The Author Online Book Forums are Moving

The Author Online Book Forums will soon redirect to Manning's liveBook and liveVideo. All book forum content will migrate to liveBook's discussion forum and all video forum content will migrate to liveVideo. Log in to liveBook or liveVideo with your Manning credentials to join the discussion!

Thank you for your engagement in the AoF over the years! We look forward to offering you a more enhanced forum experience.

447276 (13) [Avatar] Offline
#1
I'll update this thread with some of the issues I found (or what seemed like an issue to me) as I go on reading the book.

  • In the CNN chapter, the term 'kernel' is mentioned. But its never explained what that means. To a first timer who is not familiar with image processing, it might be better to explain what it means first...or maybe just say its fancy term for weights at the least?.

  • Similarly, the codes are just there -- there's no explanation on what anything does. It should have comments going along with it. For example, its not explained why random weight initialization is being multiplied and subtracted by the terms until way later in the book.

  • The term soft-max is mentioned but not explained..somewhat introduced in later pages..but I think you should be explaining the term when you mention it or maybe not mention it at all until you get around to explaining what it is? It just adds to the confusion.

  • The part explaining why we are multiplying with derivative of relu or other activation functions is not that clear (to me, from beginner's perspective).

  • The codes included in the github repo or that can be downloaded does not give the same accuracy as shown in the book. Using tensorflow engine, I am only getting like ~30% accuracy. Update: The accuracy being different is apparently because of Python 2.x's float division instead of true division. Solved the issue by importing from future (from __future__ import division) Maybe it should be mentioned on the book?

  • There's a mistake for the one hot encoded word vector image on page 189. It should say "1101", not "1010" (according to the image on page 187).

  • Page 199 mentions "cross entropy", but it is never explained what that term means. It goes back to what I said above, it just confuses readers.

  • MNIST images are divided by 255 but it is never explained why it is done like that. To a first timer (as the book is meant for beginners), you wouldn't expect them to know why 255 is the magic number. Perhaps it should say that MNIST consists of only black and white images and the max value any pixel can have is 255 that is why we are dividing by 255 to normalize. Also, it would be better if there were some paragraphs about how to handle if the computation has to be done on a colored image (repeat for each channel) or maybe an example of the code. Some intuition about image processing in general or what values pixels can hold or how images are made up might be helpful. Just a suggestion.

  • Anyhow, my major issue with what i've read so far is not having comments on the code. If there could be comments alongside the code explaining what was being done (take for example, as it is done in the Machine Learning in Action book), it would be really helpful.
    447276 (13) [Avatar] Offline
    #2
    Btw, the accuracy being different is apparently because of Python 2.x's float division instead of true division. Solved the issue by importing from future.

    from __future__ import division


    Maybe it should be mentioned on the book? Thanks!