Feature Transfer Learning Research Articles

Correlation feature and instance weights transfer learning for cross project software defect prediction

Due to the differentiation between training and testing data in the feature space, cross-project defect prediction (CPDP) remains unaddressed within the field of traditional machine learning. Recently, transfer learning has become a research hot-spot for building classifiers in the target domain using the data from the related source domains. To implement better CPDP models, recent studies focus on either feature transferring or instance transferring to weaken the impact of irrelevant cross-project data. Instead, this work proposes a dual weighting mechanism to aid the learning process, considering both feature transferring and instance transferring. In our method, a local data gravitation between source and target domains determines instance weight, while features that are highly correlated with the learning task, uncorrelated with other features and minimizing the difference between the domains are rewarded with a higher feature weight. Experiments on 25 real-world datasets indicate that the proposed approach outperforms the existing CPDP methods in most cases. By assigning weights based on the different contribution of features and instances to the predictor, the proposed approach is able to build a better CPDP model and demonstrates substantial improvements over the state-of-the-art CPDP models.

Improving prognostic performance in resectable pancreatic ductal adenocarcinoma using radiomics and deep learning features fusion in CT images

As an analytic pipeline for quantitative imaging feature extraction and analysis, radiomics has grown rapidly in the past decade. On the other hand, recent advances in deep learning and transfer learning have shown significant potential in the quantitative medical imaging field, raising the research question of whether deep transfer learning features have predictive information in addition to radiomics features. In this study, using CT images from Pancreatic Ductal Adenocarcinoma (PDAC) patients recruited in two independent hospitals, we discovered most transfer learning features have weak linear relationships with radiomics features, suggesting a potential complementary relationship between these two feature sets. We also tested the prognostic performance for overall survival using four feature fusion and reduction methods for combining radiomics and transfer learning features and compared the results with our proposed risk score-based feature fusion method. It was shown that the risk score-based feature fusion method significantly improves the prognosis performance for predicting overall survival in PDAC patients compared to other traditional feature reduction methods used in previous radiomics studies (40% increase in area under ROC curve (AUC) yielding AUC of 0.84).

Multi-Type Diesel Engines Operating Condition Recognition Method Based on Stacked Auto-Encoder and Feature Transfer Learning

It is of great significance to diagnose the fault of diesel engine, which is widely used in many important fields as key power equipment. The accuracy of fault diagnosis can be effectively improved by obtaining the complex and changeable operating conditions, which can result in the change of monitoring signals. This study proposes a variable operating conditions recognition method based on stacked auto-encoder (SAE) and feature transfer learning. In this method, the vibration in the firing angle domain collected from multi-sensor signals is reconstructed. Then a feature set sensitive to working conditions is extracted from the recombinant signals by a well-constructed stack auto-encoder. According to the dataset test, the softmax classifier can effectively get a high recognition accuracy. Considering that the fault may affect the condition identification, the misfire fault that has a great influence on firing angle domain signals is used to test the robustness of the proposed method. Besides, to enable a well-trained test rig with a large amount of data to be effectively applied to another unit that lacks data, the BDA transfer learning method is used to map the operating conditions of two different engines to the same feature space. The results of experiments conducted on two large power marine multi-cylinder diesel engines show that BDA is capable of transferring the sensitive features of operating conditions.

Exploiting Transfer Learning and Hand-Crafted Features in a Unified Neural Model for Identifying Actionable Informative Tweets

Exploiting Transfer Learning and Hand-Crafted Features in a Unified Neural Model for Identifying Actionable Informative Tweets

On Lemon Defect Recognition with Visual Feature Extraction and Transfers Learning

Applying machine learning to lemon defect recognition can improve the efficiency of lemon quality detection. This paper proposes a deep learning-based classification method with visual feature extraction and transfer learning to recognize defect lemons (i.e., green and mold defects). First, the data enhancement and brightness compensation techniques are used for data prepossessing. The visual feature extraction is used to quantify the defects and determine the feature variables as the bandit basis for classification. Then we construct a convolutional neural network with an embedded Visual Geometry Group 16 based (VGG16-based) network using transfer learning. The proposed model is compared with many benchmark models such as K-nearest Neighbor (KNN) and Support Vector Machine (SVM). Results show that the proposed model achieves the highest accuracy (95.44%) in the testing data set. The research provides a new solution for lemon defect recognition.

