multiple kernel learning presentation 思路

Overall introduction of the topic, the keywords, the references.

Part 1： Introduction on svm(5 min)

　　first show that svm is what kind of a classifier. via graphs and examples. The notion of support vectors . the use of support vectors.

　　svm characteristic:largest margin,geometric margin and function margin, induction,finally reduced to a QP problem.

 

　　svm applications

Part 2: kernel method and kernel matrix (5 min)

　　why kernel method

　　what's kernel method for 

　　the notion of kernel functions.

　　the notion of similarity function

　　inner product

　　example

Part 3: the SMO algorithm

 

Part 4:  muliple kernel learning(20 min)

　　why MLK-------the need for flexibility to combine different kernel matrix linearly.

　　what MLK yields and where the problem lies

　　the formulation of MLK problem

　　the notion of SKM and the induction of it.   Primal problem --> Dual problem-->KKT-->some conclusions 

　　kernelization-->the kernelized problem formulation

　　Equavalence of the two formulation

　　why we can't solve it

　　regularize SKM: the dual problem -->solving the MY-regularized problem using SMO

Conclusion

　　propose a novel dual formulation to apply SMO to MY-regularized non-smooth convex problem.

　　and the author's simulation show that the SMO will be faster than the normal Mosek