Performance Of Transfer Learning Research Articles

Enhancing Domain Diversity of Transfer Learning-Based SSVEP-BCIs by the Reconstruction of Channel Correlation.

The application of transfer learning, specifically pre-training and fine-tuning, in steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) has been demonstrated to effectively improve the classification performance of deep learning methods with limited calibration data. However, effectively learning task-related knowledge from source domains during the pre-training phase remains challenging. To address this issue, this study proposes an effective data augmentation method called Reconstruction of Channel Correlation (RCC) to optimize the utilization of the source domain data. Concretely, RCC reconstructs training samples using probabilistically mixed eigenvector matrices derived from covariance matrices across source domains. This process manipulates the channel correlation of training samples, implicitly creating novel synthesized domains. By increasing the diversity of source domains, RCC aims to enhance the domain generalizability of the pre-trained model. The effectiveness of RCC is validated through subject-independent and subject-adaptive classification experiments. The results of subject-independent classification demonstrate that RCC significantly improves the classification performance of the pre-trained model on unseen target subjects. Moreover, when compared to the fine-tuning process using the RCC-absent pre-trained model, the fine-tuning process using the RCC-enhanced pre-trained model yields significantly improved performance in the subject-adaptive classification. RCC proves to enhance the performance of transfer learning by optimizing the utilization of the source domain data. The RCC-enhanced transfer learning has the potential to facilitate the practical implementation of SSVEP-BCIs in real-world scenarios.

Self-Supervised In-Domain Representation Learning for Remote Sensing Image Scene Classification

Transferring the ImageNet pre-trained weights to the various remote sensing tasks has produced acceptable results and reduced the need for labeled samples. However, the domain differences between ground imageries and remote sensing images cause the performance of such transfer learning to be limited. The difficulty of annotating remote sensing images is well-known as it requires domain experts and more time, whereas unlabelled data is readily available. Recently, self-supervised learning, which is a subset of unsupervised learning, emerged and significantly improved representation learning. Recent research has demonstrated that self-supervised learning methods capture visual features that are more discriminative and transferable than the supervised ImageNet weights. We are motivated by these facts to pre-train the in-domain representations of remote sensing imagery using contrastive self-supervised learning and transfer the learned features to other related remote sensing datasets. Specifically, we used the SimSiam algorithm to pre-train the in-domain knowledge of remote sensing datasets and then transferred the obtained weights to the other scene classification datasets. Thus, we have obtained state-of-the-art results on five land cover classification datasets with varying numbers of classes and spatial resolutions. In addition, by conducting appropriate experiments, including feature pre-training using datasets with different attributes, we have identified the most influential factors that make a dataset a good choice for obtaining in-domain features. We have transferred the features obtained by pre-training SimSiam on remote sensing datasets to various downstream tasks and used them as initial weights for fine-tuning. Moreover, we have linearly evaluated the obtained representations in cases where the number of samples per class is limited. Our experiments have demonstrated that using a higher-resolution dataset during the self-supervised pre-training stage results in learning more discriminative and general representations.

Adaptive fault diagnosis for high-purity carbonate process based on unsupervised and transfer learning

Reactive distillation integrates reaction and distillation to achieve process intensification. And the increasing process complexity urgently requires fault diagnosis system to ensure its safe and efficient operation. However, establishing fault diagnosis model is facing intractable challenges, e.g. scarce labeled data, operating mode changeover due to product proportion variation. To tackle absent labeled data problem, pseudo-labeled database is established by integrating data mining, density-based spatial clustering of applications with noise algorithm and process knowledge, based on which unsupervised deep learning is developed. Besides, to make intelligent fault diagnosis system adaptive for multimode operation changeover, transfer learning with domain adaptation strategy is adopted, which could mitigate different data distributions in various operating states and transfer knowledge learned from source domain to target domain. Taking carbonate ester process with reactive distillation as benchmark, the performance and superiority of unsupervised learning and transfer learning is verified and demonstrated in solving the corresponding puzzle.

Adaptive fine-tuning based transfer learning for the identification of MGMT promoter methylation status.