Skin Lesion Classification by Multi-View Filtered Transfer Learning

Skin cancer is one of the most deadly cancer types with considerable number of patients. Image analysis has largely improved the automated diagnosis accuracy for malignant melanoma and other pigmented skin lesions, compared to unaided visual examination. Recent popular solution for automated skin lesion classification is using deep neural networks, trained from large amounts of professional annotated data, but that largely limits the model’s scalability. This paper exploits transfer learning for skin lesion classification task with the help of labeled data from another domain (source), and proposes a multi-view filtered transfer learning network to strongly represent discriminative information from different image views with reasonable weighing strategy. This method also evaluates the importance for each source images, which can learn useful knowledge with neglecting negative samples from source domain. The extensive skin lesion classification experiments demonstrate our method can effectively solve Melanoma and Seborrheic Keratosis classification tasks with outstanding extensibility, and the discussion of the major components also testifies the improvements of our proposed multi-view filtered transfer learning approach.

Bridging multimedia heterogeneity gap via Graph Representation Learning for cross-modal retrieval

Information retrieval among different modalities becomes a significant issue with many promising applications. However, inconsistent feature representation of various multimedia data causes the “heterogeneity gap” among various modalities, which is a challenge in cross-modal retrieval. For bridging the “heterogeneity gap,” the popular methods attempt to project the original data into a common representation space, which needs great fitting ability of the model. To address the above issue, we propose a novel Graph Representation Learning (GRL) method for bridging the heterogeneity gap, which does not project the original feature into an aligned representation space but adopts a cross-modal graph to link different modalities. The GRL approach consists of two subnetworks, Feature Transfer Learning Network (FTLN) and Graph Representation Learning Network (GRLN). Firstly, FTLN model finds a latent space for each modality, where the cosine similarity is suitable to describe their similarity. Then, we build a cross-modal graph to reconstruct the original data and their relationships. Finally, we abandon the features in the latent space and turn into embedding the graph vertexes into a common representation space directly. During the process, the proposed Graph Representation Learning method bypasses the most challenging issue by utilizing a cross-modal graph as a bridge to link the “heterogeneity gap” among different modalities. This attempt utilizes a cross-modal graph as an intermediary agent to bridge the “heterogeneity gap” in cross-modal retrieval, which is simple but effective. Extensive experiment results on six widely-used datasets indicate that the proposed GRL outperforms other state-of-the-art cross-modal retrieval methods.

Bucket of Deep Transfer Learning Features and Classification Models for Melanoma Detection.

Malignant melanoma is the deadliest form of skin cancer and, in recent years, is rapidly growing in terms of the incidence worldwide rate. The most effective approach to targeted treatment is early diagnosis. Deep learning algorithms, specifically convolutional neural networks, represent a methodology for the image analysis and representation. They optimize the features design task, essential for an automatic approach on different types of images, including medical. In this paper, we adopted pretrained deep convolutional neural networks architectures for the image representation with purpose to predict skin lesion melanoma. Firstly, we applied a transfer learning approach to extract image features. Secondly, we adopted the transferred learning features inside an ensemble classification context. Specifically, the framework trains individual classifiers on balanced subspaces and combines the provided predictions through statistical measures. Experimental phase on datasets of skin lesion images is performed and results obtained show the effectiveness of the proposed approach with respect to state-of-the-art competitors.

