Federated Active Learning for Target Domain Generalisation

Supervised learning methods require sufficient labeled examples to learn a good model for classification or regression. However, available labeled data are insufficient in many applications. Active learning (AL) and domain adaptation (DA) are two strategies to minimize the required amount of labeled data for model training. AL requires the domain expert to label a small number of highly informative examples to facilitate classification, while DA involves tuning the source domain knowledge for classification on the target domain. In this paper, we demonstrate how AL can efficiently minimize the required amount of labeled data for DA. Since the source and target domains usually have different distributions, it is possible that the domain expert may not have sufficient knowledge to answer each query correctly. We exploit our active DA framework to handle incorrect labels provided by domain experts. Experiments with multimedia data demonstrate the efficiency of our proposed framework for active DA with noisy labels.

This paper presents a novel system for automatically updating land-cover maps by classifying multitemporal remote sensing images (i.e., satellite images acquired on the same area at different times). The proposed system assumes that a training set is not available for the image to be classified (i.e., the target domain), but it is accessible for another image (i.e., the source domain) acquired on the same area at a different time. Under this assumption the proposed method aims at defining a reliable training set for the target domain on the basis of three steps. In the first step, unsupervised change detection is applied to target and source domains. Then class labels of detected unchanged training samples are transferred from the source to the target domain to initialize the training set for the target domain. In the second step, the training set is enriched by labeling new samples according to a novel active learning method. In the last step, the target domain is classified by a cascade classifier that exploits the temporal correlation between domains.

Cross-domain video concept detection: A joint discriminative and generative active learning approach

How can a doctor diagnose new diseases with little historical knowledge, which are emerging over time? Active learning is a promising way to address the problem by querying the most informative samples. Since the diagnosed cases for new disease are very limited, gleaning knowledge from other domains (classical prescriptions) to prevent the bias of active leaning would be vital for accurate diagnosis. In this paper, a framework that attempts to glean knowledge from multiple domains for active learning by querying the most uncertain and representative samples from the target domain and calculating the importance weights for re-weighting the source data in a single unified formulation is proposed. The weights are optimized by both a supervised classifier and distribution matching between the source domain and target domain with maximum mean discrepancy. Besides, a multiple domains active learning method is designed based on the proposed framework as an example. The proposed method is verified with newsgroups and handwritten digits data recognition tasks, where it outperforms the state-of-the-art methods.

Low-resource entity resolution with domain generalization and active learning

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.

In this chapter, we address the visual learning of automatic concept detectors from web video as available from services like YouTube. While allowing a much more efficient, flexible, and scalable concept learning compared to expert labels, web-based detectors perform poorly when applied to different domains (such as specific TV channels). We address this domain change problem using a novel approach, which – after an initial training on web content – performs a highly efficient online adaptation on the target domain. In quantitative experiments on data from YouTube and from the TRECVID campaign, we first validate that domain change appears to be the key problem for web-based concept learning, with a much more significant impact than other phenomena like label noise. Second, the proposed adaptation approach is shown to improve the accuracy of web-based detectors significantly, even over SVMs trained on the target domain. Finally, we extend our approach with active learning such that adaptation can be interleaved with manual annotation for an efficient exploration of novel domains.

Deep learning has recently attracted significant attention in the field of hyperspectral images (HSIs) classification. However, the construction of an efficient deep neural network (DNN) mostly relies on a large number of labeled samples being available. To address this problem, this paper proposes a unified deep network, combined with active transfer learning that can be well-trained for HSIs classification using only minimally labeled training data. More specifically, deep joint spectral-spatial feature is first extracted through hierarchical stacked sparse autoencoder (SSAE) networks. Active transfer learning is then exploited to transfer the pre-trained SSAE network and the limited training samples from the source domain to the target domain, where the SSAE network is subsequently fine-tuned using the limited labeled samples selected from both source and target domain by corresponding active learning strategies. The advantages of our proposed method are threefold: 1) the network can be effectively trained using only limited labeled samples with the help of novel active learning strategies; 2) the network is flexible and scalable enough to function across various transfer situations, including cross-dataset and intra-image; 3) the learned deep joint spectral-spatial feature representation is more generic and robust than many joint spectral-spatial feature representation. Extensive comparative evaluations demonstrate that our proposed method significantly outperforms many state-of-the-art approaches, including both traditional and deep network-based methods, on three popular datasets.