Background.Glioblastoma Multiforme (GBM) is an aggressive form of malignant brain tumor with a generally poor prognosis.O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation has been shown to be a predictive bio-marker for resistance to treatment of GBM, but it is invasive and time-consuming to determine methylation status. There has been effort to predict the MGMT methylation status through analyzing MRI scans using machine learning, which only requires pre-operative scans that are already part of standard-of-care for GBM patients.Purpose.To improve the performance of conventional transfer learning in the identification of MGMT promoter methylation status, we developed a 3D SpotTune network with adaptive fine-tuning capability. Using the pretrained weights of MedicalNet with the SpotTune network, we compared its performance with a randomly initialized network for different combinations of MR modalities.Methods.Using a ResNet50 as the base network, three categories of networks are created: (1) A 3D SpotTune network to process volumetric MR images, (2) a network with randomly initialized weights, and (3) a network pre-trained on MedicalNet. These three networks are trained and evaluated using a public GBM dataset provided by the University of Pennsylvania. The MRI scans from 240 patients are used, with 11 different modalities corresponding to a set of perfusion, diffusion, and structural scans. The performance is evaluated using 5-fold cross validation with a hold-out testing dataset.Results.The SpotTune network showed better performance than the randomly initialized network. The best performing SpotTune model achieved an area under the Receiver Operating Characteristic curve (AUC), average precision of the precision-recall curve (AP), sensitivity, and specificity values of 0.6604, 0.6179, 0.6667, and 0.6061 respectively.Conclusions.SpotTune enables transfer learning to be adaptive to individual patients, resulting in improved performance in predicting MGMT promoter methylation status in GBM using equivalent MRI modalities as compared to a randomly initialized network.

Fixed-Wing UAV Pose Estimation Using a Self-Organizing Map and Deep Learning

In many Unmanned Aerial Vehicle (UAV) operations, accurately estimating the UAV’s position and orientation over time is crucial for controlling its trajectory. This is especially important when considering the landing maneuver, where a ground-based camera system can estimate the UAV’s 3D position and orientation. A Red, Green, and Blue (RGB) ground-based monocular approach can be used for this purpose, allowing for more complex algorithms and higher processing power. The proposed method uses a hybrid Artificial Neural Network (ANN) model, incorporating a Kohonen Neural Network (KNN) or Self-Organizing Map (SOM) to identify feature points representing a cluster obtained from a binary image containing the UAV. A Deep Neural Network (DNN) architecture is then used to estimate the actual UAV pose based on a single frame, including translation and orientation. Utilizing the UAV Computer-Aided Design (CAD) model, the network structure can be easily trained using a synthetic dataset, and then fine-tuning can be done to perform transfer learning to deal with real data. The experimental results demonstrate that the system achieves high accuracy, characterized by low errors in UAV pose estimation. This implementation paves the way for automating operational tasks like autonomous landing, which is especially hazardous and prone to failure.

Solving deep-learning density functional theory via variational autoencoders

In recent years, machine learning models, chiefly deep neural networks, have revealed suited to learn accurate energy-density functionals from data. However, problematic instabilities have been shown to occur in the search of ground-state density profiles via energy minimization. Indeed, any small noise can lead astray from realistic profiles, causing the failure of the learned functional and, hence, strong violations of the variational property. In this article, we employ variational autoencoders (VAEs) to build a compressed, flexible, and regular representation of the ground-state density profiles of various quantum models. Performing energy minimization in this compressed space allows us to avoid both numerical instabilities and variational biases due to excessive constraints. Our tests are performed on one-dimensional single-particle models from the literature in the field and, notably, on a three-dimensional disordered potential. In all cases, the ground-state energies are estimated with errors below the chemical accuracy and the density profiles are accurately reproduced without numerical artifacts. Furthermore, we show that it is possible to perform transfer learning, applying pre-trained VAEs to different potentials.

Global contrast-masked autoencoders are powerful pathological representation learners

Using digital pathology slide scanning technology, artificial intelligence algorithms, particularly deep learning, have achieved significant results in the field of computational pathology. Compared to other medical images, pathology images are more difficult to annotate, and thus, there is an extreme lack of available datasets for conducting supervised learning to train robust deep learning models. In this paper, we introduce a self-supervised learning (SSL) model, the Global Contrast-masked Autoencoder (GCMAE), designed to train encoders to capture both local and global features of pathological images and significantly enhance the performance of transfer learning across datasets. Our study demonstrates the capability of the GCMAE to learn transferable representations through extensive experiments on three distinct disease-specific hematoxylin and eosin (H&E)-stained pathology datasets: Camelyon16, NCT-CRC, and BreakHis. Moreover, we propose an effective automated pathology diagnosis process based on the GCMAE for clinical applications. The source code of this paper is publicly available at https://github.com/StarUniversus/gcmae.

