An Introduction to Support Vector Machines and other kernel-based learning methods
Nello Cristianini, John Shawer-Taylor
Tom M. Mitchell
曾华军 张银奎 等译
 An introduction to Kernel-Based Learning Algorithms
Klaus-Robert Muller, Sebastian Mika, Gunnar Ratsch, Koji Tsuda, and Bernhard Scholkopf
1.Some background of computational learning theory
Empirical Risk Minimization (ERM)
Probably Approximately Correct Learning (PAC)
--risk function and expected risk ( p131)
 Chapter4 Generalization Theory
 p161 section4.8
 chapter seven
*They tend to solve the problem:
Roughly speaking, the VC dimension measures how many (training) points can be shattered (i.e., separated) for all possible labelings using functions of the class. ( section II)
 chapter 5
 chapter 1
The problems addressed in 1 have a similar form: the hypothesis function should be chosen to minimize (or maximize) a certain functional. Optimization theory is the branch of mathematics concerned with characterizing the solutions of classes of such problems, and developing effective algorithms for finding them.
*Frequently the target concept cannot be expressed as a simple linear combination of the given attributes, but in general requires that more abstract features of the data be exploited. Kernel representations offer alternative solution by projecting the data into a high dimensional feature space to increase the computational power of the linear learning machines. p26
[comment:] "the advantage of using the machines in the dual representation derives from the fact that in this representation the number of tunable parameters does not depend on the number of attributes being used."
-- The number of parameters depends on the definition of the kernel. if you define a "bad" kernel, there will be many parameters as well.
With kernels we can compute the linear hyperplane classifier in a Hilbert space with higher dimension. We prefer linear models because:
1) there is the intuition that a "simple" (e.g., linear) function that explains most of the data is preferable to a complex one (Occam's razor). ( section II)
2) In practice the bound on the general expected error computed by VC dimension is often neither easily computable nor very helpful. Typical problems are that the upper bound on the expected test error might be trivial (i.e., larger than one), the VC dimension of the function class is unknown or it is infinite (in which case one would need an infinite amount of training data). (, section II )
For linear methods, the VC dimension can be estimated as the number of free parameters, but for most other methods (including k-nearest neighbors), accurate estimates for VC-dimension are not available.
3) for the class of hyperplanes the VC dimension itself can be bounded in terms of an other quantity, ther margin. (the margin is defined as trhe minimal distance of a sample to the decision surface) ( section II )
4. Existing Kernels
- Polynomial (homogeneous):
- Polynomial (inhomogeneous):
- Radial Basis Function: , for γ > 0
- Gaussian Radial basis function:
- Sigmoid: , for some (not every) κ > 0 and c <>
In short, not the mimensionalitybut the complexity of the function class matters. , all one needs for separation is a linear hyperplane. However, it becomes rather tricky to control the feature space for large real-world problems. So even i one could control the statistical complexity of this function class, one would still run into intractability problems while executing an algorithm in this space. Fortunately, for certain feature spaces F and corresponding mapping phai, there is a highly effective trick for computing scalar products in feature spaces using kernel functions. ( section II)