Limited Training Set Research Articles

A Learning Approach to Evaluate the Quality of 3D City Models

The automatic generation of three-dimensional (3D) building models from geospatial data is now a standard procedure. An abundance of literature covers the last two decades, and several solutions are now available. However, urban areas are very complex environments. Inevitably, practitioners still have to visually assess, at a city-scale, the correctness of these models and detect frequent reconstruction errors. Such a process relies on experts and is highly time-consuming, with approximately two hours/km 2 per expert. This work proposes an approach for automatically evaluating the quality of 3D building models. Potential errors are compiled in a novel hierarchical and versatile taxonomy. This allows, for the first time, to disentangle fidelity and modeling errors, whatever the level of details of the modeled buildings. The quality of models is predicted using the geometric properties of buildings and, when available, Very High Resolution images and Digital Surface Models. A baseline of handcrafted, yet generic, features is fed into a Random Forest classifier. Both multiclass and multilabel cases are considered: due to the interdependence between classes of errors, it is possible to retrieve all errors at the same time while simply predicting correct and erroneous buildings. The proposed framework was tested on three distinct urban areas in France with more than 3000 buildings. 80%–99% F-score values are attained for the most frequent errors. For scalability purposes, the impact of the urban area composition on the error prediction was also studied, in terms of transferability, generalization, and representativeness of the classifiers. It showed the necessity of multimodal remote sensing data and mixing training samples from various cities to ensure a stability of the detection ratios, even with very limited training set sizes.

Genetics of male reproductive performance in White Leghorns

The ability to produce viable progeny is a complex trait, involving both male and female components. In poultry, mating ratios are usually 1 male to 6 to 12 females. Consequently, the impact of male reproductive failure is much greater than that for a female. In this study, the genetic determination of male reproductive performance, by natural mating and artificial insemination (AI), was evaluated. Semen quality was studied in 1,575 pre-selected (using a selection index of multiple egg production and quality traits) White Leghorn males of a single pure line from multiple generations. A subset of individuals with satisfactory semen quality (based on sperm count and motility) were further tested for subsequent fertility and hatchability. Genetic parameters for fertility (FER), hatch of fertile (HOF), hatch of set (HOS), sperm motility (SM), sperm count (SC), and fertility using AI (FER-AI) were estimated using single- and multi-trait animal models, with generation as fixed effect. Selected birds were genotyped using the 600K Affymetrix SNP chip. Genomic data were analyzed with the BayesB method. FER, HOS, and HOF were highly correlated, both genetically (0.82 to 0.99) and phenotypically (0.28 to 0.99), but genetic correlations with semen quality traits were not strong (0.05 to 0.43) and phenotypic correlations varied between generations (-0.13 to 0.14). Birds used for fertility and hatchability tests were pre-selected based on SM and SC, which could contribute to the lack of strong correlations between these traits (due to truncation of the distribution). Based on pedigree information, low to moderate heritabilities were estimated for reproductive traits (0.08 to 0.21). Markers explained a low proportion of phenotypic variance (0.04 to 0.15), probably due to stringent selection of genotyped individuals and the limited training set size. No genes with large effects were identified. Genomic estimated breeding values were more accurate than pedigree-based estimates but only for HOF and FERT-AI. Despite low estimates of accuracy in validation, genetic trends were positive for all analyzed traits. In conclusion, continued long-term selection can result in genetic improvement of reproductive performance of roosters.

Abundance-Indicated Subspace for Hyperspectral Classification With Limited Training Samples

The imbalance between the (often limited) number of available training samples and the high data dimensionality, together with the presence of mixed pixels, often complicates the classification of remotely sensed hyperspectral data. In this paper, we tackle these problems by developing a new method that combines spectral unmixing and classification techniques in a subspace-based approach. The proposed method is developed under the assumption that the spectral signature of a land cover class is associated with a given set of pure spectral signatures (called endmembers in spectral unmixing terminology), which define a low-dimensional subspace with clear physical meaning. We aim to exploit this relationship to learn the class-dependent subspaces and integrate them with a multinomial logistic regression procedure. Experiments on synthetic datasets and real hyperspectral images show that our method is able to obtain competitive performances in comparison with other approaches, particularly when very limited training sets are available.