Automated detection of pneumoconiosis with multilevel deep features learned from chest X-Ray radiographs

Early detection of pneumoconiosis in X-Rays has been a challenging task that leads to high inter- and intra-reader variability. Motivated by the success of deep learning in general and medical image classification, this paper proposes an approach to automatically detect pneumoconiosis using a deep feature based binary classifier. The features are extracted from X-rays using deep transfer learning, comprising both low and high-level feature sets. For this, a CNN model pre-trained with a transfer learning from a CheXNet model was initially used to extract deep features from the X-Ray images, then the deep features were mapped to higher-dimensional feature spaces for classification using Support Vector Machine (SVM) and CNN based feature aggregation methods. In order to cross validate the proposed method, the training and testing images were randomly split into three folds before each experiment. Nine evaluation metrics were employed to compare the performance of the proposed method and state-of-the-art methods from the literature that used the same datasets. The experimental results show that the proposed framework outperformed others, achieving an accuracy of 92.68% in the automated detection of pneumoconiosis.

COVID-19 classification by CCSHNet with deep fusion using transfer learning and discriminant correlation analysis

Aim: COVID-19 is a disease caused by a new strain of coronavirus. Up to 18th October 2020, worldwide there have been 39.6 million confirmed cases resulting in more than 1.1 million deaths. To improve diagnosis, we aimed to design and develop a novel advanced AI system for COVID-19 classification based on chest CT (CCT) images. Methods: Our dataset from local hospitals consisted of 284 COVID-19 images, 281 community-acquired pneumonia images, 293 secondary pulmonary tuberculosis images; and 306 healthy control images. We first used pretrained models (PTMs) to learn features, and proposed a novel (L, 2) transfer feature learning algorithm to extract features, with a hyperparameter of number of layers to be removed (NLR, symbolized as L). Second, we proposed a selection algorithm of pretrained network for fusion to determine the best two models characterized by PTM and NLR. Third, deep CCT fusion by discriminant correlation analysis was proposed to help fuse the two features from the two models. Micro-averaged (MA) F1 score was used as the measuring indicator. The final determined model was named CCSHNet. Results: On the test set, CCSHNet achieved sensitivities of four classes of 95.61%, 96.25%, 98.30%, and 97.86%, respectively. The precision values of four classes were 97.32%, 96.42%, 96.99%, and 97.38%, respectively. The F1 scores of four classes were 96.46%, 96.33%, 97.64%, and 97.62%, respectively. The MA F1 score was 97.04%. In addition, CCSHNet outperformed 12 state-of-the-art COVID-19 detection methods. Conclusions: CCSHNet is effective in detecting COVID-19 and other lung infectious diseases using first-line clinical imaging and can therefore assist radiologists in making accurate diagnoses based on CCTs.

On null arguments and phi-features in second language acquisition

AbstractSaito, Mamoru. 2007. Notes on East Asian argument ellipsis.Language Research43. 203–227 argues thatargument ellipsis(AE) is available only in languages that lack phi-feature agreement. Accordingly, Japanese, but not English, permits AE. Under Saito’s theoretical framework, this paper compares experimental data from L1 Japanese learners of L2 English (J-EFL) and L1 English learners of L2 Japanese (E-JFL). Given that sloppy and quantificational reading arises from an ellipsis operation (Hankamer, Jorge & Sag, Ivan. 1976. Deep and surface anaphora.Linguistic Inquiry7. 391–426, Takahashi, Daiko. 2008. Noun phrase ellipsis. In Miyagawa, Shigeru & Saito, Mamoru (eds.),The Oxford handbook of Japanese linguistics, 394–422. Oxford: Oxford University Press, among others), we hypothesize that J-EFL learners, but not E-JFL learners, allow the reading in point with null arguments: AE is available only in the grammar of J-EFL learners, forced by the lack of phi-features in their L2 English grammar, due to L1 transfer. The results from our main study adopting a truth value judgement task supported the hypothesis. Based on our finding, we suggest that correct L2 phi-feature specification can ultimately be obtained when no phi-features are present in L1 (Ishino, Nao. 2012.Feature transfer and feature learning in universal grammar: A comparative study of the syntactic mechanism for second language acquisition. Doctoral dissertation: Kwansei Gakuin University, Miyamoto, Yoichi. 2012.Dainigengo-ni okeru hikenzaiteki-na yōso-ni kansuru Ichikōsatsu[A study on null elements in second language acquisition]. Paper presented at the 84th ELSJ annual general meeting: Senshu University, 26 May).

