I found an Error or Typo in the Book — what now?

Excellent – well… in any case, please let me know so that I can add it to the list of errata and fix it in a future edition of the book. There are two ways for doing this:

• Alternative 1: send me an e-mail with a brief description of the problem: mds-book ∂ proton.proton.me
  
• Alternative 2: Go to the book's GitHub repository and create an "issue" consisting of a short title (e.g., "typo in chapter 8") and a description (e.g., "equation 3 on page 33 should start with a capital X").
  
  
  Note that for submitting a GitHub issue you need to be logged in to your Github account.

List of Errata

Below are the most important corrections:

• page 48, in Definition 3.11 an index j should be an i. Correct sentence "The record number $i$ consists of the input data $x_i =$ ...". [found by GitHub user zym10031]
• page 103: John von Neumann died in 1957 (and not 1857). [thanks to GitHub user Meisenbart]
• page 451: Example 16.3 last line: "Therefore, it has slightly...:" should be removed
  [found by GitHub user IAM-DW]
• page 503-504 [spotted by GitHub user IAM-DW]:
  • $\partial J/\partial b_3$ ? from the bias of the last layer is missing in the text;
  • $b_3$ to be added to eq. 18.7, 18.14 ; in eq. 18.9 the bias $b_3$ is included.
  • equation 18.39: it should be $W^{(k)T}$ transposed and also in its later occurrence (eq. 18.47) $\partial J/\partial y_i^{(k-1)}=\partial J/\partial y_r^{(k)} w_{ir}^{(k)} $ in index notation
  • equation 18.51: the derivative of the activation function got lost. The equation should read: $\partial J/\partial \boldsymbol{y}^{(k-1)} = ... =\partial J/\partial \boldsymbol{y} ^{(k)} {\varphi}'(\boldsymbol{x}^{(k)}) $ where the derivative of the activation function is applied element-wise to all components of $\boldsymbol{x}^{(k)}$.
• page 507, Listing 18.1, line 28 and 29: It should be dJdb2 instead of dJdb1
• page 512:
  • the superscripts in eq. 18.30 should be (k+1) instead of (i+1)
  • eq. 18.31 should have no summation symbol. Corrected equation: $\mathcal{J} = \frac{1}{2}\left(y^\text{train} - y^\text{pred}\right)^2$
• page 518, In [2]: for consistency, weights/bias should be updated after the computation of dJdy_prev.