Robust URL Classification With Generative Adversarial Networks

Classifying URLs is essential for different applications, such as parental control, URL filtering and Ads/tracking protection. Such systems historically identify URLs by means of regular expressions, even if machine learning alternatives have been proposed to overcome the time-consuming maintenance of classification rules. Classical machine learning algorithms, however, require large samples of URLs to train the models, covering the diverse classes of URLs (i.e., a ground truth), which somehow limits the applicability of the approach. We here give a first step towards the use of Generative Adversarial Neural Networks (GANs) to classify URLs. GANs are attractive for this problem for two reasons. First, GANs can produce samples of URLs belonging to specific classes even if exposed to a limited training set, outputting both synthetic traces and a robust discriminator. Second, a GAN can be trained to discriminate a class of URLs without being exposed to all other URLs classes - i.e., GANs are robust even if not exposed to uninteresting URL classes during training. Experiments on real data show that not only the generated synthetic traces are somehow realistic, but also the URL classification is accurate with GANs.

The integration of pharmacophore-based 3D QSAR modeling and virtual screening in safety profiling: A case study to identify antagonistic activities against adenosine receptor, A2A, using 1,897 known drugs.

Safety pharmacology screening against a wide range of unintended vital targets using in vitro assays is crucial to understand off-target interactions with drug candidates. With the increasing demand for in vitro assays, ligand- and structure-based virtual screening approaches have been evaluated for potential utilization in safety profiling. Although ligand based approaches have been actively applied in retrospective analysis or prospectively within well-defined chemical space during the early discovery stage (i.e., HTS screening and lead optimization), virtual screening is rarely implemented in later stage of drug discovery (i.e., safety). Here we present a case study to evaluate ligand-based 3D QSAR models built based on in vitro antagonistic activity data against adenosine receptor 2A (A2A). The resulting models, obtained from 268 chemically diverse compounds, were used to test a set of 1,897 chemically distinct drugs, simulating the real-world challenge of safety screening when presented with novel chemistry and a limited training set. Due to the unique requirements of safety screening versus discovery screening, the limitations of 3D QSAR methods (i.e., chemotypes, dependence on large training set, and prone to false positives) are less critical than early discovery screen. We demonstrated that 3D QSAR modeling can be effectively applied in safety assessment prior to in vitro assays, even with chemotypes that are drastically different from training compounds. It is also worth noting that our model is able to adequately make the mechanistic distinction between agonists and antagonists, which is important to inform subsequent in vivo studies. Overall, we present an in-depth analysis of the appropriate utilization and interpretation of pharmacophore-based 3D QSAR models for safety screening.

Energy-Efficient Scheduling for Real-Time Systems Based on Deep Q-Learning Model

Energy saving is a critical and challenging issue for real-time systems in embedded devices because of their limited energy supply. To reduce the energy consumption, a hybrid dynamic voltage and frequency scaling (DVFS) scheduling based on Q-learning (QL-HDS) was proposed by combining energy-efficient DVFS techniques. However, QL-HDS discretizes the system state parameters with a certain step size, resulting in a poor distinction of the system states. More importantly, it is difficult for QL-HDS to learn a system for various task sets with a Q-table and limited training sets. In this paper, an energy-efficient scheduling scheme based on deep Q-learning model is proposed for periodic tasks in real-time systems (DQL-EES). Specially, a deep Q-learning model is designed by combining a stacked auto-encoder and a Q-learning model. In the deep Q-learning model, the stacked auto-encoder is used to replace the Q-function for learning the Q-value of each DVFS technology for any system state. Furthermore, a training strategy is devised to learn the parameters of the deep Q-learning model based on the experience replay scheme. Finally, the performance of the proposed scheme is evaluated by comparison with QL-HDS on different simulation task sets. Results demonstrated that the proposed algorithm can save average <inline-formula><tex-math notation="LaTeX">$4.2\%$</tex-math></inline-formula> energy than QL-HDS.

