Many SVM formulations uses some or other smoothing function for example as in (https://pdfs.semanticscholar.org/fb7e/5403c219b9a49135c21a4580608f8ef1520f.pdf)
How does it impact computational time in solving the problem?
And accuracy of the problem in solving the problem?
The intuition that I have about this is that generalized derivatives cannot be further taken derivative. Because of this may be some of the optimization problem solvers may not work. When we do smoothing the functions may become more closer to convex but in case of non-smooth versions the the condition of convergence should be like (if 2 consecutive solutions shouldn't be better than some epsilon then the iterations should be stopped.) but in case of smoothed version the solution achieved must be quick.
Dear Mohamed Azab , I am not interested in another smoothing function instead my interest lies in benefits of using smoothed version over non-smooth version.
- Badges
- Science topic