On Gleaning Knowledge from Multiple Domains for Active Learning

In many real-world applications, the domain of model learning (referred as source domain) is usually inconsistent with or even different from the domain of testing (referred as target domain), which makes the learnt model degenerate in target domain, i.e., the test domain. To alleviate the discrepancy between source and target domains, we propose a domain adaptation method, named as Bi-shifting Auto-Encoder network (BAE). The proposed BAE attempts to shift source domain samples to target domain, and also shift the target domain samples to source domain. The non-linear transformation of BAE ensures the feasibility of shifting between domains, and the distribution consistency between the shifted domain and the desirable domain is constrained by sparse reconstruction between them. As a result, the shifted source domain is supervised and follows similar distribution as target domain. Therefore, any supervised method can be applied on the shifted source domain to train a classifier for classification in target domain. The proposed method is evaluated on three domain adaptation scenarios of face recognition, i.e., domain adaptation across view angle, ethnicity, and imaging sensor, and the promising results demonstrate that our proposed BAE can shift samples between domains and thus effectively deal with the domain discrepancy.

Adversarial discriminative domain adaptation (ADDA) is an efficient framework for unsupervised domain adaptation in image classification, where the source and target domains are assumed to have the same classes, but no labels are available for the target domain. While ADDA has already achieved better training efficiency and competitive accuracy on image classification in comparison to other adversarial based methods, we investigate whether we can improve its performance with a new framework and new loss formulations. Following the framework of semi-supervised GANs, we first extend the discriminator output over the source classes, in order to model the joint distribution over domain and task. We thus leverage on the distribution over the source encoder posteriors (which is fixed during adversarial training) and propose maximum mean discrepancy (MMD) and reconstruction-based loss functions for aligning the target encoder distribution to the source domain. We compare and provide a comprehensive analysis of how our framework and loss formulations extend over simple multi-class extensions of ADDA and other discriminative variants of semi-supervised GANs. In addition, we introduce various forms of regularization for stabilizing training, including treating the discriminator as a denoising autoencoder and regularizing the target encoder with source examples to reduce overfitting under a contraction mapping (i.e., when the target per-class distributions are contracting during alignment with the source). Finally, we validate our framework on standard datasets like MNIST, USPS, SVHN, MNIST-M and Office-31. We additionally examine how the proposed framework benefits recognition problems based on sensing modalities that lack training data. This is realized by introducing and evaluating on a neuromorphic vision sensing (NVS) sign language recognition dataset, where the source domain constitutes emulated neuromorphic spike events converted from conventional pixel-based video and the target domain is experimental (real) spike events from an NVS camera. Our results on all datasets show that our proposal is both simple and efficient, as it competes or outperforms the state-of-the-art in unsupervised domain adaptation, such as DIFA and MCDDA, whilst offering lower complexity than other recent adversarial methods.

Multisource unsupervised domain adaptation (MUDA) is an important and challenging topic for target classification with the assistance of labeled data in source domains. When we have several labeled source domains, it is difficult to map all source domains and target domain into a common feature space for classifying the targets well. In this article, a new progressive multisource domain adaptation network (PMSDAN) is proposed to further improve the classification performance. PMSDAN mainly consists of two steps for distribution alignment. First, the multiple source domains are integrated as one auxiliary domain to match the distribution with the target domain. By doing this, we can generally reduce the distribution discrepancy between each source and target domains, as well as the discrepancy between different source domains. It can efficiently explore useful knowledge from the integrated source domain. Second, to mine assistance knowledge from each source domain as much as possible, the distribution of the target domain is separately aligned with that of each source domain. A weighted fusion method is employed to combine the multiple classification results for making the final decision. In the optimization of domain adaption, weighted hybrid maximum mean discrepancy (WHMMD) is proposed, and it considers both the interclass and intraclass discrepancies. The effectiveness of the proposed PMSDAN is demonstrated in the experiments comparing with some state-of-the-art methods.