An LSTM-based adversarial variational autoencoder framework for self-supervised neural decoding of behavioral choices

Objective.This paper presents data-driven solutions to address two challenges in the problem of linking neural data and behavior: (1) unsupervised analysis of behavioral data and automatic label generation from behavioral observations, and (2) extraction of subject-invariant features for the development of generalized neural decoding models. Approach. For behavioral analysis and label generation, an unsupervised method, which employs an autoencoder to transform behavioral data into a cluster-friendly feature space is presented. The model iteratively refines the assigned clusters with soft clustering assignment loss, and gradually improves the learned feature representations. To address subject variability in decoding neural activity, adversarial learning in combination with a long short-term memory-based adversarial variational autoencoder (LSTM-AVAE) model is employed. By using an adversary network to constrain the latent representations, the model captures shared information among subjects’ neural activity, making it proper for cross-subject transfer learning. Main results. The proposed approach is evaluated using cortical recordings of Thy1-GCaMP6s transgenic mice obtained via widefield calcium imaging during a motivational licking behavioral experiment. The results show that the proposed model achieves an accuracy of 89.7% in cross-subject neural decoding, outperforming other well-known autoencoder-based feature learning models. These findings suggest that incorporating an adversary network eliminates subject dependency in representations, leading to improved cross-subject transfer learning performance, while also demonstrating the effectiveness of LSTM-based models in capturing the temporal dependencies within neural data. Significance. Results demonstrate the feasibility of the proposed framework in unsupervised clustering and label generation of behavioral data, as well as achieving high accuracy in cross-subject neural decoding, indicating its potentials for relating neural activity to behavior.

A systematic evaluation of Euclidean alignment with deep learning for EEG decoding

Objective: Electroencephalography signals are frequently used for various Brain–Computer interface (BCI) tasks. While deep learning (DL) techniques have shown promising results, they are hindered by the substantial data requirements. By leveraging data from multiple subjects, transfer learning enables more effective training of DL models. A technique that is gaining popularity is Euclidean alignment (EA) due to its ease of use, low computational complexity, and compatibility with DL models. However, few studies evaluate its impact on the training performance of shared and individual DL models. In this work, we systematically evaluate the effect of EA combined with DL for decoding BCI signals. Approach: We used EA as a pre-processing step to train shared DL models with data from multiple subjects and evaluated their transferability to new subjects. Main results: Our experimental results show that it improves decoding in the target subject by 4.33% and decreases convergence time by more than 70%. We also trained individual models for each subject to use as a majority-voting ensemble classifier. In this scenario, using EA improved the 3-model ensemble accuracy by 3.71%. However, when compared to the shared model with EA, the ensemble accuracy was 3.62% lower. Significance: EA succeeds in the task of improving transfer learning performance with DL models and, could be used as a standard pre-processing technique.

Learning-Based Approach for Automated Surface Inspection with Industrial Tomography Imaging

Abstract In recent years, advanced deep learning techniques have emerged as pivotal tools in enabling the development of robust vision-based solutions for steel surface inspection. This resulted in enhanced inspection accuracy, all while significantly reducing costs in the manufacturing industry. However, the lack of actual steel surface defects datasets currently places a certain constraint on further research into classifying those anomalies. As a consequence, the Convolutional Neural Network (CNN) technique, known for its prowess in image-related tasks, faces certain challenges, especially in classifying less common defects. This work proposes a novel hybrid CNN model with a Support Vector Machine (SVM) classifier at the output layer for surface defects classification. The features extracted from the pre-trained ResNet152 and EfficientB0 CNN algorithms are concatenated and fed to the SVM layer for classification. Extensive experiments on a merged dataset consisting of the publicly available Northeastern University (NEU) dataset and Xsteel surface defect dataset (X-SDD) are carried out and the accuracy and F1 scores are calculated for performance evaluation. The merged dataset contains eleven typical defect types with a total of 2660 defect images. Then, the adopted algorithm is compared with ten fine-tuned deep learning models to evaluate the performance of transfer learning for steel defect detection and identification. The evaluation results show that the deep feature extraction and SVM classification produced better results than the transfer learning. Finally, the proposed classifier model is validated on a newly collected dataset from a Computed Tomography scanner with an accuracy reaching over 96%.

