A Memory-Efficient Encoding Method for Processing Mixed-Type Data on Machine Learning

The most common machine-learning methods solve supervised and unsupervised problems based on datasets where the problem’s features belong to a numerical space. However, many problems often include data where numerical and categorical data coexist, which represents a challenge to manage them. To transform categorical data into a numeric form, preprocessing tasks are compulsory. Methods such as one-hot and feature-hashing have been the most widely used encoding approaches at the expense of a significant increase in the dimensionality of the dataset. This effect introduces unexpected challenges to deal with the overabundance of variables and/or noisy data. In this regard, in this paper we propose a novel encoding approach that maps mixed-type data into an information space using Shannon’s Theory to model the amount of information contained in the original data. We evaluated our proposal with ten mixed-type datasets from the UCI repository and two datasets representing real-world problems obtaining promising results. For demonstrating the performance of our proposal, this was applied for preparing these datasets for classification, regression, and clustering tasks. We demonstrate that our encoding proposal is remarkably superior to one-hot and feature-hashing encoding in terms of memory efficiency. Our proposal can preserve the information conveyed by the original data.     

An Approach to Growth Delimitation of Straight Line Segment Classifiers Based on a Minimum Bounding Box

Several supervised machine learning algorithms focused on binary classification for solving daily problems can be found in the literature. The straight-line segment classifier stands out for its low complexity and competitiveness, compared to well-knownconventional classifiers. This binary classifier is based on distances between points and two labeled sets of straight-line segments. Its training phase consists of finding the placement of labeled straight-line segment extremities (and consequently, their lengths) which gives the minimum mean square error. However, during the training phase, the straight-line segment lengths can grow significantly, giving a negative impact on the classification rate. Therefore, this paper proposes an approach for adjusting the placements of labeled straight-line segment extremities to build reliable classifiers in a constrained search space (tuned by a scale factor parameter) in order to restrict their lengths. Ten artificial and eight datasets from the UCI Machine Learning Repository were used to prove that our approach shows promising results, compared to other classifiers. We conclude that this classifier can be used in industry for decision-making problems, due to the straightforward interpretation and classification rates.     

A Manifold Learning Perspective on Representation Learning: Learning Decoder and Representations without an Encoder

Autoencoders are commonly used in representation learning. They consist of an encoder and a decoder, which provide a straightforward method to map n-dimensional data in input space to a lower m-dimensional representation space and back. The decoder itself defines an m-dimensional manifold in input space. Inspired by manifold learning, we showed that the decoder can be trained on its own by learning the representations of the training samples along with the decoder weights using gradient descent. A sum-of-squares loss then corresponds to optimizing the manifold to have the smallest Euclidean distance to the training samples, and similarly for other loss functions. We derived expressions for the number of samples needed to specify the encoder and decoder and showed that the decoder generally requires much fewer training samples to be well-specified compared to the encoder. We discuss the training of autoencoders in this perspective and relate it to previous work in the field that uses noisy training examples and other types of regularization. On the natural image data sets MNIST and CIFAR10, we demonstrated that the decoder is much better suited to learn a low-dimensional representation, especially when trained on small data sets. Using simulated gene regulatory data, we further showed that the decoder alone leads to better generalization and meaningful representations. Our approach of training the decoder alone facilitates representation learning even on small data sets and can lead to improved training of autoencoders. We hope that the simple analyses presented will also contribute to an improved conceptual understanding of representation learning.     

Co-Training for Visual Object Recognition Based on Self-Supervised Models Using a Cross-Entropy Regularization

Automatic recognition of visual objects using a deep learning approach has been successfully applied to multiple areas. However, deep learning techniques require a large amount of labeled data, which is usually expensive to obtain. An alternative is to use semi-supervised models, such as co-training, where multiple complementary views are combined using a small amount of labeled data. A simple way to associate views to visual objects is through the application of a degree of rotation or a type of filter. In this work, we propose a co-training model for visual object recognition using deep neural networks by adding layers of self-supervised neural networks as intermediate inputs to the views, where the views are diversified through the cross-entropy regularization of their outputs. Since the model merges the concepts of co-training and self-supervised learning by considering the differentiation of outputs, we called it Differential Self-Supervised Co-Training (DSSCo-Training). This paper presents some experiments using the DSSCo-Training model to well-known image datasets such as MNIST, CIFAR-100, and SVHN. The results indicate that the proposed model is competitive with the state-of-art models and shows an average relative improvement of 5% in accuracy for several datasets, despite its greater simplicity with respect to more recent approaches.     