This paper addresses the problem of land-cover map updating by classification of multitemporal remote-sensing images in the context of domain adaptation (DA). The basic assumptions behind the proposed approach are twofold. The first one is that training data (ground reference information) are available for one of the considered multitemporal acquisitions (source domain) whereas they are not for the other (target domain). The second one is that multitemporal acquisitions (i.e., target and source domains) may be characterized by different sets of classes. Unlike other approaches available in the literature, the proposed DA Bayesian classifier based on maximum a posteriori decision rule (DA-MAP) automatically identifies whether there exist differences between the set of classes in the target and source domains and properly handles these differences in the updating process. The proposed method was tested in different scenarios of increasing complexity related to multitemporal image classification. Experimental results on medium-resolution and very high resolution multitemporal remote-sensing data sets confirm the effectiveness and the reliability of the proposed DA-MAP classifier.

This paper introduces a new method to solve the cross-domain recognition problem. Different from the traditional domain adaption methods which rely on a global domain shift for all classes between the source and target domains, the proposed method is more flexible to capture individual class variations across domains. By adopting a natural and widely used assumption that the data samples from the same class should lay on an intrinsic low-dimensional subspace, even if they come from different domains, the proposed method circumvents the limitation of the global domain shift, and solves the cross-domain recognition by finding the joint subspaces of the source and target domains. Specifically, given labeled samples in the source domain, we construct a subspace for each of the classes. Then we construct subspaces in the target domain, called anchor subspaces, by collecting unlabeled samples that are close to each other and are highly likely to belong to the same class. The corresponding class label is then assigned by minimizing a cost function which reflects the overlap and topological structure consistency between subspaces across the source and target domains, and within the anchor subspaces, respectively. We further combine the anchor subspaces to the corresponding source subspaces to construct the joint subspaces. Subsequently, one-versus-rest support vector machine classifiers are trained using the data samples belonging to the same joint subspaces and applied to unlabeled data in the target domain. We evaluate the proposed method on two widely used datasets: 1) object recognition dataset for computer vision tasks and 2) sentiment classification dataset for natural language processing tasks. Comparison results demonstrate that the proposed method outperforms the comparison methods on both datasets.

Multi-source adversarial transfer learning for ultrasound image segmentation with limited similarity

Domain-transfer learning is a machine learning task to explore a source domain data set to help the learning problem in a target domain. Usually, the source domain has sufficient labeled data, while the target domain does not. In this paper, we propose a novel domain-transfer convolutional model by mapping a target domain data sample to a proxy in the source domain and applying a source domain model to the proxy for the purpose of prediction. In our framework, we firstly represent both source and target domains to feature vectors by two convolutional neural networks and then construct a proxy for each target domain sample in the source domain space. The proxy is supposed to be matched to the corresponding target domain sample convolutional representation vector well. To measure the matching quality, we proposed to maximize their squared-loss mutual information (SMI) between the proxy and target domain samples. We further develop a novel neural SMI estimator based on a parametric density ratio estimation function. Moreover, we also propose to minimize the classification error of both source domain samples and target domain proxies. The classification responses are also smoothened by manifolds of both the source domain and proxy space. By minimizing an objective function of SMI, classification error, and manifold regularization, we learn the convolutional networks of both source and target domains. In this way, the proxy of a target domain sample can be matched to the source domain data and thus benefits from the rich supervision information of the source domain. We design an iterative algorithm to update the parameters alternately and test it over benchmark data sets of abnormal behavior detection in video, Amazon product reviews sentiment analysis, etc.

