Generate Training Sets Research Articles

Optimizing SEM parameters for segmentation with AI – Part 2: Designing and training a regression model

Selecting the best microscope parameters for optimal image quality currently relies on microscopists; there exist no procedures or guidelines for tuning parameters to ensure the desired image quality is achieved. More importantly, for quantitative analysis purposes, adequate image quality for segmentation should be prioritized. This paper is the second of two parts, describing a regression model, mixed input, multiple output with Keras TensorFlow, trained to predict the beam energy and probe current, two important parameters for image quality. Specifically, parameters are predicted to optimize the image quality for segmentation, using a generated training set, as described in Part 1 of this paper. Model performance is then tested on models trained with multiple different training sets, and with different proportions of simulated and acquired data. First, to examine the impact of the training set on the prediction accuracy and then, to evaluate the importance of including real data during training. The model successfully predicted the beam energy and probe current to set on the microscope to improve image quality for segmentation. Models trained with both simulated and acquired data performed the best, as evaluated by their efficacy at improving the image quality for feature segmentation.

Sparse linear array synthesis using artificial neural network and nonconvex optimization

AbstractThe concept of sparse array synthesis has attracted considerable attention due to its potential to reduce hardware costs in array design and enhance the efficiency of array aperture utilization. In contrast to conventional artificial neural networks (ANNs), a more efficient encoder–decoding ANN framework is proposed in this letter for designing sparse linear arrays. The decoder is pretrained with a randomly generated training set, enabling its ability to rapidly predictiction of far‐field patterns. This process considers the actual coupling effect among the antenna elements. The output of the encoder is used as the input of the decoder to synthesize the sparse arrays by providing the desired pattern to the encoder. During the training process, a novel ‐norm minimization nonconvex optimization method is integrated into the proposed framework to reduce the number of activated elements. The effectiveness and efficiency of the method are demonstrated through numerical and full‐wave simulations for synthesizing various linear sparse arrays, including large‐scale arrays. In comparison to existing algorithms, our proposed method exhibits superior computational efficiency and predicted accuracy with fewer elements.

Active Crowdsourcing for Multilabel Annotation.

Multilabel annotation is a critical step to generate training sets when learning classification models in various application domains, but asking domain experts to provide labels is usually time-consuming and expensive, which cannot meet the current requirement of the fast evolution of the models in the big data era. Although crowdsourcing provides a fast solution to acquire labels for multilabel learning, it faces the risk of high data acquisition cost and low label quality. This article proposes a novel one-coin label-dependent active crowdsourcing (OCLDAC) method to iteratively query noisy labels from crowd workers and learn multilabel classification models. In each iteration of active learning, integrated labels of instances are first inferred by a novel one-coin label-dependent model, which utilizes a mixture of multiple independent Bernoulli distributions to explore and exploit correlations among the labels to increase the accuracy of truth inference. Then, instances, labels, and workers are selected according to the novel strategies that incorporate the distribution of noisy labels, the prediction probability of learning models, label correlations, and the reliability of crowd workers. Simulations on eight multilabel datasets and evaluation on one real-world crowdsourcing dataset consistently show that the proposed OCLDAC significantly outperforms the state-of-the-art methods and their variants.

As Simple as Possible but No Simpler: Optimizing the Performance of Neural Net Emulators for Galaxy SED Fitting

Artificial neural network emulators have been demonstrated to be a very computationally efficient method to rapidly generate galaxy spectral energy distributions, for parameter inference or otherwise. Using a highly flexible and fast mathematical structure, they can learn the nontrivial relationship between input galaxy parameters and output observables. However, they do so imperfectly, and small errors in flux prediction can yield large differences in recovered parameters. In this work, we investigate the relationship between an emulator’s execution time, uncertainties, correlated errors, and ability to recover accurate posteriors. We show that emulators can recover consistent results to traditional fits, with a precision of 25%–40% in posterior medians for stellar mass, stellar metallicity, star formation rate, and stellar age. We find that emulation uncertainties scale with an emulator’s width N as ∝N −1, while execution time scales as ∝N 2, resulting in an inherent tradeoff between execution time and emulation uncertainties. We also find that emulators with uncertainties smaller than observational uncertainties are able to recover accurate posteriors for most parameters without a significant increase in catastrophic outliers. Furthermore, we demonstrate that small architectures can produce flux residuals that have significant correlations, which can create dangerous systematic errors in colors. Finally, we show that the distributions chosen for generating training sets can have a large effect on an emulator’s ability to accurately fit rare objects. Selecting the optimal architecture and training set for an emulator will minimize the computational requirements for fitting near-future large-scale galaxy surveys. We release our emulators on GitHub (http://github.com/elijahmathews/MathewsEtAl2023).