Energy-Efficient Classification for Resource-Constrained Biomedical Applications

Biomedical applications often require classifiers that are both accurate and cheap to implement. Today, deep neural networks achieve the state-of-the-art accuracy in most learning tasks that involve large data sets of unstructured data. However, the application of deep learning techniques may not be beneficial in problems with limited training sets and computational resources, or under domain-specific test time constraints. Among other algorithms, ensembles of decision trees, particularly the gradient boosted models have recently been very successful in machine learning competitions. Here, we propose an efficient hardware architecture to implement gradient boosted trees in applications under stringent power, area, and delay constraints, such as medical devices. Specifically, we introduce the concepts of asynchronous tree operation and sequential feature extraction to achieve an unprecedented energy and area efficiency. The proposed architecture is evaluated in automated seizure detection for epilepsy, using 3074 h of intracranial EEG data from 26 patients with 393 seizures. Average F1 scores of 99.23% and 87.86% are achieved for random and block-wise splitting of data into train/test sets, respectively, with an average detection latency of 1.1 s. The proposed classifier is fabricated in a 65-nm TSMC process, consuming 41.2 nJ/class in a total area of $540\times 1850\,\,\mathrm {\mu m}^{2}$ . This design improves the state-of-the-art by $27\times $ reduction in energy-area-latency product. Moreover, the proposed gradient-boosting architecture offers the flexibility to accommodate variable tree counts specific to each patient, to trade the predictive accuracy with energy. This patient-specific and energy-quality scalable classifier holds great promise for low-power sensor data classification in biomedical applications.

Deep Multi-Feature Learning for Water Body Extraction from Landsat Imagery

Water body extraction from remote sensing image data has been becoming a really hot topic. Recently, researchers put forward numerous methods for water body extraction, while most of them rely on elaborative feature selection and enough number of training samples. Convolution Neural Network (CNN), one of the implementation models of deep learning, has strong capability for two-dimension images’ classification. A new water body extraction model based on CNNs is established for deep multi-feature learning. Before experiment, image enhancement will be done by Dark Channel Prior. Then we concatenate three kinds of features: spectral information, spatial information that is extracted by Extended Multi-attribute Profile (EMAP) and various water indexes firstly. Next, feature matrixes are acted as the input of CNN-based model for training and classifying. The experimental results showed that the proposed model has better classification performance than Support Vector Machine (SVM) and artificial neural network (ANN). On very limited training set, our model could learn unique and representative features for better water body extraction.

Machine Learning in Risk Measurement: Gaussian Process Regression for Value-at-Risk and Expected Shortfall

While machine learning and its many variants are becoming established tools in quantitative finance, their application in a risk measurement context is less developed. This paper uses a scheme from probability theory and statistics – Gaussian Processes – and applies the corresponding non-parametric technique of Gaussian Process Regression to “train” a system suitable for revaluing instruments as required to determine a portfolio’s Value-at-Risk and Expected Shortfall. Time series of historical valuation parameters and prices of the portfolio’s constituents serve as the only inputs. On the example of a variety of portfolios consisting of vanilla and barrier options, it is demonstrated that, even with limited training sets, Gaussian Process Regression leads to risk figures identical to those from full revaluation and outperforms Taylor expansion. Applications for risk management and regulatory capital calculations are apparent. Research into an extension to related areas such as counterparty credit risk measurement is promising.

Biological image analysis using deep learning-based methods: Literature review

Automatic processing large amount of microscopic images is important for medical and biological studies. Deep learning has demonstrated better performance than traditional machine learning methods for processing massive quantities of images; therefore, it has attracted increasing attention from the research and industry fields. This paper summarizes the latest progress of deep learning methods in biological microscopic image processing, including image classification, object detection, and image segmentation. Compared to the traditional machine learning methods, deep neural networks achieved better accuracy without tedious feature selection procedure. Obstacles of the biological image analysis with deep learning methods include limited training set and imperfect image quality. Viable solutions to these obstacles are discussed at the end of the paper. With this survey, we hope to provide a reference for the researchers conducting biological microscopic image processing.