Unsupervised Anomaly Detection with Distillated Teacher-Student Network Ensemble

We address the problem of unsupervised anomaly detection for multivariate data. Traditional machine learning based anomaly detection algorithms rely on specific assumptions of normal patterns and fail to model complex feature interactions and relations. Recently, existing deep learning based methods are promising for extracting representations from complex features. These methods train an auxiliary task, e.g., reconstruction and prediction, on normal samples. They further assume that anomalies fail to perform well on the auxiliary task since they are never trained during the model optimization. However, the assumption does not always hold in practice. Deep models may also perform the auxiliary task well on anomalous samples, leading to the failure detection of anomalies. To effectively detect anomalies for multivariate data, this paper introduces a teacher-student distillation based framework Distillated Teacher-Student Network Ensemble (DTSNE). The paradigm of the teacher-student distillation is able to deal with high-dimensional complex features. In addition, an ensemble of student networks provides a better capability to avoid generalizing the auxiliary task performance on anomalous samples. To validate the effectiveness of our model, we conduct extensive experiments on real-world datasets. Experimental results show superior performance of DTSNE over competing methods. Analysis and discussion towards the behavior of our model are also provided in the experiment section.     

New Interfaces and Approaches to Machine Learning When Classifying Gestures within Music

Interactive music uses wearable sensors (i.e., gestural interfaces—GIs) and biometric datasets to reinvent traditional human–computer interaction and enhance music composition. In recent years, machine learning (ML) has been important for the artform. This is because ML helps process complex biometric datasets from GIs when predicting musical actions (termed performance gestures). ML allows musicians to create novel interactions with digital media. Wekinator is a popular ML software amongst artists, allowing users to train models through demonstration. It is built on the Waikato Environment for Knowledge Analysis (WEKA) framework, which is used to build supervised predictive models. Previous research has used biometric data from GIs to train specific ML models. However, previous research does not inform optimum ML model choice, within music, or compare model performance. Wekinator offers several ML models. Thus, we used Wekinator and the Myo armband GI and study three performance gestures for piano practice to solve this problem. Using these, we trained all models in Wekinator and investigated their accuracy, how gesture representation affects model accuracy and if optimisation can arise. Results show that neural networks are the strongest continuous classifiers, mapping behaviour differs amongst continuous models, optimisation can occur and gesture representation disparately affects model mapping behaviour; impacting music practice.     

Whole volume brain extraction for multi-centre,multi-disease FLAIR MRI datasets

Automatic segmentation of the brain from magnetic resonance images (MRI) is a fundamental step in many neuroimaging processing frameworks. There are mature technologies for this task for T1- and T2-weighted MRI; however, a widely-accepted brain extraction method for Fluid-Attenuated Inversion Recovery (FLAIR) MRI has yet to be established. FLAIR MRI are becoming increasingly important for the analysis of neurodegenerative diseases and tools developed for this sequence would have clinical value. To maximize translation opportunities and for large scale research studies, algorithms for brain extraction in FLAIR MRI should generalize to multi-centre (MC) data. To this end, this work proposes a fully automated, whole volume brain extraction methodology for MC FLAIR MRI datasets. The framework is built using a novel standardization framework which reduces acquisition artifacts, standardizes the intensities of tissues and normalizes the spatial coordinates of brain tissue across MC datasets. Using the standardized datasets, an intuitive set of features based on intensity, spatial location and gradients are extracted and classified using a random forest (RF) classifier to segment the brain tissue class. A series of experiments were conducted to optimize classifier parameters, and to determine segmentation accuracy for standardized and unstandardized (original) data, as a function of scanner vendor, feature type and disease type. The models are trained, tested and validated on 156 image volumes (∼8000 image slices) from two multi-centre, multi-disease datasets, acquired with varying imaging parameters from 30 centres and three scanner vendors. The image datasets, denoted as CAIN and ADNI for vascular and dementia disease, respectively, represent a diverse collection of MC data to test the generalization capabilities of the proposed design. Results demonstrate the importance of standardization for segmentation of MC data, as models trained on standardized data yielded a drastic improvement in brain extraction accuracy compared to the original, unstandardized data (CAIN: DSC = 91% and ADNI: DSC = 86% vs. CAIN: 78% and ADNI: 65%). It was also found that models created from one scanner vendor based on unstandardized data yielded poor segmentation results in data acquired from other scanner vendors, which was improved through standardization. These results demonstrate that to create consistency in segmentations from multi-institutional datasets it is paramount that MC variability be mitigated to improve stability and to ensure generalization of machine learning algorithms for MRI.     