Transfer learning enables identification of multiple types of RNA modifications using nanopore direct RNA sequencing

Nanopore direct RNA sequencing (DRS) has emerged as a powerful tool for RNA modification identification. However, concurrently detecting multiple types of modifications in a single DRS sample remains a challenge. Here, we develop TandemMod, a transferable deep learning framework capable of detecting multiple types of RNA modifications in single DRS data. To train high-performance TandemMod models, we generate in vitro epitranscriptome datasets from cDNA libraries, containing thousands of transcripts labeled with various types of RNA modifications. We validate the performance of TandemMod on both in vitro transcripts and in vivo human cell lines, confirming its high accuracy for profiling m6A and m5C modification sites. Furthermore, we perform transfer learning for identifying other modifications such as m7G, Ψ, and inosine, significantly reducing training data size and running time without compromising performance. Finally, we apply TandemMod to identify 3 types of RNA modifications in rice grown in different environments, demonstrating its applicability across species and conditions. In summary, we provide a resource with ground-truth labels that can serve as benchmark datasets for nanopore-based modification identification methods, and TandemMod for identifying diverse RNA modifications using a single DRS sample.

Transfer Learning for Efficient Intent Prediction in Lower-Limb Prosthetics: A Strategy for Limited Datasets.

This paper presents a transfer learning method to enhance locomotion intent prediction in novel transfemoral amputee subjects, particularly in data-sparse scenarios. Transfer learning is done with three pre-trained models trained on separate datasets: transfemoral amputees, able-bodied individuals, and a mixed dataset of both groups. Each model is subsequently fine-tuned using data from a new transfemoral amputee subject. While subject-dependent models, trained and tested using individual user data, can achieve the least error rate, they require extensive training datasets. In contrast, our transfer learning approach yields comparable error rates while requiring significantly less data. This highlights the benefit of using preexisting, pre-trained features when data is scarce. As anticipated, the performance of transfer learning improves as more data from the subject is made available. We also explore the performance of the intent prediction system under various sensor configurations. We identify that a combination of a thigh inertial measurement unit and load cell offers a practical and efficient choice for sensor setup. These findings underscore the potential of transfer learning as a powerful tool for enhancing intent prediction accuracy for new transfemoral amputee subjects, even under data-limited conditions.

Physics-Informed Transfer Learning to Enhance Sleep Staging.

At-home sleep staging using wearable medical sensors poses a viable alternative to in-hospital polysomnography due to its lower cost and lower disruption to the daily lives of patients, especially in the case of long-term monitoring. Machine learning with wearables however is difficult due to the paucity of data from wearable sensors, making automation a challenge. Transfer learning from hospital polysomnograms can boost performance, but is still hindered by differences between wearable and in-hospital EEG resulting in part from differing electrode placement. We improve transfer learning performance by using electrophysiological models of a human head to generate synthetic EEG resembling EEG from a wearable sensor. The data generation method utilizes Low-Resolution Electromagnetic Tomography Analysis (LORETA). Real EEG from standard in- hospital recordings is first mapped to point currents within the brain using LORETA, after which the point currents are used to estimate EEG that would have been recorded using a wearable sensor at any given point on the head. Augmenting source datasets with synthetic data statistically significantly boosted accuracy on a wearable sleep staging task from 80.8% to 81.3% on average, depending on the transfer learning parameters and data sources. Machine learning performance can be improved using data synthesized using physical models. Our approach represents a new form of transfer learning and demonstrates that incorporating domain knowledge of electrophysiological modeling can improve machine learning results for sleep staging tasks. We expect this approach to be particularly useful for EEG data which is hard to collect, or which is obtained using unusual electrode configurations.

Self-supervised contrastive learning for EEG-based cross-subject motor imagery recognition.