A transfer learning approach via procrustes analysis and mean shift for cancer drug sensitivity prediction

Transfer learning (TL) algorithms aim to improve the prediction performance in a target task (e.g. the prediction of cisplatin sensitivity in triple-negative breast cancer patients) via transferring knowledge from auxiliary data of a related task (e.g. the prediction of docetaxel sensitivity in breast cancer patients), where the distribution and even the feature space of the data pertaining to the tasks can be different. In real-world applications, we sometimes have a limited training set in a target task while we have auxiliary data from a related task. To obtain a better prediction performance in the target task, supervised learning requires a sufficiently large training set in the target task to perform well in predicting future test examples of the target task. In this paper, we propose a TL approach for cancer drug sensitivity prediction, where our approach combines three techniques. First, we shift the representation of a subset of examples from auxiliary data of a related task to a representation closer to a target training set of a target task. Second, we align the shifted representation of the selected examples of the auxiliary data to the target training set to obtain examples with representation aligned to the target training set. Third, we train machine learning algorithms using both the target training set and the aligned examples. We evaluate the performance of our approach against baseline approaches using the Area Under the receiver operating characteristic (ROC) Curve (AUC) on real clinical trial datasets pertaining to multiple myeloma, nonsmall cell lung cancer, triple-negative breast cancer, and breast cancer. Experimental results show that our approach is better than the baseline approaches in terms of performance and statistical significance.

Latent source mining in FMRI via restricted Boltzmann machine.

Blind source separation (BSS) is commonly used in functional magnetic resonance imaging (fMRI) data analysis. Recently, BSS models based on restricted Boltzmann machine (RBM), one of the building blocks of deep learning models, have been shown to improve brain network identification compared to conventional single matrix factorization models such as independent component analysis (ICA). These models, however, trained RBM on fMRI volumes, and are hence challenged by model complexity and limited training set. In this article, we propose to apply RBM to fMRI time courses instead of volumes for BSS. The proposed method not only interprets fMRI time courses explicitly to take advantages of deep learning models in latent feature learning but also substantially reduces model complexity and increases the scale of training set to improve training efficiency. Our experimental results based on Human Connectome Project (HCP) datasets demonstrated the superiority of the proposed method over ICA and the one that applied RBM to fMRI volumes in identifying task-related components, resulted in more accurate and specific representations of task-related activations. Moreover, our method separated out components representing intermixed effects between task events, which could reflect inherent interactions among functionally connected brain regions. Our study demonstrates the value of RBM in mining complex structures embedded in large-scale fMRI data and its potential as a building block for deeper models in fMRI data analysis.

Tile-Based Semisupervised Classification of Large-Scale VHR Remote Sensing Images

This paper deals with the problem of the classification of large-scale very high-resolution (VHR) remote sensing (RS) images in a semisupervised scenario, where we have a limited training set (less than ten training samples per class). Typical pixel-based classification methods are unfeasible for large-scale VHR images. Thus, as a practical and efficient solution, we propose to subdivide the large image into a grid of tiles and then classify the tiles instead of classifying pixels. Our proposed method uses the power of a pretrained convolutional neural network (CNN) to first extract descriptive features from each tile. Next, a neural network classifier (composed of 2 fully connected layers) is trained in a semisupervised fashion and used to classify all remaining tiles in the image. This basically presents a coarse classification of the image, which is sufficient for many RS application. The second contribution deals with the employment of the semisupervised learning to improve the classification accuracy. We present a novel semisupervised approach which exploits both the spectral and spatial relationships embedded in the remaining unlabelled tiles. In particular, we embed a spectral graph Laplacian in the hidden layer of the neural network. In addition, we apply regularization of the output labels using a spatial graph Laplacian and the random Walker algorithm. Experimental results obtained by testing the method on two large-scale images acquired by the IKONOS2 sensor reveal promising capabilities of this method in terms of classification accuracy even with less than ten training samples per class.