The use of pseudo-faults for damage location in SHM: An experimental investigation on a Piper Tomahawk aircraft wing

The application of pattern recognition-based approaches in damage localisation and quantification will eventually require the use of some kind of supervised learning algorithm. The use, and most importantly, the success of such algorithms will depend critically on the availability of data from all possible damage states for training. It is perhaps well known that the availability of damage data through destructive means cannot generally be afforded in the case of high value engineering structures outside laboratory conditions. This paper presents the attempt to use added masses in order to identify features suitable for training supervised learning algorithms and then to test the trained classifiers with damage data, with the ultimate purpose of damage localisation. In order to test the approach of adding masses, two separate cases of a dual-class classification problem, representing two distinct locations, and a three-class problem representing three distinct locations, are examined with the help of a full-scale aircraft wing. It was found that an excellent rate of correct classification could be achieved in both the dual-class and three-class cases. However, it was also found that the rate of correct classification was sensitive to the choices made in training the supervised learning algorithm. The results for the dual-class problem demonstrated a comparatively high level of robustness to these choices with a substantially lower robustness found in the three-class case.     

Learning Ordinal Embedding from Sets

Ordinal embedding is the task of computing a meaningful multidimensional representation of objects, for which only qualitative constraints on their distance functions are known. In particular, we consider comparisons of the form “Which object from the pair (j,k) is more similar to object i?”. In this paper, we generalize this framework to the case where the ordinal constraints are not given at the level of individual points, but at the level of sets, and propose a distributional triplet embedding approach in a scalable learning framework. We show that the query complexity of our approach is on par with the single-item approach. Without having access to features of the items to be embedded, we show the applicability of our model on toy datasets for the task of reconstruction and demonstrate the validity of the obtained embeddings in experiments on synthetic and real-world datasets.     

A multi-scale residual network for accelerated radial MR parameter mapping

A deep learning MR parameter mapping framework which combines accelerated radial data acquisition with a multi-scale residual network (MS-ResNet) for image reconstruction is proposed. The proposed supervised learning strategy uses input image patches from multi-contrast images with radial undersampling artifacts and target image patches from artifact-free multi-contrast images. Subspace filtering is used during pre-processing to denoise input patches. For each anatomy and relaxation parameter, an individual network is trained. in vivo T₁ mapping results are obtained on brain and abdomen datasets and in vivo T₂ mapping results are obtained on brain and knee datasets. Quantitative results for the T₂ mapping of the knee show that MS-ResNet trained using either fully sampled or undersampled data outperforms conventional model-based compressed sensing methods. This is significant because obtaining fully sampled training data is not possible in many applications. in vivo brain and abdomen results for T₁ mapping and in vivo brain results for T₂ mapping demonstrate that MS-ResNet yields contrast-weighted images and parameter maps that are comparable to those achieved by model-based iterative methods while offering two orders of magnitude reduction in reconstruction times. The proposed approach enables recovery of high-quality contrast-weighted images and parameter maps from highly accelerated radial data acquisitions. The rapid image reconstructions enabled by the proposed approach makes it a good candidate for routine clinical use.     

Task’s Choice: Pruning-Based Feature Sharing (PBFS) for Multi-Task Learning

In most of the existing multi-task learning (MTL) models, multiple tasks’ public information is learned by sharing parameters across hidden layers, such as hard sharing, soft sharing, and hierarchical sharing. One promising approach is to introduce model pruning into information learning, such as sparse sharing, which is regarded as being outstanding in knowledge transferring. However, the above method performs inefficiently in conflict tasks, with inadequate learning of tasks’ private information, or through suffering from negative transferring. In this paper, we propose a multi-task learning model (Pruning-Based Feature Sharing, PBFS) that merges a soft parameter sharing structure with model pruning and adds a prunable shared network among different task-specific subnets. In this way, each task can select parameters in a shared subnet, according to its requirements. Experiments are conducted on three benchmark public datasets and one synthetic dataset; the impact of the different subnets’ sparsity and tasks’ correlations to the model performance is analyzed. Results show that the proposed model’s information sharing strategy is helpful to transfer learning and superior to the several comparison models.     