The perceptive systems used in automated driving need to function accurately and reliably in a variety of traffic environments. These systems generally perform object detection to identify the positions and attributes of potential obstacles. Among the methods which have been proposed, object detection using three-dimensional (3D) point cloud data obtained using LiDAR has attracted much attention. However, when attempting to create a detection model, annotation must be performed on a huge amount of data. Furthermore, the accuracy of 3D object detection models is dependent on the data domains used for training, such as geographic or traffic environments, so it is necessary to train models for each domain, which requires large amounts of training data for each domain. Therefore, the objective of this study is to develop a 3D object detector for new domains, even when trained with relatively small amounts of annotated data from new domains. We propose using a model that has been trained with a large amount of labeled data for pre-trained model, and simultaneously using transfer learning with limited amount of highly effective training data, selected from the target domain by active learning. Experimental evaluations show that 3D object detection models created using the proposed method perform well at a new location. We also confirm that active learning is particularly effective only limited training data available.Keywords3D object detectionTransfer learningActive learning

Where am I in the dark: Exploring active transfer learning on the use of indoor localization based on thermal imaging

Active multi-kernel domain adaptation for hyperspectral image classification

While the current supernova (SN) photometric classification system is based on high resolution spectroscopic observations, the next generation of large scale surveys will be based on photometric light curves of supernovae gathered at an unprecedented rate. Developing an efficient method for SN photometric classification is critical to cope with the rapid growth of data volumes in current astronomical surveys. In this work, we present an adaptive mechanism that generates a predictive model to identify a particular class of SN known as Type Ia, when the source set is made of spectroscopic data, while the target set is made of photometric data. The method is applied to simulated data sets derived from the Supernova Photometric Classification Challenge, and preprocessed using Gaussian Process Regression for all objects with at least 1 observational epoch before -3 and after +24 days since the SN maximum brightness. The main difficulty lies in the compatibility of models between spectroscopic (source) data and photometric (target) data, since the underlying distributions on both, source and target domains, are expected to be significantly different. A solution is to adapt predictive models across domains. Our methodology exploits machine learning techniques by combining two concepts: 1) domain adaptation is used to transfer properties from the source domain to the target domain; and 2) active learning is used as a means to rely on a set of confident labels on the target domain. We show how a combination of both concepts leads to high generalization (i.e., predictive) performance.

We tackled the Evalita 2011 Domain Adaptation task with a strategy of active learning. The DeSR parser can be configured to provide different measures of perplexity in its own ability to parse sentences correctly. After parsing sentences in the target domain, a small number of the sentences with the highest perplexity were selected, revised manually and added to the training corpus in order to build a new parser model incorporating some knowledge from the target domain. The process was repeated a few times for building a new training resource partially adapted to the target domain. Using the new resource we trained three stacked parsers, and their combination was used to produce the final results.KeywordsDependency parserdomain adaptationactive learningstacked parserparser combination

Open set transfer learning through distribution driven active learning

How to discover potential users of new released products is an important task in marketing campaigns, especially for companies have many established users. Indeed, transfer learning is a natural choice for addressing such problem, which can leverage the user data from historical products (i.e., auxiliary domains) for recommendation of new product (i.e., target domain). However, although most of the existing solutions of transfer learning work well on predicting user ratings, few of them can be directly exploited for the top-k recommendation with respect to user preferences. To this end, in this paper, we propose a novel transfer learning approach based on metric learning and active learning to address the problem of top-k potential user recommendation. Specifically, by modeling the user behaviors on historical products, we first propose to learn a global distance metric for capturing the commonality of users and segmenting users into different neighborhoods. Furthermore, to enhance the local discriminability among different neighborhoods, we select one representative user for each neighborhood by a new type of active learning and also learn local distance metrics for better estimating the user preferences. With such representative selection and metric learning, a novel method is developed for recommending top-k potential users of new products. Finally, we perform extensive experiments on real-world datasets, and the results validate the effectiveness of the proposed methods.