Input Domain Research Articles

AdaCM: Adaptive ColorMLP for Real-Time Universal Photo-Realistic Style Transfer

Photo-realistic style transfer aims at migrating the artistic style from an exemplar style image to a content image, producing a result image without spatial distortions or unrealistic artifacts. Impressive results have been achieved by recent deep models. However, deep neural network based methods are too expensive to run in real-time. Meanwhile, bilateral grid based methods are much faster but still contain artifacts like overexposure. In this work, we propose the Adaptive ColorMLP (AdaCM), an effective and efficient framework for universal photo-realistic style transfer. First, we find the complex non-linear color mapping between input and target domain can be efficiently modeled by a small multi-layer perceptron (ColorMLP) model. Then, in AdaCM, we adopt a CNN encoder to adaptively predict all parameters for the ColorMLP conditioned on each input content and style image pair. Experimental results demonstrate that AdaCM can generate vivid and high-quality stylization results. Meanwhile, our AdaCM is ultrafast and can process a 4K resolution image in 6ms on one V100 GPU.

Meshless power diagrams

The computation of power diagrams (or weighted Voronoi diagrams) is a fundamental task in computational geometry and computer graphics. To accomplish the computation, we provide a different way from the existing ones for lifting the weighted seeds to a set of points in the space of one dimension higher, then the power cells can be directly obtained by computing the intersections of the Voronoi cells of these lifting points and the original space. This property enables us to apply the method based on the k-nearest neighbors query to the generation of power diagrams. Each power cell is obtained by sequentially clipping the input domain using the bisectors between its seed and the k-nearest neighbors. Experimental results demonstrate that our method outperforms the state-of-the-art one based on regular triangulation in terms of efficiency for general cases in the 2D and 3D spaces.

Towards global neural network abstractions with locally-exact reconstruction

Neural networks are a powerful class of non-linear functions. However, their black-box nature makes it difficult to explain their behaviour and certify their safety. Abstraction techniques address this challenge by transforming the neural network into a simpler, over-approximated function. Unfortunately, existing abstraction techniques are slack, which limits their applicability to small local regions of the input domain. In this paper, we propose Global Interval Neural Network Abstractions with Center-Exact Reconstruction (GINNACER). Our novel abstraction technique produces sound over-approximation bounds over the whole input domain while guaranteeing exact reconstructions for any given local input. Our experiments show that GINNACER is several orders of magnitude tighter than state-of-the-art global abstraction techniques, while being competitive with local ones.

Level Set Learning with Pseudoreversible Neural Networks for Nonlinear Dimension Reduction in Function Approximation

Due to the curse of dimensionality and the limitation on training data, approximating high-dimensional functions is a very challenging task even for powerful deep neural networks. Inspired by the Nonlinear Level-set Learning (NLL) method, which uses the reversible residual network (RevNet), in this paper, we propose a new method of Dimension Reduction via Learning Level Sets (DRiLLS) for function approximation. Our method contains two major components: one is the pseudoreversible neural network (PRNN) module, which effectively transforms high-dimensional input variables to low-dimensional active variables, and the other is the synthesized regression module for approximating function values based on the transformed data in the low-dimensional space. The PRNN not only relaxes the invertibility constraint of the nonlinear transformation present in the NLL method due to the use of RevNet but also adaptively weights the influence of each sample and controls the sensitivity of the function to the learned active variables. The synthesized regression uses Euclidean distance in the input space to select neighboring samples whose projections on the space of active variables are used to perform local least-squares polynomial fitting. This helps to resolve numerical oscillation issues present in traditional local and global regressions. Extensive experimental results demonstrate that our DRiLLS method outperforms both the NLL and active subspace methods, especially when the target function possesses critical points in the interior of its input domain.

Theoretical study of effect of the geometrical parameters on the dynamic properties of the elastic rings of an air journal bearing

The paper investigates theoretically the effect of the geometry of the elastic rings of an air journal bearing on the elastic rings dynamic coefficients. The physical finite element method (FEM) model used to obtain the dynamic coefficients of the rings is discussed. A theoretical model is implemented to predict the effect of the geometrical parameters on the dynamic coefficients of the elastic rings. The effect of the geometrical parameters on the dynamic coefficients at different frequencies is studied using FEM. The elastic geometry that result in desired dynamic coefficients is demonstrated. Since predicting the dynamic coefficients for all possible ring geometries using FEM would be computationally expensive. A neural network (NN) is trained to predict the dynamic coefficients for all possible ring geometries generated by the different ring geometrical parameters within a given input domain. The NN results are compared to the experimentally verified FEM results and the results are in good agreement.

