Multiple Kernel Learning for Fusion in Social Multimedia Data - Essay Example

Multiple Kernel Learning for Fusion In social Multimedia Data Subject: Computers Name: School Name:  A set of machine knowledge approaches that use predefined set of kernels and acquire an optimal linear or nonlinear, that what we refer to as multiple kernel learning. We combine kernels as the part of the algorithm. The key properties of the multiple kernel learning include the functional form, learning method, training method, target fusion, computational complexity, target function and base learner. Varma M. and Ray D. (2007) Multiple kernel learning has being so useful in social media in so many ways which include combination of texts, video, images and interactive contents also has helped in sharing of these files it. Also assist in extraction and detecting of events based on information The state of the art comprises of the alignment and face extraction, feature fusion and feature extraction. In face detection we use part based model and in the face tracking we use supervised gradient descent. The output from the tracker is sent to the tracker, incase the tracker fail the detector is reinitialized. There some reasons that makes us use this multiple learning kernel which include: is a technically simpler way of way of merging cases, there are simultaneous classifier training and combination of features, Decent learning limits, in the same formulation one can use Different data formats, Non-linear learning According to Yang, Tian, Duan and Gao (2009), Ability to collect parameters from a pool of kernels which in turns reduces the bias since there was kernel selection. It also helps in the combination of data from different sources with different notions of similarity this leads to demanding of different kernels. Combinations of already established kernels for each individual data source we use multiple kernel algorithm. Some application of multiple kernel learning include: Biomedical data fusion, Event recognition in video and Object recognition in images Uses of multiple kernel learning Has helped in developing classic support vector machines which helped in improvement in quality prediction. According to calpain by the use of kernel learning we are able to highlight the role and importance of each feature Used when there are heterogeneous representations of data for the task at hand as explained at Dietterich (2000) The (SVM) is a classifier which uses classification of binaries founded on the philosophy of risk reduction This is how SVM works by maximizing boundary2 F(x) = (w, Φ(x)) +b By solving this quadratic optimization we can get the classifier By minimizing 1 from ½||w||1+c N∑i=1 ξi W = ∈ Rs , ξ ∈ RN+, b ∈ R Yi(w, Φ(xi)) +b)≥ 1-ξi X is the dimension input (D) and yi ∈ {-1, +1}, ξ = vector of slack, c is a predefined +ve trade off parameter c= +ve parameter, W represent weight coefficients and b = bias term. Salakhutdinov R (2012) The process which helps in effective retrieval performance on video and images is the multimodal approach. There are two ways of combining strategies, which include: late fusion and early fusion. Early fusion consists of combining the features before classification is done, which may include multi-kernel learning. Its results are always better. Late fusion consists of combining the. Late fusion is preferred where hard work is needed to be tackled and it’s still robust In Double Fusion we explore the features which include text, audio and visual features For audio feature we use, Automatic Speech Recognition (ASR) feature, while for textual we use Optical Character Recognition (OCR) in visual features. The five visual features are CSIFT SIFT, Mo SIFT, STIP and GIST. Which have predefined functions In order to achieve high fusion presentation in CBIR, the presentation of the image should be diversified and good also require their corresponding similarity measures. We use six representations which are in pairs of three levels which include: patch-based level, global based level, and region-based levels. Fusion Image and Text Data When dealing with multimedia data, we apply fusion for image and text through combination images and text features. We make use of the data stored in the database that consist both text and photos to be used in the proposed event of mining system by Zafarani Abbasi R. (2014) This mining system consists of properties like event photo classification, event keyword detection, and photo selection. In tweet it can be either classified as text only or image, this is determined by the text mining score. If is less than a threshold it’s an image only and if its more than then it’s a text In calculating mining accuracy we prefer this formula A=TP+TN/TP+TN+FP+FN A = accuracy, TP= true positive, TN = true negative, FP = false positive and FN = false negative How multiple kernel learning work to fuse different sources (use mathematics) I am using labeled data with a set of sources which may be related or have different data sets and it assumes that all the tasks are in a single domain. Here we are using matrices to integrate data, there’s a method called symmetric positive definite. Kernel matrices are used to scale heterogeneous data making it to disappear; the main objective in in kernel fusion method is to construct datasets of same representation. Methods like linear combination which not mostly used since it can only work with small number,’ Am are going to use symmetric positive definite since it can handle large number of data sets. There are two SPDs one is explicit formula and the second one is when there are many matrices, in this condition some natural and GM remain for long. Karcher mean is considered to be the most appropriate instance when using the matrix GM, A1 to Ak becomes the barycenter for the matrices in the SPD. This is acquired by finding the minimum of the optimization as shown below GM is mostly used because of its properties and the most useful is its invariance under inversion while AM lacks invariance under inverse. Let’s see how it goes: G (K1, K2) = K11/2(K1-1/2K2K1-1/2)1/2K11/2 All this brings about the fusion in kernel learning of K2, K1 this represents the FðK1, K2Þ ¼ GðK1,K2Þ. By use of this description it helps in computing the GM of Matrices related to SPD. The approximations are in the discussed below Minimize, G(A1,,,,,,,,,,,,AK) = X2P,𝑖=1-𝑘-.||log(A1-1/2xA1-1/2)||2F Where pn is the set of symmetric +ve definite and the ||,||f becomes the frobenius norm The main problem related to use kernel learning in data sources includes the cost of computation increases as the number of data sources increases. REFERENCES Srivastava N, Salakhutdinov R (2012) Multimodal learning with deep Boltzmann machines. In: NIPS Bach .F. R, Lanckriet .G. R. G. and Jordan .M. I., (2004) “Multiple kernel learning, conic duality, and the SMO algorithm,” in Proc. Int. Conf. Mach. Learn. Schölkopf .B and Smola .A. J. (2002) with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. Cambridge, MA: The MIT Press. Hanjalic .A and Xu .L.-Q, (2005) “Affective video content representation and modeling,” IEEE Trans. Multimedia, vol. 7. Cao. L, Luo .J, Liang .F, and Huang .T. S, (2009,) “Heterogeneous feature machines for visual recognition,” in Proc. IEEE Int. Conf. Comput. Vis., pp. 1095–1102 Seo H. J. and Milanfar P. (2011) “Action recognition from one example,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 5, pp. 867–882. Dietterich T. G. (2000) “Ensemble methods in machine learning,” in Proc. Workshop Multiple Classifers Syst. Varma M. and Ray D. (2007) “Learning the discriminative power-invariance tradeoff,” in Proc. IEEE Int. Conf. Comput. Vis., Dec, pp. 1–8 Yang.J, Y. Li, Tian. Y, Duan L and Gao W. (2009) “Group- Sensitive Multiple Kernel Learning for Object Categorization,” in ICCV. ZafaraniAbbasi R. (2014) Social Media Mining: Cambridge University Press, Read More