Domain adaptation allows knowledge from a source domain to be transferred different but related target domain. Intuitively, discovering good feature representation across domains is crucial. In this paper, we first propose find such through new learning method, transfer component analysis (TCA), for adaptation. TCA tries learn some components in reproducing kernel Hilbert space using maximum mean miscrepancy. the subspace spanned by these components, data properties are preserved and...

Standard SVM training has O(m3) time and O(m2) space complexities, where m is the set size. It thus computationally infeasible on very large data sets. By observing that practical implementations only approximate optimal solution by an iterative strategy, we scale up kernel methods exploiting such approximateness in this paper. We first show many can be equivalently formulated as minimum enclosing ball (MEB) problems computational geometry. Then, adopting efficient MEB algorithm, obtain...

Machine learning has been highly successful in data-intensive applications but is often hampered when the data set small. Recently, Few-Shot Learning (FSL) proposed to tackle this problem. Using prior knowledge, FSL can rapidly generalize new tasks containing only a few samples with supervised information. In paper, we conduct thorough survey fully understand FSL. Starting from formal definition of FSL, distinguish several relevant machine problems. We then point out that core issue...

In this paper, we address the problem of finding pre-image a feature vector in space induced by kernel. This is central importance some kernel applications, such as on using principal component analysis (PCA) for image denoising. Unlike traditional method which relies nonlinear optimization, our proposed directly finds location based distance constraints space. It noniterative, involves only linear algebra and does not suffer from numerical instability or local minimum problems. Evaluations...

It is well-known that exploiting label correlations important to multi-label learning. Existing approaches either assume the are global and shared by all instances; or local only a data subset. In fact, in real-world applications, both cases may occur some globally applicable group of instances. Moreover, it also usual case partial labels observed, which makes exploitation much more difficult. That is, hard estimate when many absent. this paper, we propose new approach GLOCAL dealing with...

Low-rank matrix approximation is an effective tool in alleviating the memory and computational burdens of kernel methods sampling, as mainstream such algorithms, has drawn considerable attention both theory practice. This paper presents detailed studies on Nyström sampling scheme particular, error analysis that directly relates quality with encoding powers landmark points summarizing data. The resultant bound suggests a simple efficient scheme, k-means clustering algorithm, for low-rank...

Principal component analysis (PCA) has been proven to be an efficient method in pattern recognition and image analysis. Recently, PCA extensively employed for face-recognition algorithms, such as eigenface fisherface. The encouraging results have reported discussed the literature. Many PCA-based systems also developed last decade. However, existing are hard scale up because of computational cost memory-requirement burden. To overcome this limitation, incremental approach is usually adopted....

Kernel (or similarity) matrix plays a key role in many machine learning algorithms such as kernel methods, manifold learning, and dimension reduction. However, the cost of storing manipulating complete makes it infeasible for large problems. The Nyström method is popular sampling-based low-rank approximation scheme reducing computational burdens handling matrices. In this paper, we analyze how approximating quality depends on choice landmark points, particular encoding powers points...

The core vector machine (CVM) is a recent approach for scaling up kernel methods based on the notion of minimum enclosing ball (MEB). Though conceptually simple, an efficient implementation still requires sophisticated numerical solver. In this paper, we introduce (EB) problem where ball's radius fixed and thus does not have to be minimized. We develop (1 + e)-approximation algorithms that are simple implement do require any For Gaussian in particular, suitable choice (fixed) easy determine,...

Many real world learning problems can be recast as multi-task which utilize correlations among different tasks to obtain better generalization performance than each task individually. The feature selection problem in setting has many applications fields of computer vision, text classification and bio-informatics. Generally, it realized by solving a L-1-infinity regularized optimization problem. And the solution automatically yields joint sparsity tasks. However, due nonsmooth nature norm,...

Kernel methods, such as the support vector machine (SVM), are often formulated quadratic programming (QP) problems. However, given <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m$</tex> training patterns, a naive implementation of QP solver takes xmlns:xlink="http://www.w3.org/1999/xlink">$O(m^3)$</tex> time and at least xmlns:xlink="http://www.w3.org/1999/xlink">$O(m^2)$</tex> space. Hence, scaling up these QPs is major stumbling block in...

Previous chapter Next Full AccessProceedings Proceedings of the 2009 SIAM International Conference on Data Mining (SDM)Multiple Kernel ClusteringBin Zhao, James T. Kwok, and Changshui ZhangBin Zhangpp.638 - 649Chapter DOI:https://doi.org/10.1137/1.9781611972795.55PDFBibTexSections ToolsAdd to favoritesExport CitationTrack CitationsEmail SectionsAboutAbstract Maximum margin clustering (MMC) has recently attracted considerable interests in both data mining machine learning communities. It...

Deep neural network models, though very powerful and highly successful, are computationally expensive in terms of space time. Recently, there have been a number attempts on binarizing the weights activations. This greatly reduces size, replaces underlying multiplications to additions or even XNOR bit operations. However, existing binarization schemes based simple matrix approximation ignore effect loss. In this paper, we propose proximal Newton algorithm with diagonal Hessian that directly...