A black box based model for phase change heat exchanger in refrigeration system simulations using Kriging interpolation method

In the present study, a black box model for phase change heat exchangers is developed based on Kriging interpolation. The input and output of the heat exchanger black box model were established based on the mass-flow-guided method. Methods for the input domain determination of are established respectively for the application of system simulation and separate heat exchanger simulation. A training data set sampling method is established based on the Latin hypercube sampling method to achieve significant reduction in sample points while maintaining the integrity of the sample data characteristics. The heat exchanger calculation trend was determined by establishing a regression model; the accurate prediction of heat exchanger performance was achieved by making local corrections to the prediction results through a stochastic distribution process model. Validation results show that, for the condenser model, the average pressure drop prediction error is 0.25%, and the average heat transfer prediction error is 1.18%; for the evaporator model, the average pressure drop prediction error is 0.71%, and the average heat transfer prediction error is 0.80%. The model can achieve high prediction accuracy with a small amount of training data and has a good fitting performance in the range of general applications.

Exploring the reform of the environmental design education model under the OBE theory model of professional and creative depth integration

Abstract To be able to better improve the professional ability of environmental design students and cultivate applied technical talents, this paper proposes an analysis of the OBE theory for the reform of specialized and deeply integrated environmental design education. By establishing a mathematical model, the database is divided into different classes based on the similarity between data. The error value of the objective function is gradually reduced by repeated iterative operations. The data points are randomly selected to calculate the center distance, and the dense area of the input domain is obtained in the data set obeying Gaussian mixture distribution. Through iteration to find the optimal solution, adjust its speed and position so that the fitness function takes the maximum value. The speed is constantly revised according to its own experience and group information to enhance the global seeking ability, improve the local seeking ability of particles and avoid the oscillation phenomenon, and select the particle with the highest adaptability to calculate its average fluctuation value, to analyze the environmental design education reform under OBE theory. The result analysis shows that the correct rate of the algorithm analysis performance of this paper reaches 90.1%, which ensures more complete and accurate environmental design education data.

On-Device Deep Multi-Task Inference via Multi-Task Zipping

Future mobile devices are anticipated to perceive, understand and react to the world on their own by running multiple correlated deep neural networks locally on-device. Yet the complexity of deep models needs to be trimmed down to fit in mobile storage and memory. Previous studies squeeze the redundancy within a single model. In this work, we reduce the redundancy across multiple models. We propose Multi-Task Zipping (MTZ), a framework to merge correlated, pre-trained deep neural networks for cross-model compression. Central in MTZ is a layer-wise neuron sharing and incoming weight updating scheme that induces a minimal change in the error function. MTZ inherits information from eachmodel and demands light retraining to re-boost the accuracy of individual tasks. MTZ supports typical network layers and applies to inference tasks with different input domains. Evaluations show that MTZ can fully merge the hidden layers of two VGG-16 network. Moreover, MTZ can effectively merge nine residual networks for diverse inference tasks and models for different input domains. Withthe joint model merged by MTZ, the latency to switch between these tasks on memory-constrained devices is reduced by 8.71.

Regularized Multi-Output Gaussian Convolution Process With Domain Adaptation.

Multi-output Gaussian process (MGP) has been attracting increasing attention as a transfer learning method to model multiple outputs. Despite its high flexibility and generality, MGP still faces two critical challenges when applied to transfer learning. The first one is negative transfer, which occurs when there exists no shared information among the outputs. The second challenge is the input domain inconsistency, which is commonly studied in transfer learning yet not explored in MGP. In this paper, we propose a regularized MGP modeling framework with domain adaptation to overcome these challenges. More specifically, a sparse covariance matrix of MGP is constructed using convolution process, where penalization terms are added to adaptively select the most informative outputs for knowledge transfer. To deal with the domain inconsistency, a domain adaptation method is proposed by marginalizing inconsistent features and expanding missing features to align the input domains among different outputs. Statistical properties of the proposed method are provided to guarantee the performance practically and asymptotically. The proposed framework outperforms state-of-the-art benchmarks in comprehensive simulation studies and one real case study of a ceramic manufacturing process. The results demonstrate the effectiveness of our method in dealing with both the negative transfer and the domain inconsistency.

