mml-book / mml-book.github.io

Companion webpage to the book "Mathematics For Machine Learning"
12.91k stars 2.39k forks source link

Optimal Parameters in SVM #721

Open PBrdng opened 2 years ago

PBrdng commented 2 years ago

Describe the mistake I think I found an error in the SVM section: you write in the remark after Eq. (12.42) that the optimal value for the off-set for the Soft-Margin SVM is given by the median of the numbers

$$ | y_n - \langle w^*, x_n\rangle| $$

for all $n$ such that $\alpha_n\neq 0$. Instead, I think it should be the median of $(y_n - \langle w^*, x_n\rangle)$ (without absolute values).

Location

  1. ©2021 M. P. Deisenroth, A. A. Faisal, C. S. Ong. Published by Cambridge University Press (2020).
  2. 12
  3. 385
  4. Remark just after Equation (12.42)

Proposed solution As explained above, I think one has to remove the absolute values.

Additional context I typed a proof in Prop. 3.30 in https://pbrdng.github.io/MathData-SoSe22/MathData.pdf. Feel free to copy it.