data interaction. We further suggest that techniques of development and visualization in deep learning can benefit by integrating virtual reality.
Although there has been a rapid development of practical applications, theoretical explanations of deep learning are in their infancy. Deep learning performs a sophisticated coarse graining. Since coarse graining is a key ingredient of the renormalization group (RG), RG may provide a useful theoretical framework directly relevant to deep learning. In this study we pursue this possibility. A statistical mechanics model for a magnet, the Ising model, is used to train an unsupervised restricted Boltzmann machine (RBM). The patterns generated by the trained RBM are compared to the configurations generated through an RG treatment of the Ising model. Although we are motivated by the connection between deep learning and RG flow, in this study we focus mainly on comparing a single layer of a deep network to a single step in the RG flow. We argue that correlation functions between hidden and visible neurons are capable of diagnosing RG-like coarse graining. Numerical experiments show the presence of RG-like patterns in correlators computed using the trained RBMs. The observables we consider are also able to exhibit important differences between RG and deep learning.
Recent works have shown the effectiveness of randomized smoothing as a scalable technique for building neural network-based classifiers that are provably robust to $\ell_2$-norm adversarial perturbations. In this paper, we employ adversarial training to improve the performance of randomized smoothing. We design an adapted attack for smoothed classifiers, and we show how this attack can be used in an adversarial training setting to boost the provable robustness of smoothed classifiers. We demonstrate through extensive experimentation that our method consistently outperforms all existing provably $\ell_2$-robust classifiers by a significant margin on ImageNet and CIFAR-10, establishing the state-of-the-art for provable $\ell_2$-defenses. Moreover, we find that pre-training and semi-supervised learning boost adversarially trained smoothed classifiers even further. Our code and trained models are available at http://github.com/Hadisalman/smoothing-adversarial .
Esports have become major international sports with hundreds of millions of spectators. Esports games generate massive amounts of telemetry data. Using these to predict the outcome of esports matches has received considerable attention, but micro-predictions, which seek to predict events inside a match, is as yet unknown territory. Micro-predictions are however of perennial interest across esports commentators and audience, because they provide the ability to observe events that might otherwise be missed: esports games are highly complex with fast-moving action where the balance of a game can change in the span of seconds, and where events can happen in multiple areas of the playing field at the same time. Such events can happen rapidly, and it is easy for commentators and viewers alike to miss an event and only observe the following impact of events. In Dota 2, a player hero being killed by the opposing team is a key event of interest to commentators and audience. We present a deep learning network with shared weights which provides accurate death predictions within a five-second window. The network is trained on a vast selection of Dota 2 gameplay features and professional/semi-professional level match dataset. Even though death events are rare within a game (1\% of the data), the model achieves 0.377 precision with 0.725 recall on test data when prompted to predict which of any of the 10 players of either team will die within 5 seconds. An example of the system applied to a Dota 2 match is presented. This model enables real-time micro-predictions of kills in Dota 2, one of the most played esports titles in the world, giving commentators and viewers time to move their attention to these key events.
Purpose: To assess the utility of deep learning in the detection of geographic atrophy (GA) from color fundus photographs; secondary aim to explore potential utility in detecting central GA (CGA). Design: A deep learning model was developed to detect the presence of GA in color fundus photographs, and two additional models to detect CGA in different scenarios. Participants: 59,812 color fundus photographs from longitudinal follow up of 4,582 participants in the AREDS dataset. Gold standard labels were from human expert reading center graders using a standardized protocol. Methods: A deep learning model was trained to use color fundus photographs to predict GA presence from a population of eyes with no AMD to advanced AMD. A second model was trained to predict CGA presence from the same population. A third model was trained to predict CGA presence from the subset of eyes with GA. For training and testing, 5-fold cross-validation was employed. For comparison with human clinician performance, model performance was compared with that of 88 retinal specialists. Results: The deep learning models (GA detection, CGA detection from all eyes, and centrality detection from GA eyes) had AUC of 0.933-0.976, 0.939-0.976, and 0.827-0.888, respectively. The GA detection model had accuracy, sensitivity, specificity, and precision of 0.965, 0.692, 0.978, and 0.584, respectively. The CGA detection model had equivalent values of 0.966, 0.763, 0.971, and 0.394. The centrality detection model had equivalent values of 0.762, 0.782, 0.729, and 0.799. Conclusions: A deep learning model demonstrated high accuracy for the automated detection of GA. The AUC was non-inferior to that of human retinal specialists. Deep learning approaches may also be applied to the identification of CGA. The code and pretrained models are publicly available at https://github.com/ncbi-nlp/DeepSeeNet.