An ensemble approach for emotion cause detection with event extraction and multi-kernel SVMs

In this paper, we present a new challenging task for emotion analysis, namely emotion cause extraction. In this task, we focus on the detection of emotion cause a.k.a the reason or the stimulant of an emotion, rather than the regular emotion classification or emotion component extraction. Since there is no open dataset for this task available, we first designed and annotated an emotion cause dataset which follows the scheme of W3C Emotion Markup Language. We then present an emotion cause detection method by using event extraction framework, where a tree structure-based representation method is used to represent the events. Since the distribution of events is imbalanced in the training data, we propose an under-sampling-based bagging algorithm to solve this problem. Even with a limited training set, the proposed approach may still extract sufficient features for analysis by a bagging of multi-kernel based SVMs method. Evaluations show that our approach achieves an F-measure 7.04% higher than the state-of-the-art methods.

Abstract 4296: A real-time annexin V method for monitoring programmed cell death

Abstract Efficacious and durable anti-cancer responses are driven by both the selective death of malignant cells and the induction of immunostimulatory activities. Limited but promising clinical evidence suggests that provocation of an unknown balance of inflammatory- and non-inflammatory cell death may be key for orchestrating these positive outcomes. Therefore, identifying and characterizing new clinically useful small molecules and biologic inducers (or combinations thereof) which promote a spectrum of programmed cell death in in vitro screening environments remains critically important. Unfortunately, current screening methods are either insufficiently robust, cost- or resource-prohibitive, or provide no means for initial characterization of the kinetics of programmed cell death. To address this unmet need, we developed a real-time, live cell assay method that utilizes a fully homogeneous, bioluminescent annexin V reagent. The method does not require laborious washing and sample preparation steps associated with traditional annexin methods and is fully compatible with plate-based multimodal signal detection systems. The system contains two annexin proteins which have been engineered to contain separate and distinct complementing domains of a binary luciferase. Additionally, the system contains a novel time-released luciferase substrate and a cell impermeable, fluorogenic DNA dye for monitoring necrosis. Because the annexin-luciferase fusion pairs have only modest affinity for each other, luminescence remains low until phosphatidylserine exposure, a hallmark of the programmed cell death phenotype, brings annexin monomers into close proximity facilitating complementation of the luciferase sensor. The assay reagent can be applied at dosing for real-time measurement of the dose-dependency and magnitude of programmed cell death progression. This work describes our efforts to characterize and validate the performance of the bioluminescent annexin assay using relevant cell death induction models. First, the assay was shown to be functionally concordant with a flow cytometry annexin method in a dose-response model with bortezomib at exposure periods known to produce both early and late-apoptosis (with necrosis) phenotypes. Next, we assessed the performance of the assay using a limited training set of small molecule and biologic inducers of programmed cell death that utilize different mechanisms of action (i.e., apoptosis and necroptosis) and by using a number of diverse but representative cancer cell lines. We conclude that the bioluminescent annexin method provides a new kinetic approach to efficient and effective detection of programmed cell death mechanisms in real time in a convenient homogeneous format. Citation Format: Kevin Kupcho, Andrew Niles, John Shultz, Jamison Grailer, Wenhui Zhou, Robin Hurst, Jim Hartnett, Terry Riss, Dan Lazar, James Cali. A real-time annexin V method for monitoring programmed cell death [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 4296. doi:10.1158/1538-7445.AM2017-4296

Measuring air traffic complexity based on small samples

Air traffic complexity is an objective metric for evaluating the operational condition of the airspace. It has several applications, such as airspace design and traffic flow management. Therefore, identifying a reliable method to accurately measure traffic complexity is important. Considering that many factors correlate with traffic complexity in complicated nonlinear ways, researchers have proposed several complexity evaluation methods based on machine learning models which were trained with large samples. However, the high cost of sample collection usually results in limited training set. In this paper, an ensemble learning model is proposed for measuring air traffic complexity within a sector based on small samples. To exploit the classification information within each factor, multiple diverse factor subsets (FSSs) are generated under guidance from factor noise and independence analysis. Then, a base complexity evaluator is built corresponding to each FSS. The final complexity evaluation result is obtained by integrating all results from the base evaluators. Experimental studies using real-world air traffic operation data demonstrate the advantages of our model for small-sample-based traffic complexity evaluation over other state-of-the-art methods.