Comparative Analysis of Supervised and Unsupervised Approaches Applied to Large-Scale “In The Wild” Face Verification

Deep learning-based feature extraction methods and transfer learning have become common approaches in the field of pattern recognition. Deep convolutional neural networks trained using tripled-based loss functions allow for the generation of face embeddings, which can be directly applied to face verification and clustering. Knowledge about the ground truth of face identities might improve the effectiveness of the final classification algorithm; however, it is also possible to use ground truth clusters previously discovered using an unsupervised approach. The aim of this paper is to evaluate the potential improvement of classification results of state-of-the-art supervised classification methods trained with and without ground truth knowledge. In this study, we use two sufficiently large data sets containing more than 200,000 “taken in the wild” images, each with various resolutions, visual quality, and face poses which, in our opinion, guarantee the statistical significance of the results. We examine several clustering and supervised pattern recognition algorithms and find that knowledge about the ground truth has a very small influence on the Fowlkes–Mallows score (FMS) of the classification algorithm. In the case of the classification algorithm that obtained the highest accuracy in our experiment, the FMS improved by only 5.3% (from 0.749 to 0.791) in the first data set and by 6.6% (from 0.652 to 0.718) in the second data set. Our results show that, beside highly secure systems in which face verification is a key component, face identities discovered by unsupervised approaches can be safely used for training supervised classifiers. We also found that the Silhouette Coefficient (SC) of unsupervised clustering is positively correlated with the Adjusted Rand Index, V-measure score, and Fowlkes–Mallows score and, so, we can use the SC as an indicator of clustering performance when the ground truth of face identities is not known. All of these conclusions are important findings for large-scale face verification problems. The reason for this is the fact that skipping the verification of people’s identities before supervised training saves a lot of time and resources.

Discriminative Transfer Feature and Label Consistency for Cross-Domain Image Classification.

Visual domain adaptation aims to seek an effective transferable model for unlabeled target images by benefiting from the well-labeled source images following different distributions. Many recent efforts focus on extracting domain-invariant image representations via exploring target pseudo labels, predicted by the source classifier, to further mitigate the conditional distribution shift across domains. However, two essential factors are overlooked by most existing methods: 1) the learned transferable features should be not only domain invariant but also category discriminative; and 2) the target pseudo label is a two-edged sword to cross-domain alignment. In other words, the wrongly predicted target labels may hinder the class-wise domain matching. In this article, to address these two issues simultaneously, we propose a discriminative transfer feature and label consistency (DTLC) approach for visual domain adaptation problems, which can naturally unify cross-domain alignment with discriminative information preserved and label consistency of source and target data into one framework. To be specific, DTLC first incorporates class discriminative information by penalizing the maximum distance of data pair in the same class and the minimum distance of data pair sharing the different labels for each data into the distribution alignment of both domains. The target pseudo labels are then refined based on the label consistency within the domains. Thus, the transfer feature learning and coarse-to-fine target labels would be coupled to benefit each other in an iterative way. Comprehensive experiments on several visual cross-domain benchmarks verify that DTLC can gain remarkable margins over state-of-the-art (SOTA) nondeep visual domain adaptation methods and even be comparable to competitive deep domain adaptation ones.