Interpretable Deep Learning Applied to Rip Current Detection and Localization

A rip current is a strong, localized current of water which moves along and away from the shore. Recent studies have suggested that drownings due to rip currents are still a major threat to beach safety. Identification of rip currents is important for lifeguards when making decisions on where to designate patrolled areas. The public also require information while deciding where to swim when lifeguards are not on patrol. In the present study we present an artificial intelligence (AI) algorithm that both identifies whether a rip current exists in images/video, and also localizes where that rip current occurs. While there have been some significant advances in AI for rip current detection and localization, there is a lack of research ensuring that an AI algorithm can generalize well to a diverse range of coastal environments and marine conditions. The present study made use of an interpretable AI method, gradient-weighted class-activation maps (Grad-CAM), which is a novel approach for amorphous rip current detection. The training data/images were diverse and encompass rip currents in a wide variety of environmental settings, ensuring model generalization. An open-access aerial catalogue of rip currents were used for model training. Here, the aerial imagery was also augmented by applying a wide variety of randomized image transformations (e.g., perspective, rotational transforms, and additive noise), which dramatically improves model performance through generalization. To account for diverse environmental settings, a synthetically generated training set, containing fog, shadows, and rain, was also added to the rip current images, thus increased the training dataset approximately 10-fold. Interpretable AI has dramatically improved the accuracy of unbounded rip current detection, which can correctly classify and localize rip currents about 89% of the time when validated on independent videos from surf-cameras at oblique angles. The novelty also lies in the ability to capture some shape characteristics of the amorphous rip current structure without the need of a predefined bounding box, therefore enabling the use of remote technology like drones. A comparison with well-established coastal image processing techniques is also presented via a short discussion and easy reference table. The strengths and weaknesses of both methods are highlighted and discussed.

Detecting non-natural language artifacts for de-noising bug reports

Textual documents produced in the software engineering process are a popular target for natural language processing (NLP) and information retrieval (IR) approaches. However, issue tickets often contain artifacts such as code snippets, log outputs and stack traces. These artifacts not only inflate the issue ticket sizes, but also can this noise constitute a real problem for some NLP approaches, and therefore has to be removed in the pre-processing of some approaches. In this paper, we present a machine learning based approach to classify textual content into natural language and non-natural language artifacts at line level. We show how data from GitHub issue trackers can be used for automated training set generation, and present a custom preprocessing approach for the task of artifact removal. The training sets are automatically created from Markdown annotated issue tickets and project documentation files. We use these generated training sets to train a Markdown agnostic model that is able to classify un-annotated content. We evaluate our approach on issue tickets from projects written in C++, Java, JavaScript, PHP, and Python. Our approach achieves ROC-AUC scores between 0.92 and 0.96 for language-specific models. A multi-language model trained on the issue tickets of all languages achieves ROC-AUC scores between 0.92 and 0.95. The provided models are intended to be used as noise reduction pre-processing steps for NLP and IR approaches working on issue tickets.

Feature transforms for image data augmentation

A problem with convolutional neural networks (CNNs) is that they require large datasets to obtain adequate robustness; on small datasets, they are prone to overfitting. Many methods have been proposed to overcome this shortcoming with CNNs. In cases where additional samples cannot easily be collected, a common approach is to generate more data points from existing data using an augmentation technique. In image classification, many augmentation approaches utilize simple image manipulation algorithms. In this work, we propose some new methods for data augmentation based on several image transformations: the Fourier transform (FT), the Radon transform (RT), and the discrete cosine transform (DCT). These and other data augmentation methods are considered in order to quantify their effectiveness in creating ensembles of neural networks. The novelty of this research is to consider different strategies for data augmentation to generate training sets from which to train several classifiers which are combined into an ensemble. Specifically, the idea is to create an ensemble based on a kind of bagging of the training set, where each model is trained on a different training set obtained by augmenting the original training set with different approaches. We build ensembles on the data level by adding images generated by combining fourteen augmentation approaches, with three based on FT, RT, and DCT, proposed here for the first time. Pretrained ResNet50 networks are finetuned on training sets that include images derived from each augmentation method. These networks and several fusions are evaluated and compared across eleven benchmarks. Results show that building ensembles on the data level by combining different data augmentation methods produce classifiers that not only compete competitively against the state-of-the-art but often surpass the best approaches reported in the literature.