A deep learning approach for iris sensor model identification

The aim of this paper is to propose an algorithm based on convolutional neural networks (CNN) for iris sensor model identification. This task is important in forensics applications as well as to face the problem of sensor interoperability in large scale systems. When different sensor models are involved in a recognition system, in fact, the overall performance can strongly decrease. A possible solution consists in first identifying the sensor model and then mapping the features extracted from the image from one sensor to the other. To keep low both complexity and memory requirements we propose a simple network architecture and the use of transfer learning to speed-up the training phase and tackle the problem of limited training set availability. Experiments are carried out on several public iris databases. First, we show that the proposed solution outperforms the state-of-the art approaches used for the model identification task. Then, we test the performance of a biometric recognition system and show that improving the sensor model identification step can benefit the iris sensor interoperability.

Optimal sheet metal fixture locating layout by combining radial basis function neural network and bat algorithm

Considering that sheet metal part has the properties of thin wall, low rigidity, easy to deform, and difficult to locate, this article proposes a new approach to optimizing sheet metal fixture locating layout by combining radial basis function neural network and bat algorithm. First, taking fixture locating layout as design variables based on the “ N-2-1” locating principle, this article generates limited training and testing sample sets by Latin hypercube sampling and finite element analysis. Second, the radial basis function neural network prediction model with the description of the nonlinear mapping relationship between the fixture locating layout and the corresponding sheet metal deformation is constructed through learning from the training sample sets. Third, bat algorithm is applied to search the optimal layout of the “ N” fixture locators for the minimum sheet metal deformation. Finally, two case studies are presented to demonstrate the optimization procedure and the effectiveness of the proposed method.

Individual-free representation-based classification for facial expression recognition

Facial expression recognition (FER) plays a significant role in human–computer interaction. However, in FER applications, the samples are usually corrupted by individual differences, which affect the classification result to some extent. This paper proposes an individual-free representation-based classification, which utilizes the variation training set (VTS) and the virtual variation training set (VVTS) to remit the side-effect caused by individual differences. The VTS and VVTS are both generated from the original training set and show possible variation of the expression. The new approach performs low-rank decomposition-based singular value decomposition for both VTS and VVTS, and then integrates them to determine the label of the query sample. This promising performance is mainly attributed to the fact that VTS and VVTS used in the proposed method can exploit limited original training set to produce a large possible expression variation. Experimental results show that the proposed method can achieve better performance than most of the competitive FER methods, e.g., SVM, SRC, CRC, LRC and the method in Lee et al.

Kernel Supervised Ensemble Classifier for the Classification of Hyperspectral Data Using Few Labeled Samples

Kernel-based methods and ensemble learning are two important paradigms for the classification of hyperspectral remote sensing images. However, they were developed in parallel with different principles. In this paper, we aim to combine the advantages of kernel and ensemble methods by proposing a kernel supervised ensemble classification method. In particular, the proposed method, namely RoF-KOPLS, combines the merits of ensemble feature learning (i.e., Rotation Forest (RoF)) and kernel supervised learning (i.e., Kernel Orthonormalized Partial Least Square (KOPLS)). In particular, the feature space is randomly split into K disjoint subspace and KOPLS is applied to each subspace to produce the new features set for the training of decision tree classifier. The final classification result is assigned to the corresponding class by the majority voting rule. Experimental results on two hyperspectral airborne images demonstrated that RoF-KOPLS with radial basis function (RBF) kernel yields the best classification accuracies due to the ability of improving the accuracies of base classifiers and the diversity within the ensemble, especially for the very limited training set. Furthermore, our proposed method is insensitive to the number of subsets.