Using Copula distributions to support more accurate imaging-based diagnostic classifiers for neuropsychiatric disorders

Many investigators have tried to apply machine learning techniques to magnetic resonance images (MRIs) of the brain in order to diagnose neuropsychiatric disorders. Usually the number of brain imaging measures (such as measures of cortical thickness and measures of local surface morphology) derived from the MRIs (i.e., their dimensionality) has been large (e.g. > 10) relative to the number of participants who provide the MRI data (< 100). Sparse data in a high dimensional space increase the variability of the classification rules that machine learning algorithms generate, thereby limiting the validity, reproducibility, and generalizability of those classifiers. The accuracy and stability of the classifiers can improve significantly if the multivariate distributions of the imaging measures can be estimated accurately. To accurately estimate the multivariate distributions using sparse data, we propose to estimate first the univariate distributions of imaging data and then combine them using a Copula to generate more accurate estimates of their multivariate distributions. We then sample the estimated Copula distributions to generate dense sets of imaging measures and use those measures to train classifiers. We hypothesize that the dense sets of brain imaging measures will generate classifiers that are stable to variations in brain imaging measures, thereby improving the reproducibility, validity, and generalizability of diagnostic classification algorithms in imaging datasets from clinical populations. In our experiments, we used both computer-generated and real-world brain imaging datasets to assess the accuracy of multivariate Copula distributions in estimating the corresponding multivariate distributions of real-world imaging data. Our experiments showed that diagnostic classifiers generated using imaging measures sampled from the Copula were significantly more accurate and more reproducible than were the classifiers generated using either the real-world imaging measures or their multivariate Gaussian distributions. Thus, our findings demonstrate that estimated multivariate Copula distributions can generate dense sets of brain imaging measures that can in turn be used to train classifiers, and those classifiers are significantly more accurate and more reproducible than are those generated using real-world imaging measures alone.     

Experimental Study and Comparison of Imbalance Ensemble Classifiers with Dynamic Selection Strategy

Imbalance ensemble classification is one of the most essential and practical strategies for improving decision performance in data analysis. There is a growing body of literature about ensemble techniques for imbalance learning in recent years, the various extensions of imbalanced classification methods were established from different points of view. The present study is initiated in an attempt to review the state-of-the-art ensemble classification algorithms for dealing with imbalanced datasets, offering a comprehensive analysis for incorporating the dynamic selection of base classifiers in classification. By conducting 14 existing ensemble algorithms incorporating a dynamic selection on 56 datasets, the experimental results reveal that the classical algorithm with a dynamic selection strategy deliver a practical way to improve the classification performance for both a binary class and multi-class imbalanced datasets. In addition, by combining patch learning with a dynamic selection ensemble classification, a patch-ensemble classification method is designed, which utilizes the misclassified samples to train patch classifiers for increasing the diversity of base classifiers. The experiments’ results indicate that the designed method has a certain potential for the performance of multi-class imbalanced classification.     

Investigation on Identifying Implicit Learning Event from EEG Signal Using Multiscale Entropy and Artificial Bee Colony

The way people learn will play an essential role in the sustainable development of the educational system for the future. Utilizing technology in the age of information and incorporating it into how people learn can produce better learners. Implicit learning is a type of learning of the underlying rules without consciously seeking or understanding the rules; it is commonly seen in small children while learning how to speak their native language without learning grammar. This research aims to introduce a processing system that can systematically identify the relationship between implicit learning events and their Encephalogram (EEG) signal characteristics. This study converted the EEG signal from participants while performing cognitive task experiments into Multiscale Entropy (MSE) data. Using MSE data from different frequency bands and channels as features, the system explored a wide range of classifiers and observed their performance to see how they classified the features related to participants’ performance. The Artificial Bee Colony (ABC) method was used for feature selection to improve the process to make the system more efficient. The results showed that the system could correctly identify the differences between participants’ performance using MSE data and the ABC method with 95% confidence.     

Subspace ensembles for classification

Ensemble learning constitutes one of the principal current directions in machine learning and data mining. In this paper, we explore subspace ensembles for classification by manipulating different feature subspaces. Commencing with the nature of ensemble efficacy, we probe into the microcosmic meaning of ensemble diversity, and propose to use region partitioning and region weighting to implement effective subspace ensembles. Individual classifiers possessing eminent performance on a partitioned region reflected by high neighborhood accuracies are deemed to contribute largely to this region, and are assigned large weights in determining the labels of instances in this area. A robust algorithm “Sena” that incarnates the mechanism is presented, which is insensitive to the number of nearest neighbors chosen to calculate neighborhood accuracies. The algorithm exhibits improved performance over the well-known ensembles of bagging, AdaBoost and random subspace. The difference of its effectivity with varying base classifiers is also investigated.     