Meme Kanseri Tespiti için Destek Vektör Makinası ile Alexnet Kullanarak Transfer Öğrenimi

Breast cancer is one of the leading causes of women death worldwide currently. Developing a computer-aided diagnosis system for breast cancer detection became an interesting problem for many researchers in recent years. Researchers focused on deep learning techniques for classification problems, including Convolutional Neural Networks (CNNs), which achieved great success. CNN is a particular type of deep, feedforward network that has gained attention from the research community and achieved great successes, especially in biomedical image processing. In this paper, transfer learning and deep feature extraction methods are used which adapt a pre-trained CNN model to classify breast cancer histopathological images from the publically available (BreakHis dataset). The data set includes both benign and malignant images with four different magnification factors. A patch strategy method proposed based on the extraction of image patches for training the CNN and the combination of these patches for final classification. AlexNet model is considered in this work with patch strategy, and pre-trained AlexNet is used for further fine-tuning. The obtained features are then classified by using support vector machines (SVM). The evaluation results show that the pre-trained Alexnet with SVM classification and patch strategy yields the best accuracy. Accuracy between 92% and 96% was achieved using five-fold cross-validation technique for different magnification factors.

How much off-the-shelf knowledge is transferable from natural images to pathology images?

Deep learning has achieved a great success in natural image classification. To overcome data-scarcity in computational pathology, recent studies exploit transfer learning to reuse knowledge gained from natural images in pathology image analysis, aiming to build effective pathology image diagnosis models. Since transferability of knowledge heavily depends on the similarity of the original and target tasks, significant differences in image content and statistics between pathology images and natural images raise the questions: how much knowledge is transferable? Is the transferred information equally contributed by pre-trained layers? If not, is there a sweet spot in transfer learning that balances transferred model's complexity and performance? To answer these questions, this paper proposes a framework to quantify knowledge gain by a particular layer, conducts an empirical investigation in pathology image centered transfer learning, and reports some interesting observations. Particularly, compared to the performance baseline obtained by a random-weight model, though transferability of off-the-shelf representations from deep layers heavily depend on specific pathology image sets, the general representation generated by early layers does convey transferred knowledge in various image classification applications. The trade-off between transferable performance and transferred model's complexity observed in this study encourages further investigation of specific metric and tools to quantify effectiveness of transfer learning in future.

A regression framework to head-circumference delineation from US fetal images

Background and ObjectivesMeasuring head-circumference (HC) length from ultrasound (US) images is a crucial clinical task to assess fetus growth. To lower intra- and inter-operator variability in HC length measuring, several computer-assisted solutions have been proposed in the years. Recently, a large number of deep-learning approaches is addressing the problem of HC delineation through the segmentation of the whole fetal head via convolutional neural networks (CNNs). Since the task is a edge-delineation problem, we propose a different strategy based on regression CNNs. MethodsThe proposed framework consists of a region-proposal CNN for head localization and centering, and a regression CNN for accurately delineate the HC. The first CNN is trained exploiting transfer learning, while we propose a training strategy for the regression CNN based on distance fields. ResultsThe framework was tested on the HC18 Challenge dataset, which consists of 999 training and 335 testing images. A mean absolute difference of 1.90 ( ± 1.76) mm and a Dice similarity coefficient of 97.75 ( ± 1.32) % were achieved, overcoming approaches in the literature. ConclusionsThe experimental results showed the effectiveness of the proposed framework, proving its potential in supporting clinicians during the clinical practice.

Outdoor positioning based on deep learning and wireless network fingerprint technology

Localization has attracted considerable attention in recent years because of the growth in location-based services and computer-aided technology. In particular, wireless fingerprint-based positioning techniques have received considerable attention due to their satisfactory positioning performance with smartphones. In this work, a two-level hierarchical structure positioning system is proposed to achieve satisfactory positioning accuracy. First, long-time-evolution (LTE) mobile networks are used to provide positioning information, and the collected measurements of LTE signals are converted into grayscale fingerprint images to construct the fingerprint database. To overcome the instability of LTE signals, several data enhancement methods are leveraged to increase the diversity of the fingerprint database. Second, a modified deep residual network (MResNet) coarse localizer is used to learn reliable features from the fingerprint image database. Then, inspired by the ideas of feature fusion and transfer learning, a multilayer perceptron (MLP)-based fine localizer is used to further learn the features of the fingerprint images and achieve a better positioning performance. Our experimental results convincingly reveal that the proposed positioning system can achieve satisfactory positioning performance in a variety of outdoor environments.

