Conventional zero-shot learning (ZSL) methods generally learn an embedding, e.g., visual-semantic mapping, to handle the unseen visual samples via indirect manner. In this paper, we take advantage of generative adversarial networks (GANs) and propose a novel method, named leveraging invariant side GAN (LisGAN), which can directly generate features from random noises are conditioned by semantic descriptions. Specifically, train conditional Wasserstein GANs in generator synthesizes fake...

Currently, unsupervised heterogeneous domain adaptation in a generalized setting, which is the most common scenario real-world applications, under insufficient exploration. Existing approaches either are limited to special cases or require labeled target samples for training. This paper aims overcome these limitations by proposing framework, named as transfer independently together (TIT). Specifically, we learn multiple transformations, one each (independently), map data onto shared latent...

Unsupervised domain adaptation addresses the problem of transferring knowledge from a well-labeled source to an unlabeled target where two domains have distinctive data distributions. Thus, essence is mitigate distribution divergence between domains. The state-of-the-art methods practice this very idea by either conducting adversarial training or minimizing metric which defines gaps. In paper, we propose new method named Adversarial Tight Match (ATM) enjoys benefits both and learning....

Domain adaptation aims to leverage knowledge from a well-labeled source domain poorly labeled target domain. A majority of existing works transfer the at either feature level or sample level. Recent studies reveal that both paradigms are essentially important, and optimizing one them can reinforce other. Inspired by this, we propose novel approach jointly exploit with distribution matching landmark selection. During transfer, also take local consistency between samples into consideration so...

In real-world transfer learning tasks, especially in cross-modal applications, the source domain and target often have different features distributions, which are well known as heterogeneous adaptation (HDA) problem. Yet, existing HDA methods focus on either alleviating feature discrepancy or mitigating distribution divergence due to challenges of HDA. fact, optimizing one them can reinforce other. this paper, we propose a novel method that optimize both unified objective function....

Unsupervised Domain Adaptation (UDA) aims to generalize the knowledge learned from a well-labeled source domain an unlabled target domain. Recently, adversarial adaptation with two distinct classifiers (biclassifier) has been introduced into UDA which is effective align distributions between different domains. Previous bi-classifier learning methods only focus on similarity outputs of classifiers. However, cannot guarantee accuracy samples, i.e., traget samples may match wrong categories...

Zero-shot learning (ZSL) and cold-start recommendation (CSR) are two challenging problems in computer vision recommender system, respectively. In general, they independently investigated different communities. This paper, however, reveals that ZSL CSR extensions of the same intension. Both them, for instance, attempt to predict unseen classes involve spaces, one direct feature representation other supplementary description. Yet there is no existing approach which addresses from perspective....

This paper focuses on the specific problem of multiview learning where samples have same feature set but different probability distributions, e.g., viewpoints or modalities. Since lying in distributions cannot be compared directly, this aims to learn a latent subspace shared by multiple views assuming that input are generated from subspace. Previous approaches usually common either maximizing empirical likelihood, preserving geometric structure. However, considering complementarity between...

Conventional machine learning algorithms suffer the problem that model trained on existing data fails to generalize well sampled from other distributions. To tackle this issue, unsupervised domain adaptation (UDA) transfers knowledge learned a well-labeled source different but related target where labeled is unavailable. The majority of UDA methods assume and are available complete during training. Thus, divergence between two domains can be formulated minimized. In paper, we consider more...

It is widely acknowledged that the success of deep learning built upon large-scale training data and tremendous computing power. However, power are not always available for many real-world applications. In this paper, we address machine problem where it lacks limits Specifically, investigate domain adaptation which able to transfer knowledge from one labeled source an unlabeled target domain, so do need much domain. At same time, consider situation running environment confined, e.g., in edge...