Objective. The extensive application of electroencephalography (EEG) in brain-computer interfaces (BCIs) can be attributed to its non-invasive nature and capability to offer high-resolution data. The acquisition of EEG signals is a straightforward process, but the datasets associated with these signals frequently exhibit data scarcity and require substantial resources for proper labeling. Furthermore, there is a significant limitation in the generalization performance of EEG models due to the substantial inter-individual variability observed in EEG signals.Approach. To address these issues, we propose a novel self-supervised contrastive learning framework for decoding motor imagery (MI) signals in cross-subject scenarios. Specifically, we design an encoder combining convolutional neural network and attention mechanism. In the contrastive learning training stage, the network undergoes training with the pretext task of data augmentation to minimize the distance between pairs of homologous transformations while simultaneously maximizing the distance between pairs of heterologous transformations. It enhances the amount of data utilized for training and improves the network's ability to extract deep features from original signals without relying on the true labels of the data.Main results. To evaluate our framework's efficacy, we conduct extensive experiments on three public MI datasets: BCI IV IIa, BCI IV IIb, and HGD datasets. The proposed method achieves cross-subject classification accuracies of 67.32%, 82.34%, and 81.13%on the three datasets, demonstrating superior performance compared to existing methods.Significance. Therefore, this method has great promise for improving the performance of cross-subject transfer learning in MI-based BCI systems.

CT-based intratumoral and peritumoral deep transfer learning features prediction of lymph node metastasis in non-small cell lung cancer.

The main metastatic route for lung cancer is lymph node metastasis, and studies have shown that non-small cell lung cancer (NSCLC) has a high risk of lymph node infiltration. This study aimed to compare the performance of handcrafted radiomics (HR) features and deep transfer learning (DTL) features in Computed Tomography (CT) of intratumoral and peritumoral regions in predicting the metastatic status of NSCLC lymph nodes in different machine learning classifier models. We retrospectively collected data of 199 patients with pathologically confirmed NSCLC. All patients were divided into training (n = 159) and validation (n = 40) cohorts, respectively. The best HR and DTL features in the intratumoral and peritumoral regions were extracted and selected, respectively. Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Light Gradient Boosting Machine (Light GBM), Multilayer Perceptron (MLP), and Logistic Regression (LR) models were constructed, and the performance of the models was evaluated. Among the five models in the training and validation cohorts, the LR classifier model performed best in terms of HR and DTL features. The AUCs of the training cohort were 0.841 (95% CI: 0.776-0.907) and 0.955 (95% CI: 0.926-0.983), and the AUCs of the validation cohort were 0.812 (95% CI: 0.677-0.948) and 0.893 (95% CI: 0.795-0.991), respectively. The DTL signature was superior to the handcrafted radiomics signature. Compared with the radiomics signature, the DTL signature constructed based on intratumoral and peritumoral areas in CT can better predict NSCLC lymph node metastasis.

SegIns: A simple extension to instance discrimination task for better localization learning

Recent self-supervised learning methods, where instance discrimination task is a fundamental way of pretraining convolutional neural networks (CNN), excel in transfer learning performance. Even though instance discrimination task is a well suited pretraining method for classification with its image-level learning, lack of dense representation learning makes it sub-optimal for localization tasks such as object detection. In this paper, we aim to mitigate this shortcoming of instance discrimination task by extending it to jointly learn dense representations alongside image-level representations. We add a segmentation branch parallel to the image-level learning to predict class-agnostic masks, enhancing location-awareness of the representations. We show the effectiveness of our pretraining approach on localization tasks by transferring the learned representations to object detection and segmentation tasks, providing relative improvements by up to 1.7% AP on PASCAL VOC and 0.8% AP on COCO object detection, 0.8% AP on COCO instance segmentation and 3.6% mIoU on PASCAL VOC semantic segmentation respectively.

Improved Adversarial Transfer Network for Bearing Fault Diagnosis under Variable Working Conditions

Bearings are one of the critical components of rotating machinery, and their failure can cause catastrophic consequences. In this regard, previous studies have proposed a variety of intelligent diagnosis methods. Most existing bearing fault diagnosis methods implicitly assume that the training and test sets are from the same distribution. However, in real scenarios, bearings have been working in complex and changeable working environments for a long time. The data during their working processes and the data used for model training cannot meet this condition. This paper proposes an improved adversarial transfer network for fault diagnosis under variable working conditions. Specifically, this paper combines an adversarial transfer network with a short-time Fourier transform to obtain satisfactory results with the lighter network. Then, this paper employs a channel attention module to enhance feature fusion. Moreover, this paper designs a novel domain discrepancy hybrid metric loss to improve model transfer learning performance. Finally, this paper verifies the method’s effectiveness on three datasets, including dual-rotor, a Case Western Reserve University dataset and the Ottawa dataset. The proposed method achieves average accuracy, surpassing other methods, and shows better domain alignment capabilities.