Ranking loss and sequestering learning for reducing image search bias in histopathology

Recently, deep learning has started to play an essential role in healthcare applications, including image search in digital pathology. Despite the recent progress in computer vision, significant issues remain for image searching in histopathology archives. A well-known problem is AI bias and lack of generalization. A more particular shortcoming of deep models is the ignorance toward search functionality. The former affects every model, the latter only search and matching. Due to the lack of ranking-based learning, researchers must train models based on the classification error and then use the resultant embedding for image search purposes. Moreover, deep models appear to be prone to internal bias even if using a large image repository of various hospitals. This paper proposes two novel ideas to improve image search performance. First, we use a ranking loss function to guide feature extraction toward the matching-oriented nature of the search. By forcing the model to learn the ranking of matched outputs, the representation learning is customized toward image search instead of learning a class label. Second, we introduce the concept of sequestering learning to enhance the generalization of feature extraction. By excluding the images of the input hospital from the matched outputs, i.e., sequestering the input domain, the institutional bias is reduced. The proposed ideas are implemented and validated through the largest public dataset of whole slide images. The experiments demonstrate superior results compare to the-state-of-art.

Input Distribution Coverage: Measuring Feature Interaction Adequacy in Neural Network Testing

Testing deep neural networks (DNNs) has garnered great interest in the recent years due to their use in many applications. Black-box test adequacy measures are useful for guiding the testing process in covering the input domain. However, the absence of input specifications makes it challenging to apply black-box test adequacy measures in DNN testing. The Input Distribution Coverage (IDC) framework addresses this challenge by using a variational autoencoder to learn a low dimensional latent representation of the input distribution, and then using that latent space as a coverage domain for testing. IDC applies combinatorial interaction testing on a partitioning of the latent space to measure test adequacy. Empirical evaluation demonstrates that IDC is cost-effective, capable of detecting feature diversity in test inputs, and more sensitive than prior work to test inputs generated using different DNN test generation methods. The findings demonstrate that IDC overcomes several limitations of white-box DNN coverage approaches by discounting coverage from unrealistic inputs and enabling the calculation of test adequacy metrics that capture the feature diversity present in the input space of DNNs.

Dynamic-Pix2Pix: Medical image segmentation by injecting noise to cGAN for modeling input and target domain joint distributions with limited training data

Learning to translate images from a source to a target domain with applications such as converting simple line drawing to oil painting has attracted significant attention. The quality of translated images is directly related to two crucial issues. First, the consistency of the output distribution with that of the target is essential. Second, the generated output should have a high correlation with the input. Conditional Generative Adversarial Networks, cGANs, are the most common models for translating images. The performance of a cGAN drops when we use a limited training dataset. In this work, we study the obstacles that prevent a U-shaped model from learning the target domain distribution from limited data by using noise as input. This study helps to increase the Pix2Pix (a form of cGAN) target distribution modeling ability from limited data with the help of dynamic neural network theory. Our model has two learning cycles. The model learns the correlation between input and ground truth in the first cycle. Then, the model's architecture is refined in the second cycle to learn the target distribution from noise input. These processes are executed in each iteration of the training procedure. Helping the cGAN learn the target distribution from noise input results in a better model generalization during the test time and allows the model to fit almost perfectly to the target domain distribution. As a result, our model surpasses the Pix2Pix model in segmenting HC18 and Montgomery's chest x-ray images. Both qualitative and Dice scores show the superiority of our model. Although our proposed method does not use thousand of additional data for pretraining, it produces comparable results for the in and out-domain generalization compared to the state-of-the-art methods.

Subset simulation for probabilistic computer models

Reliability analysis can be performed efficiently through subset simulation. Through Markov chain Monte Carlo, subset simulation progressively samples from the input domain of a performance function (typically a computer model) to find the failure domain, that is, the set of input configurations that result in an output higher than a prescribed threshold. Recently, a probabilistic framework for numerical analysis was proposed, whereby computation is treated as a statistical inference problem. The framework, called probabilistic numerics, treats the output of a computer code as a random variable. This paper presents a generalisation of subset simulation, which enables reliability analysis for probabilistic numerical models. The advantages and challenges of the method are discussed, and an example with industrial application is presented.