Conventional zero-shot learning (ZSL) methods generally learn an embedding, e.g., visual-semantic mapping, to handle the unseen visual samples via indirect manner. In this paper, we take advantage of generative adversarial networks (GANs) and propose a novel method, named leveraging invariant side GAN (LisGAN), which can directly generate features from random noises are conditioned by semantic descriptions. Specifically, train conditional Wasserstein GANs in generator synthesizes fake...

Domain adaptation investigates the problem of cross-domain knowledge transfer where labeled source domain and unlabeled target have distinctive data distributions. Recently, adversarial training been successfully applied to achieved state-of-the-art performance. However, there is still a fatal weakness existing in current models which raised from equilibrium challenge training. Specifically, although most methods are able confuse discriminator, they cannot guarantee that sufficiently...

The number of "hits" has been widely regarded as the lifeblood many web systems, e.g., e-commerce advertising systems and multimedia consumption systems. However, users would not hit an item if they cannot see it, or are interested in item. Recommender system plays a critical role discovering interesting items from near-infinite inventory exhibiting them to potential users. Yet, two issues crippling recommender One is "how handle new users", other surprise users". former well-known...

Lately, generative adversarial networks (GANs) have been successfully applied to zero-shot learning (ZSL) and achieved state-of-the-art performance. By synthesizing virtual unseen visual features, GAN-based methods convert the challenging ZSL task into a supervised problem. However, since real features are not available at training stage, train GAN generator on seen categories further apply it instances. An inevitable issue of such paradigm is that synthesized prone references incapable...

Zero-shot learning (ZSL) is a pretty intriguing topic in the computer vision community since it handles novel instances and unseen categories. In typical ZSL setting, there main visual space an auxiliary semantic space. Most existing methods handle problem by either visual-to-semantic mapping or semantic-to-visual mapping. other words, they investigate unilateral connection from one end to other. However, between are bilateral reality, that is, depicts space; space, on hand, describes this...

Domain adaptation is proposed to deal with the challenging problem where probability distribution of training source different from testing target. Recently, adversarial learning has become dominating technique for domain adaptation. Usually, methods simultaneously train a feature learner and discriminator learn domain-invariant features. Accordingly, how effectively domain-adversarial model features becomes challenge in community. To this end, we propose article novel scheme named entropy...

Researchers have proposed various machine learning algorithms for traffic sign recognition, which is a supervised multicategory classification problem with unbalanced class frequencies and appearances. We present novel graph embedding algorithm that strikes balance between local manifold structures global discriminative information. A structure designed to depict explicitly the of signs appearances intuitively model between-class Through this structure, our effectively learns compact...

The number of "hits" has been widely regarded as the lifeblood many web systems, e.g., e-commerce advertising systems and multimedia consumption systems. However, users would not hit an item if they cannot see it, or are interested in item. Recommender system plays a critical role discovering items from near-infinite inventory exhibiting them to potential users. Yet, two issues crippling recommender One is "how handle new users", other surprise users". former well-known cold-start...

Learning the relationship between multi-modal data, e.g., texts, images and videos, is a classic task in multimedia community. Cross-modal retrieval (CMR) typical example where query corresponding results are different modalities. Yet, majority of existing works investigate CMR with an ideal assumption that training samples every modality sufficient complete. In real-world applications, however, this does not always hold. Mismatch common datasets. There high chance some modalities either...

Generalized zero-shot learning (GZSL) aims to recognize samples whose categories may not have been seen at training. Recognizing unseen classes as ones or vice versa often leads poor performance in GZSL. Therefore, distinguishing and domains is naturally an effective yet challenging solution for In this paper, we present a novel method which leverages both visual semantic modalities distinguish categories. Specifically, our deploys two variational autoencoders generate latent representations...

In industry, accurate remaining useful life (RUL) prediction is critical in improving system reliability and reducing downtime accident risk. Numerous deep-learning approaches have been proposed achieved impressive performance RUL prediction. Nevertheless, most of them are based on an unrealistic assumption, that is, the training (source) testing (target) data follow similar distributions. real-world applications, source target domains usually different distributions, which degrades model...