A BP Neural Network‐Based GIS‐Data‐Driven Automated Valuation Framework for Benchmark Land Price

The automated valuation of benchmark land price plays an essential role in regulating land demand in Chinese real‐estate market as the big data are currently accumulated rapidly. However, this problem becomes highly challenging due to the multidimension, large volume, and nonlinearity of the land price‐influencing factors. In this paper, an effective data‐driven automated valuation framework is proposed for valuing real estate assets by combining a GIS (geographic information system) and neural network technologies. This framework can automatically obtain the values of spatial factors affecting land price from GIS and generate training set data for training the neural network to identify the complex relationship between all kinds of factors and benchmark land prices. The effectiveness and universality of the framework is verified via the data of benchmark land prices in Wuhan. The framework can be applied for automated benchmark land price valuation in other cities.

Efficient Prediction of Human Motion for Real-Time Robotics Applications With Physics-Inspired Neural Networks

Generating accurate and efficient predictions for the motion of the humans present in the scene is key to the development of effective motion planning algorithms for robots moving in promiscuous areas, where wrong planning decisions could generate safety hazard or simply make the presence of the robot "socially" unacceptable. Our approach to predict human motion is based on a neural network of a peculiar kind. Contrary to conventional deep neural networks, our network embeds in its structure the popular Social Force Model, a dynamic equation describing the motion in physical terms. This choice allows us to concentrate the learning phase in the aspects, which are really unknown (i.e., the model's parameters) and to keep the structure of the network simple and manageable. As a result, we are able to obtain a good prediction accuracy with a small synthetically generated training set, and the accuracy remains acceptable even when the network is applied in scenarios quite different from those for which it was trained. Finally, the choices of the network are "explainable", as they can be interpreted in physical terms. Comparative and experimental results prove the effectiveness of the proposed approach.

Аугментация обучающего набора при обучении нейросетевой языковой модели для наполнения онтологии

This paper is a continuation of the research focused on solving the problem of ontology population using training on an automatically generated training set and the subsequent use of a neural-network language model for analyzing texts in order to discover new concepts to add to the ontology. The article is devoted to the text data augmentation - increasing the size of the training set by modification of its samples. Along with this, a solution to the problem of clarifying concepts (i.e. adjusting their boundaries in sentences), which were found during the automatic formation of the training set, is considered. A brief overview of existing approaches to text data augmentation, as well as approaches to extracting so-called nested named entities (nested NER), is presented. A procedure is proposed for clarifying the boundaries of the discovered concepts of the training set and its augmentation for subsequent training a neural-network language model in order to identify new concepts of ontology in the domain texts. The results of the experimental evaluation of the trained model and the main directions of further research are considered.

Argos: A toolkit for tracking multiple animals in complex visual environments.

Automatically tracking the positions of multiple animals is often necessary for studying behaviours. This task involves multiple object tracking, a challenging problem in computer vision. Recent advances in machine learning applied to video analysis have been helpful for animal tracking. However, existing tools work well only in homogeneous environments with uniform illumination, features rarely found in natural settings. Moreover, available algorithms cannot effectively process discontinuities in animal motion such as sudden jumps, thus requiring laborious manual review.Here we present Argos, a software toolkit for tracking multiple animals in inhomogeneous environments. Argos includes tools for compressing videos based on animal movement, for generating training sets for a convolutional neural network (CNN) to detect animals, for tracking multiple animals in a video and for facilitating review and correction of the tracks manually, with simple graphical user interfaces.We demonstrate that Argos can help reduce the amount of video data to be stored and analysed, speed up analysis and allow analysing difficult and ambiguous conditions in a scene.Thus, Argos supports multiple approaches to animal tracking suited for varying recording conditions and available computational resources. Together, these tools allow the recording and tracking of movements of multiple markerless animals in inhomogeneous environments over many hours.

Cortical Thickness from MRI to Predict Conversion from Mild Cognitive Impairment to Dementia in Parkinson Disease: A Machine Learning-based Model.