An INN-Based Solution for the Underdetermined Image Reconstruction Problem in EIT

Electrical impedance tomography (EIT) has been widely concerned in online nondestructive testing of carbon fiber-reinforced polymers (CFRPs) due to its significant advantages of high real-time performance, zero radiation, simple structure, and low cost. However, its inverse problem is highly underdetermined and nonlinear, thus limiting its imaging quality and application scope. Based on the architecture of an invertible neural network (INN), an INN image reconstruction algorithm with latent output variables is proposed in this article. With latent output variables, the information lost during training is reused, and a unique input can be determined by sampling latent output variables and predicted outputs. In addition, the forward and reverse processes are combined together to weaken the dependency of effective output and latent output variables, reduce the errors between input and output domains, and solve the underdetermination and nonlinearity problems of EIT. The simulation and experimental results showed that compared with the traditional algorithms and the convolution neural network (CNN) network algorithm, the INN can accurately reconstruct the damage location, yield a clearer damaged edge, and reduce artifacts. Noise-added simulations and prototype experiments showed that the INN achieves both good imaging quality and enhanced noise immunity, to improve the performance of the EIT system in CFRP laminate inspection.

Zone-wise surrogate modelling (ZSM) of univariate systems

Many complex systems display distinctly different behaviors across regions, zones, or sub-domains. A single surrogate may not suffice in modelling such systems. A better approach would be to identify the various zones and model them individually. In this work, we propose a zone-wise surrogate modelling (ZSM) algorithm to identify various zones in a system's input domain based on a user-specified acceptable goodness of fit and recommend the best surrogate for each identified zone from a library of potential surrogates. We have assessed ZSM on ten case studies involving complex 1-D functions and compared its modelling performance against some non-linear and piecewise models. We also show how ZSM can help in global optimization using five complex multimodal functions and found that a ZSM-based approach successfully identifies the true global optima of these functions. In future, we aim to extend ZSM for the modelling and optimization of complex multi-input single-output (MISO) systems.

Enhancing response of a protein conformational switch by using two disordered ligand binding domains.

Introduction: Protein conformational switches are often constructed by fusing an input domain, which recognizes a target ligand, to an output domain that establishes a biological response. Prior designs have employed binding-induced folding of the input domain to drive a conformational change in the output domain. Adding a second input domain can in principle harvest additional binding energy for performing useful work. It is not obvious, however, how to fuse two binding domains to a single output domain such that folding of both binding domains combine to effect conformational change in the output domain. Methods: Here, we converted the ribonuclease barnase (Bn) to a switchable enzyme by duplicating a C-terminal portion of its sequence and appending it to its N-terminus, thereby establishing a native fold (OFF state) and a circularly permuted fold (ON state) that competed for the shared core in a mutually exclusive fashion. Two copies of FK506 binding protein (FKBP), both made unstable by the V24A mutation and one that had been circularly permuted, were inserted into the engineered barnase at the junctions between the shared and duplicated sequences. Results: Rapamycin-induced folding of FK506 binding protein stretched and unfolded the native fold of barnase via the mutually exclusive folding effect, and rapamycin-induced folding of permuted FK506 binding protein stabilized the permuted fold of barnase by the loop-closure entropy principle. These folding events complemented each other to turn on RNase function. The cytotoxic switching mechanism was validated in yeast and human cells, and in vitro with purified protein. Discussion: Thermodynamic modeling and experimental results revealed that the dual action of loop-closure entropy and mutually exclusive folding is analogous to an engine transmission in which loop-closure entropy acts as the low gear, providing efficient switching at low ligand concentrations, and mutually exclusive folding acts as the high gear to allow the switch to reach its maximum response at high ligand concentrations.

Bayesian clustering of spatial functional data with application to a human mobility study during COVID-19

The coronavirus (COVID-19) global pandemic has made a significant impact on people’s social activities. Cell phone mobility data provide unique and rich information on studying this impact. The motivating dataset of this study is the daily leaving-home index data at Harris County in Texas provided by SafeGraph. To study changes in daily leaving-home index and how they relate to public policy and sociodemographic variables, we propose a new Bayesian wavelet model for modeling and clustering spatial functional data, where domain partitioning is achieved by operating on the spanning trees. The resulting clusters can have arbitrary shapes and are spatially contiguous in the input domain. An efficient tailored reversible jump Markov chain Monte Carlo algorithm is proposed to implement the model. The method is applied to the spatial functional data of the daily percentages of people who left home. We focus on the time period covering both lockdown and phased reopening in Texas during the COVID-19 pandemic and study the changing behaviors of those functional curves. By linking the clustering results with the sociodemographic information, we identify several covariates of census blocks that have a noticeable impact on the clustering patterns of people’s mobility behaviors.