一种基于GAN和自适应迁移学习的样本生成方法

研究了基于生成式对抗网络(GAN)和跨域自适应迁移学习的样本生成和自动标注方法。该方法利用自适应迁移学习网络,基于已有的少量可见光图像样本集,挖掘目标在红外和可见光图像中特征内在相关性,构建自适应的转换迁移学习网络模型,生成标注好的目标图像。提出的方法解决了红外图像样本数量少且标注费时的问题,为后续多频段协同目标检测和识别获得了足够的样本数据。实验结果表明:自动标注算法对实际采集的装甲目标图像和生成的装甲目标图像各1 000张进行自动标注测试,对实际装甲目标图像的标注准确率达到95%以上,对生成的装甲目标标注准确率达到83%以上;利用真实图像和生成图像的混合数据集训练的分类器的性能和使用纯真实图像时基本一致。     

Multi-Scale Aggregation Graph Neural Networks Based on Feature Similarity for Semi-Supervised Learning

The problem of extracting meaningful data through graph analysis spans a range of different fields, such as social networks, knowledge graphs, citation networks, the World Wide Web, and so on. As increasingly structured data become available, the importance of being able to effectively mine and learn from such data continues to grow. In this paper, we propose the multi-scale aggregation graph neural network based on feature similarity (MAGN), a novel graph neural network defined in the vertex domain. Our model provides a simple and general semi-supervised learning method for graph-structured data, in which only a very small part of the data is labeled as the training set. We first construct a similarity matrix by calculating the similarity of original features between all adjacent node pairs, and then generate a set of feature extractors utilizing the similarity matrix to perform multi-scale feature propagation on graphs. The output of multi-scale feature propagation is finally aggregated by using the mean-pooling operation. Our method aims to improve the model representation ability via multi-scale neighborhood aggregation based on feature similarity. Extensive experimental evaluation on various open benchmarks shows the competitive performance of our method compared to a variety of popular architectures.     

Transfer-Learning-Based Approach for the Diagnosis of Lung Diseases from Chest X-ray Images

Using chest X-ray images is one of the least expensive and easiest ways to diagnose patients who suffer from lung diseases such as pneumonia and bronchitis. Inspired by existing work, a deep learning model is proposed to classify chest X-ray images into 14 lung-related pathological conditions. However, small datasets are not sufficient to train the deep learning model. Two methods were used to tackle this: (1) transfer learning based on two pretrained neural networks, DenseNet and ResNet, was employed; (2) data were preprocessed, including checking data leakage, handling class imbalance, and performing data augmentation, before feeding the neural network. The proposed model was evaluated according to the classification accuracy and receiver operating characteristic (ROC) curves, as well as visualized by class activation maps. DenseNet121 and ResNet50 were used in the simulations, and the results showed that the model trained by DenseNet121 had better accuracy than that trained by ResNet50.     

Deep learning for image reconstruction in thermoacoustic tomography

Microwave-induced thermoacoustic tomography(TAT)is a rapidly-developing noninvasive imaging technique that integrates the advantages of microwave imaging and ultrasound imaging.While an image reconstruction algorithm is critical for the TAT,current reconstruction methods often creates significant artifacts and are computationally costly.In this work,we propose a deep learning-based end-to-end image reconstruction method to achieve the direct reconstruction from the sinogram data to the initial pressure density image.We design a new network architecture TAT-Net to transfer the sinogram domain to the image domain with high accuracy.For the scenarios where realistic training data are scarce or unavailable,we use the finite element method(FEM)to generate synthetic data where the domain gap between the synthetic and realistic data is resolved through the signal processing method.The TAT-Net trained with synthetic data is evaluated through both simulations and phantom experiments and achieves competitive performance in artifact removal and robustness.Compared with other state-of-the-art reconstruction methods,the TAT-Net method can reduce the root mean square error to 0.0143,and increase the structure similarity and peak signal-to-noise ratio to 0.988 and 38.64,respectively.The results obtained indicate that the TAT-Net has great potential applications in improving image reconstruction quality and fast quantitative reconstruction.     

Federated Quantum Machine Learning

Distributed training across several quantum computers could significantly improve the training time and if we could share the learned model, not the data, it could potentially improve the data privacy as the training would happen where the data is located. One of the potential schemes to achieve this property is the federated learning (FL), which consists of several clients or local nodes learning on their own data and a central node to aggregate the models collected from those local nodes. However, to the best of our knowledge, no work has been done in quantum machine learning (QML) in federation setting yet. In this work, we present the federated training on hybrid quantum-classical machine learning models although our framework could be generalized to pure quantum machine learning model. Specifically, we consider the quantum neural network (QNN) coupled with classical pre-trained convolutional model. Our distributed federated learning scheme demonstrated almost the same level of trained model accuracies and yet significantly faster distributed training. It demonstrates a promising future research direction for scaling and privacy aspects.