Transfer learning for bearing performance degradation assessment based on deep hierarchical features

Bearing design, production and complicated operating condition can lead to the scattered life cycle degradation distribution, which will bring a challenge of generalization for performance degradation assessment models. And it is costly and time-consuming to collect a large amount of labeled data for supervised diagnosis, especially when the task comes from a new operating condition. Thus in this paper, a novel bearing degradation assessment model is proposed based on transfer learning and deep hierarchical features extraction. The research of degradation assessment is transformed to the classification task of degradation pattern, which divides degradation process into the normal, slight fault, fault development and damage patterns. The hierarchical network with random weight parameters is introduced to extract the local sub-band characteristics of spectrum, in which the multiple alternately convolution and pooling layers without supervised fine-tuning are employed. Joint Geometrical and Statistical Alignment method is then utilized to obtain projected sharing feature space, and thus the knowledge of bearing degradation process is transferred to accomplish degradation pattern assessment under different operating conditions. Results of the experiments on bearing fault severity and degradation process show that the proposed method reduces the feature distribution divergence between the degradation processes and accomplishes bearing performance degradation assessment in different operating condition.

Rolling Bearing Fault Diagnosis Based on Sensitive Feature Transfer Learning and Local Maximum Margin Criterion under Variable Working Condition

In real industrial scenarios, the working conditions of bearings are variable, and it is therefore difficult for data-driven diagnosis methods based on conventional machine-learning techniques to guarantee the desirable performance of diagnosis models, as the models assume that the distributions of both the training and testing data are the same. To enhance the performance of the fault diagnosis of bearings under different working conditions, a novel diagnosis framework inspired by feature extraction, transfer learning (TL), and feature dimensionality reduction is proposed in this work, and dual-tree complex wavelet packet transform (DTCWPT) is used for signal processing. Additionally, transferable sensitive feature selection by ReliefF and the sum of mean deviation (TSFSR) is proposed to reduce the redundant information of the original feature set, to select sensitive features for fault diagnosis, and to reduce the difference between the marginal distributions of the training and testing feature sets. Furthermore, a modified feature reduction method, the local maximum margin criterion (LMMC), is proposed to acquire low-dimensional mapping for high-dimensional feature spaces. Finally, bearing vibration signals collected from two test rigs are analyzed to demonstrate the adaptability, effectiveness, and practicability of the proposed diagnosis framework. The experimental results show that the proposed method can achieve high diagnosis accuracy and has significant potential benefits in industrial applications.

Detection of COVID-19 using CXR and CT images using Transfer Learning and Haralick features.

Recognition of COVID-19 is a challenging task which consistently requires taking a gander at clinical images of patients. In this paper, the transfer learning technique has been applied to clinical images of different types of pulmonary diseases, including COVID-19. It is found that COVID-19 is very much similar to pneumonia lung disease. Further findings are made to identify the type of pneumonia similar to COVID-19. Transfer Learning makes it possible for us to find out that viral pneumonia is same as COVID-19. This shows the knowledge gained by model trained for detecting viral pneumonia can be transferred for identifying COVID-19. Transfer Learning shows significant difference in results when compared with the outcome from conventional classifications. It is obvious that we need not create separate model for classifying COVID-19 as done by conventional classifications. This makes the herculean work easier by using existing model for determining COVID-19. Second, it is difficult to detect the abnormal features from images due to the noise impedance from lesions and tissues. For this reason, texture feature extraction is accomplished using Haralick features which focus only on the area of interest to detect COVID-19 using statistical analyses. Hence, there is a need to propose a model to predict the COVID-19 cases at the earliest possible to control the spread of disease. We propose a transfer learning model to quicken the prediction process and assist the medical professionals. The proposed model outperforms the other existing models. This makes the time-consuming process easier and faster for radiologists and this reduces the spread of virus and save lives.