SVM: This algorithm should be the most important part of the machine learning course.
The first is the idea of SVM: To find the super plane, the two types of the nearest point in the distance between the more open the better (until they are equal).
Then there is the difference between the function interval and the definition of the geometric interval.
Then a super-planar model is established to see how it transforms the problem into a convex optimization problem.
The first difficulty of SVM: Lagrange duality. By the KKT condition, the binding constraints are on the boundary, and this will be used to explain the support vectors.
It is known from the KKT condition that in the SVM model, only points with a function interval of 1 are the support vectors.
The w,b is solved by duality. For newly emerging samples, it is only necessary and support vectors for the inner product to classify them.
SVM: Sometimes low-dimensional can not be a good classification of samples, high-dimensional resolution, the introduction of nuclear functions , the low-dimensional mapping to high-dimensional.
The linear classification method is used to solve the nonlinear problem in two steps, first using a transform to map the data of the original space to the new space, and then using the line-line classification learning method in the new space.
Learn the classification model from the training data.
If a kernel function is semi-positive, it is valid.
In order to solve the problem of outliers, penalties are introduced. The new model should not only make the interval as small as possible, but also make the number of the wrong classification points as few as possible.
Summary of machine learning algorithms (II.)