Transient stability assessment (TSA) based on deep learning has recently attracted broad attention. The offline training of deep learning model typically requires a large number of labeled samples, and the training process is very time-consuming. Moreover, it is difficult to ensure that the trained model is suitable for all the operation conditions of power systems. In order to dramatically reduce the computation cost, and further enhance the adaptability of the model, an innovative method is proposed based on the active transfer learning (TL). First, the deep belief network (DBN) is designed as the base model and trained with a large number of labeled samples to obtain better evaluation performance for TSA. Second, when the topologies or operation conditions change substantially, active learning with information entropy is exploited to select the minimum number of the most informative samples, which effectively reduces the time of samples generation. Third, Maximum Mean Discrepancy (MMD) is calculated to select different transfer schemes. Only the last BP layer needs to be fine-tuned for the scenarios with a low MMD, but for those with high MMD, the entire network needs to be fine-tuned. The transfer time is greatly shortened on the premise of ensuring the transfer effect. The comprehensive experimental results of three power systems illustrate the effectiveness of the proposed method. Note to Practitioners—This research is motivated by automatic adaptive assessment of power system transient stability. The main objective is to realize online fast and continuous prediction with limited samples. This kind of demand is very normal in the industrial systems where it is expensive, time-consuming and difficult to obtain labeled samples. This paper proposes an active transfer learning method to select critical samples and update the deep learning model, which greatly reduces the time of samples generation and improves the updating efficiency of model. The proposed method is also applicable to make classification prediction or to address regression problems of industrial scenarios based on deep learning models.

This paper addresses the problem of land-cover maps updating by classifying multitemporal remote sensing images (i.e., images acquired on the same area at different times) in the context of change-detection-driven active transfer learning. The proposed method is based on the assumption that training samples are available for one of the available multitemporal images (i.e., source domain), whereas they are not for the others (i.e., target domain). In order to effectively classify the target domain (i.e., update the maps obtained for the source domain according to the new information brought from another acquisition) we present a novel approach to automatically define a training set for the target domain taking advantage of its temporal correlation with the source domain. The proposed method is based on four steps. In the first step unsupervised change detection is applied to multitemporal images (i.e., target and source domains). Labels of detected unchanged training samples are propagated from the source to the target domain in the second step, thus becoming its initial training set. In the third step, changed areas are statistically compared with land-cover classes in the target domain training set. This information is used to drive the initial training set expansion by Active Learning (AL). In the first expansion iterations priority is given to samples detected as being changed, in the next ones the most informative samples are selected from a pool including both changed and unchanged unlabeled samples (i.e., priority is removed). At convergence of the AL process, the target image is classified (fourth step). To this, in this paper we use a Support Vector Machine classifier. Experimental results show that transferring the class-labels from source domain to target domain provides a reliable initial training set and that the priority rule for AL involves a faster convergence to the desired accuracy with respect to standard AL.

Unsupervised domain adaptation aims to transfer knowledge from labeled source domain to unlabeled target domain. Recently, multisource domain adaptation (MDA) has begun to attract attention. Its performance should go beyond simply mixing all source domains together for knowledge transfer. In this article, we propose a novel prototype-based method for MDA. Specifically, for solving the problem that the target domain has no label, we use the prototype to transfer the semantic category information from source domains to target domain. First, a feature extraction network is applied to both source and target domains to obtain the extracted features from which the domain-invariant features and domain-specific features will be disentangled. Then, based on these two kinds of features, the named inherent class prototypes and domain prototypes are estimated, respectively. Then a prototype mapping to the extracted feature space is learned in the feature reconstruction process. Thus, the class prototypes for all source and target domains can be constructed in the extracted feature space based on the previous domain prototypes and inherent class prototypes. By forcing the extracted features are close to the corresponding class prototypes for all domains, the feature extraction network is progressively adjusted. In the end, the inherent class prototypes are used as a classifier in the target domain. Our contribution is that through the inherent class prototypes and domain prototypes, the semantic category information from source domains is transformed into the target domain by constructing the corresponding class prototypes. In our method, all source and target domains are aligned twice at the feature level for better domain-invariant features and more closer features to the class prototypes, respectively. Several experiments on public data sets also prove the effectiveness of our method.