Background Group comparison results associating cortical thinning and Parkinson disease (PD) dementia (PDD) are limited in their application to clinical settings. Purpose To investigate whether cortical thickness from MRI can help predict conversion from mild cognitive impairment (MCI) to dementia in PD at an individual level using a machine learning-based model. Materials and Methods In this retrospective study, patients with PD and MCI who underwent MRI from September 2008 to November 2016 were included. Features were selected from clinical and cortical thickness variables in 10 000 randomly generated training sets. Features selected 5000 times or more were used to train random forest and support vector machine models. Each model was trained and tested in 10 000 randomly resampled data sets, and a median of 10 000 areas under the receiver operating characteristic curve (AUCs) was calculated for each. Model performances were validated in an external test set. Results Forty-two patients progressed to PDD (converters) (mean age, 71 years ± 6 [standard deviation]; 22 women), and 75 patients did not progress to PDD (nonconverters) (mean age, 68 years ± 6; 40 women). Four PDD converters (mean age, 74 years ± 10; four men) and 20 nonconverters (mean age, 67 years ± 7; 11 women) were included in the external test set. Models trained with cortical thickness variables (AUC range, 0.75-0.83) showed fair to good performances similar to those trained with clinical variables (AUC range, 0.70-0.81). Model performances improved when models were trained with both variables (AUC range, 0.80-0.88). In pair-wise comparisons, models trained with both variables more frequently showed better performance than others in all model types. The models trained with both variables were successfully validated in the external test set (AUC range, 0.69-0.84). Conclusion Cortical thickness from MRI helped predict conversion from mild cognitive impairment to dementia in Parkinson disease at an individual level, with improved performance when integrated with clinical variables. © RSNA, 2021 Online supplemental material is available for this article. See also the editorial by Port in this issue.

Towards closing the gap in weakly supervised semantic segmentation with DCNNs: Combining local and global models

Generating training sets for Deep Convolutional Neural Networks (DCNNs) is a bottleneck for modern real-world applications. This is a demanding task for applications where annotating training data is costly, such as in semantic segmentation. In the literature, there is still a gap between the performance achieved by a network trained on full and on weak annotations. In this paper, we establish a simple and natural strategy to measure this gap and to identify the components necessary to reduce it. On scribbles, we establish new state-of-the-art results: we obtain a mIoU of 75.6% without, and 75.7% with CRF post-processing. We reduce the gap by 64.2% whereas the current state-of-the-art reduces it only by 57.5%. Thanks to a formal reformulation of the weak supervision problem, a systematic study of the different components involved, and an original experimental strategy, we unravel a counter-intuitive mechanism analog to the philosophy of ensemble learning. This strategy is simple and amenable to generalizations to other weakly-supervised scenarios: averaging poor local predicted annotations with a generic naive baseline and reusing them for training a DCNN yields new state-of-the-art results. We show that our strategy accommodates effortlessly other pixel-level weak annotations such as bounding boxes and remains competitive.

Fast Multiple Edge Response Method Based on the Design of Experiment and Machine Learning

The multiple edge response (MER) method can accurately capture the nonlinear characteristics in signal integrity simulations. However, transient simulation is time-consuming to obtain all the edge responses for high-order MER. A fast method based on the design of experiment (DoE) and machine learning is proposed in this letter to obtain all the edge responses for a severe nonlinear link. The DoE is used to generate a few training sets for an artificial neural network (ANN). The ANN uses the generated training sets to build a precise model. The developed ANN model predicts the remaining edge responses. The proposed method takes 9.8 min to obtain all the responses, which is 92 min for transient simulations for the ninth-order MER. The results show that the eye height and width error between the proposed method and measurement is less than 5%.

Visual perception-based fault diagnosis in froth flotation using statistical approaches

Froth flotation is an important mineral concentration technique. Faulty conditions in flotation processes may cause the huge waste of mineral resources and reagents, and consequently, may lead to deterioration in terms of benefits of flotation plants. In this paper, we propose a computer vision-aided fault detection and diagnosis approach for froth flotation. Specifically, a joint Gabor texture feature based on the Copula model is designed to describe froth images; a rejection sampling technique is developed to generate training sets from the quality distribution of real flotation products, and then an isolation forest-based fault detector is learned; and a fault diagnosis model based on spline regression is developed for root cause identification. Simulation experiments conducted on the historical industry data show that the proposed strategy has better performance than the alternative methods. Thereafter, the entire framework has been tested on a lead-zinc flotation plant in China. Experimental results have demonstrated the effectiveness of the proposed method.

Active learning of many-body configuration space: Application to the Cs+-water MB-nrg potential energy function as a case study.