Procedural Content Generation via Knowledge Transformation (PCG-KT)

We introduce the concept of Procedural Content Generation via Knowledge Transformation (PCG-KT), a new lens and framework for characterizing PCG methods and approaches in which content generation is enabled by the process of knowledge transformation—transforming knowledge derived from one domain in order to apply it in another. Our work is motivated by a substantial number of recent PCG works that focus on generating novel content via repurposing derived knowledge. Such works have involved, for example, performing transfer learning on models trained on one game's content to adapt to another game's content, as well as recombining different generative distributions to blend the content of two or more games. Such approaches arose in part due to limitations in PCG via Machine Learning (PCGML) such as producing generative models for games lacking training data and generating content for entirely new games. In this paper, we categorize such approaches under this new lens of PCG-KT by offering a definition and framework for describing such methods and surveying existing works using this framework. Finally, we conclude by highlighting open problems and directions for future research in this area.

Variational inference-based transfer learning for profile monitoring with incomplete data

Profile monitoring is a widely used tool in quality control. The rapid development of sensor technology has created unprecedented opportunities for multi-channel profile data collection, which motivates the modeling and transfer learning for multi-profile data. The Multi-output Gaussian Process (MGP) is often used for multi-profile data, due to its flexible modeling capability and elegant mathematical properties. However, two practical concerns limit the broader application of MGP for transfer learning and monitoring of multi-profile data: high computational complexity and data incompleteness. In this article, we propose a Variational Inference (VI)-based MGP framework to facilitate transfer learning and profile monitoring using incomplete profile data. The proposed framework features a specially designed convolutional structure for constructing an explicit covariance relationship between the inducing variables in VI and the MGP in multi-profile data. This structure inspires a comprehensive solution to both computational complexity and data incompleteness in modeling multi-profile data, which facilities the transfer learning for profile monitoring. Various numerical studies and one real case study are conducted to demonstrate and compare the transfer learning and monitoring performance of the proposed method. The results show the method can achieve superior monitoring performance while maintain a very low level of computational complexity.

Fine classification of crops based on an inductive transfer learning method with compact polarimetric SAR images

ABSTRACT Compact polarimetric synthetic aperture radar (CP SAR) reduces fully polarimetric SAR system complexity and expands the imaging swath. Generally, fine classification of crop types relies on many labeled training samples. However, due to the temporal interval of crop phenology and ground environment variations over time, training samples from one dataset usually perform poorly for another. Therefore, in this study, transfer learning is introduced to crop classification to ensure classification accuracy by improving reusability of training samples. A stable and robust inductive transfer learning method, i.e. the Transfer Bagging-based Ensemble Learning (TBEL) algorithm, is proposed. The main idea is to select an adequate number of representative samples from unlabeled datasets to characterize each class in the target domain based on limited labeled samples and construct a classifier set to classify the target domain. This study investigates CP SAR data performance in transfer learning for crop classification. The proposed algorithm in the experimental study is compared with six typical methods (Subspace Alignment (SA), CORrelation ALignment (CORAL), Joint Distribution Adaptation (JDA), Balanced Distribution Adaptation (BDA), Transfer Bagging (TrBagg), and Bagging-based Ensemble Transfer Learning (BETL)). The experimental results show that the crop classification accuracy based on the TBEL algorithm is more stable, with an improved overall classification accuracy of 2–6%. Classifying the same rice harvest stage in the cross-year domain has the highest overall accuracy of 92.2%. Wheat fields in different scenes are also classified. Based on the TBEL algorithm, the overall classification accuracy improves by 1–10% compared with typical methods, with an accuracy of at least 87.6%. Furthermore, by testing the CP mode classification performance over various crops in transfer learning, we find that the circular CP mode performs better than the linear mode in most cases. This conclusion agrees with single-scene applications and was first verified in transfer learning.