Recently, domain adaptation has gained great popularity, while most researchers are focusing on domains in homogenous modalities, e.g., image domains. In reality, heterogeneous domains are pretty common and more challenging. In this paper, we present MCD-JD—a Maximum Classifier Discrepancy model which considers the joint distribution of the source and target domain data for heterogeneous domain adaption. MCD-JD derives from Generative Adversarial Networks (GAN) consisting of two parts, i.e., minimizing the discrepancy of joint distribution, and maximizing classifier discrepancy. Specifically, the first part uses the Maximum Mean Discrepancy (MMD) regularization to adapt the data distributions between source and target domains. The second part utilizes two different classifiers to maximize their discrepancy of making predictions on the target domain data, which further minimizes the discrepancy of data distributions between source and target domains. We collect a dataset depicting real-world events (e.g., protests, explosions, etc.) from multiple heterogeneous data domains, including news media textual articles, social media (Flickr) images, and YouTube videos. Extensive experiments conducted on the real-world dataset manifest the effectiveness of MCD-JD, which outperforms state-of-the-art benchmark models.

Active learning algorithms propose what data should be labeled given a pool of unlabeled data. Instead of selecting randomly what data to annotate, active learning strategies aim to select data so as to get a good predictive model with as little labeled samples as possible. Single-shot batch active learners select all samples to be labeled in a single step, before any labels are observed. We study single-shot active learners that minimize generalization bounds to select a representative sample, such as the maximum mean discrepancy (MMD) active learner. We prove that a related bound, the discrepancy, provides a tighter worst-case bound. We study these bounds probabilistically, which inspires us to introduce a novel bound, the nuclear discrepancy (ND). The ND bound is tighter for the expected loss under optimistic probabilistic assumptions. Our experiments show that the MMD active learner performs better than the discrepancy in terms of the mean squared error, indicating that tighter worst case bounds do not imply better active learning performance. The proposed active learner improves significantly upon the MMD and discrepancy in the realizable setting and a similar trend is observed in the agnostic setting, showing the benefits of a probabilistic approach to active learning. Our study highlights that assumptions underlying generalization bounds can be equally important as bound-tightness, when it comes to active learning performance. Code for reproducing our experimental results can be found at https://github.com/tomviering/NuclearDiscrepancy.

In an indoor positioning system (IPS), transfer learning (TL) methods are commonly used to predict the location of mobile devices under the assumption that all training instances of the target domain are given in advance. However, this assumption has been criticized for its shortcomings in dealing with the problem of signal distribution variations, especially in a dynamic indoor environment. The reasons are: collecting a sufficient number of training instances is costly, the training instances may arrive online, the feature spaces of the target and source domains may be different, and negative knowledge may be transferred in the case of a redundant source domain. In this work, we proposed an online heterogeneous transfer learning (OHetTLAL) algorithm for IPS-based RSS fingerprinting to improve the positioning performance in the target domain by fusing both source and target domain knowledge. The source domain was refined based on the target domain to avoid negative knowledge transfer. The co-occurrence measure of the feature spaces (Cmip) was used to derive the homogeneous new feature spaces, and the features with higher weight values were selected for training the classifier because they could positively affect the location prediction of the target. Thus, the objective function was minimized over the new feature spaces. Extensive experiments were conducted on two real-world scenarios of datasets, and the predictive power of the different modeling techniques were evaluated for predicting the location of a mobile device. The results have revealed that the proposed algorithm outperforms the state-of-the-art methods for fingerprint-based indoor positioning and is found robust to changing environments. Moreover, the proposed algorithm is not only resilient to fluctuating environments but also mitigates the model's overfitting problem.

Open set transfer learning through distribution driven active learning

Despite many recent advances, state-of-the-art recommender systems still struggle to achieve good performance with sparse datasets. To address the sparsity issue, transfer learning techniques have been investigated for recommender systems, but they tend to impose strict constraints on the content and structure of the data in the source and target domains. For transfer learning methods to work well, there should normally be homogeneity between source and target domains, or a high degree of overlap between the source and target items. In this paper we propose a novel transfer learning framework for mitigating the effects of sparsity and insufficient data. Our method requires neither homogeneity nor overlap between the source and target domains. We describe and evaluate a shared parameter auto-encoder to jointly learn representations of user/item aspects in two domains, applying Maximum Mean Discrepancy (MMD) loss during training to ensure that the source and target representations are similar in the distribution space. The approach is evaluated using a number of benchmark datasets to demonstrate improved recommendation performance when learned representations are used in collaborative filtering. The code used for this work is available on github.com.