Improving transparency and representational generalizability through parallel continual learning

This paper takes a parallel learning approach in continual learning scenarios. We define parallel continual learning as learning a sequence of tasks where the data for the previous tasks, whose distribution may have shifted over time, are also available while learning new tasks. We propose a parallel continual learning method by assigning subnetworks to each task, and simultaneously training only the assigned subnetworks on their corresponding tasks. In doing so, some parts of the network will be shared across multiple tasks. This is unlike the existing literature in continual learning which aims at learning incoming tasks sequentially, with the assumption that the data for the previous tasks have a fixed distribution. Our proposed method offers promises in: (1) Transparency in the network and in the relationship across tasks by enabling examination of the learned representations by independent and shared subnetworks, (2) Representation generalizability through sharing and training subnetworks on multiple tasks simultaneously. Our analysis shows that compared to many competing approaches such as continual learning, neural architecture search, and multi-task learning, parallel continual learning is capable of learning more generalizable representations. Also, (3)Parallel continual learning overcomes the common issue of catastrophic forgetting in continual learning algorithms. This is the first effort to train a neural network on multiple tasks and input domains simultaneously in a continual learning scenario. Our code is available at https://github.com/yours-anonym/PaRT.

COSMOS: A comprehensive framework for automatically generating domain-oriented test suite

Coverage criteria are satisfied by at least one examination of the test target, while many faults are not revealed by one execution. However, despite executing the faulty statement, the test result is correct in certain circumstances. Such coincidentally passing test cases execute the faulty statement but do not cause failures. This paper introduces the new concept of domain solver. Domain solvers attempt to detect the domain of inputs rather than a single input satisfying a path constraint. Domain coverage is a new metric to evaluate the relative accuracy of the detected domains. The promising point is that the proposed approach similarly treats nonlinear and linear constraints. Domain solver splits a path constraint into conjunctions of simple conditions comparing two expressions. Such a splitting simplifies the constraint-solving task to detect regions of the input space satisfying a comparison between two expressions. After finding a region, an improved version of an algorithm, update, is used to determine the domain of variables involved in the comparing expressions. Our proposed approach, COSMOS, is implemented using the Roslyn compiler. We compared COSMOS with well-known constraint solvers using various linear/nonlinear constraints. The results show that COSMOS improves the number of supported data types involved in a constraint and solves100% of the instances in which the other solvers fail. Besides, COSMOS achieves the best relative accuracy of 84% compared to the existing domain-oriented test suite generation approaches. Moreover, our experiment results illustrate that COSMOS improves the fault-finding capability of other existing test coverage criteria by detecting coincidentally correct test cases. Combining domain coverage and compiler as a service makes a powerful constraint satisfaction method outperforming the existing constraint solvers regarding the number of solved linear/nonlinear constraints. Augmenting other structural test coverage criteria with domain coverage reveals the coincidentally correct test cases.

Solving Inverse Problems With Deep Neural Networks - Robustness Included?

In the past five years, deep learning methods have become state-of-the-art in solving various inverse problems. Before such approaches can find application in safety-critical fields, a verification of their reliability appears mandatory. Recent works have pointed out instabilities of deep neural networks for several image reconstruction tasks. In analogy to adversarial attacks in classification, it was shown that slight distortions in the input domain may cause severe artifacts. The present article sheds new light on this concern, by conducting an extensive study of the robustness of deep-learning-based algorithms for solving underdetermined inverse problems. This covers compressed sensing with Gaussian measurements as well as image recovery from Fourier and Radon measurements, including a real-world scenario for magnetic resonance imaging (using the NYU-fastMRI dataset). Our main focus is on computing adversarial perturbations of the measurements that maximize the reconstruction error. A distinctive feature of our approach is the quantitative and qualitative comparison with total-variation minimization, which serves as a provably robust reference method. In contrast to previous findings, our results reveal that standard end-to-end network architectures are not only resilient against statistical noise, but also against adversarial perturbations. All considered networks are trained by common deep learning techniques, without sophisticated defense strategies.