The efficient selection of representative configurations that are used in high-level electronic structure calculations needed for the development of many-body molecular models poses a challenge to current data-driven approaches to molecular simulations. Here, we introduce an active learning (AL) framework for generating training sets corresponding to individual many-body contributions to the energy of an N-body system, which are required for the development of MB-nrg potential energy functions (PEFs). Our AL framework is based on uncertainty and error estimation and uses Gaussian process regression to identify the most relevant configurations that are needed for an accurate representation of the energy landscape of the molecular system under examination. Taking the Cs+-water system as a case study, we demonstrate that the application of our AL framework results in significantly smaller training sets than previously used in the development of the original MB-nrg PEF, without loss of accuracy. Considering the computational cost associated with high-level electronic structure calculations, our AL framework is particularly well-suited to the development of many-body PEFs, with chemical and spectroscopic accuracy, for molecular-level computer simulations from the gas to the condensed phase.

LPI Radar Waveform Recognition Based on Features from Multiple Images.

Detecting and classifying the modulation type of the intercepted noisy LPI (low probability of intercept) radar signals in real-time is a necessary survival technique in the electronic intelligence systems. Most radar signals have been designed to have LPI properties; therefore, the LPI radar waveform recognition technique (LWRT) has recently gained increasing attention. In this paper, we propose a multiple feature images joint decision (MFIJD) model with two different feature extraction structures that fully extract the pixel feature to obtain the pre-classification results of each feature image for the non-stationary characteristics of most LPI radar signals. The core technology of this model is combining the short-time autocorrelation feature image, double short-time autocorrelation feature image and the original signal time-frequency image (TFI) simultaneously input into the hybrid model classifier, which is suitable for non-stationary signals, and it has higher universality. We demonstrate the performance of MFIJD by simulating 11 types of the signals defined in this paper and generating training sets and test sets. The comparison with the literature shows that the proposed methods not only has a high universality for LPI radar signals, but also better adapts to LPI radar waveform recognition at low SNR (signal to noise ratio) environment. The overall recognition rate of the method reaches 87.7% when the SNR is −6 dB.

Use of Artificial Neural Network for Initial design of steel structures

Design of steel structures is an iterative process that often turns complicated when requires initial guess. A good initial design guess that comes from past design experience can considerably reduce the number of subsequent analysis and design cycles. It is difficult to form declarative rules to express human intuitions and past experience. Another problem in design process is arriving at optimal solution, where optimization process in most cases is computationally expensive and time consuming. Artificial Neural Network (ANN) is a promising computational model that can perform cognitive tasks, such as learning and optimization. This paper focuses on application of ANN in design so as to make design process more efficient. The aim is to train ANN to arrive at optimal solution by considering design constraints along with practical. This is illustrated with an example of design of compression member which primarily requires initial guess. The economy in design is tried to obtain by providing training set of optimal solutions to a multilayer neural network. The design process followed to generate training set is in accordance with IS: 800-2007. The solutions given by ANN are approximate but have reasonable agreement with the expected solutions.

Object-Based and Rule-Based Classification of Synthetic Aperture Radar Images

Since thousands of years, the land is the basic and very important requirement for humans to survive and grow. The surface area of the earth provided by nature contains many different geographical locations divided into oceans, mountains, rivers, barren land, fertile land, ice caps and many more. The huge land masses and water bodies need to be observed and analyzed for optimum utilization of resources. Remote sensing is the best possible way to observe the earth's surface from a distance through different satellites and sensors. But most of the satellite images are not clear up to the extent to classify different terrain features accurately. Hence classification of image is needed to observe different terrain features in original images. In this study, the aim is to propose a branch of natural computation for SAR image classification into different terrain features with better information retrieval and accuracy measures as compared to traditional methods for satellite image classification. The object-based analysis has been used to extract spectral reflectance of five texture measures namely urban, rocky, vegetation, water and barren to generate training set. Minimum distance to mean classifier has been used with one of the Nature Inspired computation technique i.e. bacterial foraging optimization algorithm for the satellite image classification, to extract the more accurate information about land area of Alwar district, Rajasthan, India. In the proposed study a high-quality thematic map has been generated with the 7-band multi-spectral, medium-resolution satellite images. This approach provides the greater speed and accuracy in its computation with 97.43% overall accuracy (OA) and 0.96 Kappa co-efficient.

A machine learning approach to modelling solar irradiation of urban and terrain 3D models

• Solar irradiation predictive machine learning models are proposed. • Models use SVF and surface orientation as predicting features. • Parametric simulations using RADIANCE generate training sets. • Gradient boosting is used and tuned using cross-validation. • The models perform well on terrains but caution is advised for cities.