The goal of domain adaptation (DA) is to train a good model for target domain, with large amount labeled data in source but only limited the domain. Conventional closed set (CSDA) assumes and label spaces are same. However, this not quite practical real-world applications. In work, we study problem open (OSDA), which requires space partially overlap space. Consequently, solution OSDA unknown classes detection separation, normally achieved by introducing threshold prediction classes; however,...

In the cross-domain image classification scenario, domain adaption aims to address challenge of transferring knowledge obtained from source target that is regarded as similar but different domain. To get more reliable invariant representations, recent methods start consider class-level distribution alignment across and domains by adaptively assigning pseudo labels. However, these approaches are vulnerable error accumulation hence unable preserve category consistency. Because accuracy labels...

Ontology learning is a critical task in industry, dealing with identifying and extracting concepts captured text data such that these can be used different tasks, e.g. information retrieval. non-trivial due to several reasons limited amount of prior research work automatically learns domain specific ontology from data. In our work, we propose two-stage classification system learn an unstructured We first collect candidate concepts, which are classified into irrelevant collocates by...

In model serving, having one fixed during the entire often life-long inference process is usually detrimental to performance, as data distribution evolves over time, resulting in lack of reliability trained on historical data. It important detect changes and retrain time. The existing methods generally have three weaknesses: 1) using only classification error rate signal, 2) assuming ground truth labels are immediately available after features from samples received 3) unable decide what use...

Ontology learning is a critical task in industry, dealing with identifying and extracting concepts captured text data such that these can be used different tasks, e.g. information retrieval. non-trivial due to several reasons limited amount of prior research work automatically learns domain specific ontology from data. In our work, we propose two-stage classification system learn an unstructured We first collect candidate concepts, which are classified into irrelevant collocates by...

Common domain adaptation techniques assume that the source and target share an identical label space, which is problematic since when samples are unlabeled we have no knowledge on whether two domains same space. When this not case, existing methods fail to perform well because additional unknown classes also matched with during adaptation. In paper, tackle open set problem under assumption spaces only partially overlap, task becomes exist, how detect avoid aligning them domain. We propose...

Abstract Aiming at the problem that definition of crowd abnormal behavior detection is ambiguous and difficult to combine with context semantics, an algorithm using OCC human emotion model combined entropy proposed. First calculate for crowd, determine whether value abnormal, if it further extract optical flow OF HOG. Then project into two-dimensional vector data, send CNN local feature extraction achieve description emotions. Finally, predict abnormality occurs according judgment factor....

We study the problem of visual question answering (VQA) in images by exploiting supervised domain adaptation, where there is a large amount labeled data source but only limited target with goal to train good model. A straightforward solution fine-tune pre-trained model using those data, it usually cannot work well due considerable difference between distributions and domains. Moreover, availability multiple modalities (i.e., images, questions answers) VQA poses further challenges...

At present, the detection of crowd abnormal behavior has problem unclear definition anomalies and mutual occlusion. Existing researches mostly detect appearance characteristics individuals, extract analyze movement groups. Anomaly is achieved through comparison parameters thresholds. However, context semantics cannot be effectively utilized consistent with actual situation. There a disconnection between description. In this paper, we use designed convolutional neural network to...

Practical natural language processing (NLP) tasks often exhibit long-tailed distributions accompanied by noisy labels, posing significant challenges to the generalization and robustness of complex models such as Deep Neural Networks (DNNs). Traditional resampling techniques like oversampling or undersampling, while commonly employed, can easily lead overfitting. A growing trend involves leveraging small amounts metadata learn data weights, alongside demonstrated benefits self-supervised...

In model serving, having one fixed during the entire often life-long inference process is usually detrimental to performance, as data distribution evolves over time, resulting in lack of reliability trained on historical data. It important detect changes and retrain time. The existing methods generally have three weaknesses: 1) using only classification error rate signal, 2) assuming ground truth labels are immediately available after features from samples received 3) unable decide what use...