22 comments

[ 2.6 ms ] story [ 62.8 ms ] thread
Might help if you made the actual PDF more prominent. It took me a few minutes just to figure out where, you know, the book actually was. Change the README.md or move the child tex/pdf/etc. into into sub-dirs or something.
I still can't find the actual book. Too bad.
The name of the file is White_book.pdf. I found it by downloading the repo, and sorting by file size.
I just created a PR to include a link to the raw download link.
I love living in 2017 when stuff like this is just created out of the blue making my life just a little bit easier, and at almost no cost.
The cost is the impermanence of knowledge capital
What do you mean by that?
he/she means that knowing something isn't valuable anymore. maybe that's true but I prefer this to guilds and gatekeepers.

also that's not a cost but an externality.

externalities are inherently costs/benefits, the only thing that's different is who is affected.
Yeah, everyone's machine learning Ph.D.s are useless now because you spent an afternoon reading a 100 page PDF.
You would be surprised. Deep learning usurped a ton of knowledge about SVMs and such, and then TensorFlow/Keras made it easy to write and deploy good models with little knowledge.

Clearly its more complex than that, but this is true for a good number of use cases.

Warning: font is Palatino
Ok I know about Comic Sans and Papyrus but what's the deal with Palatino?
It's math characters are nearly illegible. The font itself is OK, but the math it causes is rarely used for a reason.
I am the author of the document. My aesthetics sense is pretty limited as you have observed. I like the mathpazo font, but if you don't share my weird tastes, you can just download the repo, comment the \usepackage[sc]{mathpazo} line in formatAndDefs.tex and you're good to go.
How should one utilize the knowledge in the book? When do derivations play a role in deep learning? I am thinking it helps make decisions on what layers will help or not, or explain why this architecture improves some baseline. I would love to hear anyone's thought on the matter!
It's cool to see that much dedication. It's useful when people take the time to summarize knowledge in a book to serve as reference.

But ... I have the feeling that the author, who is relatively new to the field (by his own admission), expanded a lot of formulas and made certain parts of the theory more complicated than it should be.

Look around page 60. There are formulas with 6 summation signs in front of them, with all kinds of little indices floating around. How about page 37 ?

In a way, the whole point about the chain rule (and software libraries that implement it) is that you can stay in "math world" to do the reasoning, and not think about the job of managing the computation.

Same idea with expression as much as possible in terms linear algebra primitives. Matrix multiplication is easier to understand when it's not broken apart into sums whose indices you have to track.

See author's note excerpted below; that was an explicit goal of the project.

> This work has no benefit nor added value to the deep learning topic on its own. It is just the reformulation of ideas of brighter researchers to fit a peculiar mindset: the one of preferring formulas with ten indices but where one knows precisely what one is manipulating rather than (in my opinion sometimes opaque) matrix formulations where the dimension of the objects are rarely if ever specified.

-- I think that having those things written out explicitly is of great help to those not fully comfortable with formal manipulations. It is particularly useful when implementing those operations in low-level code. I say this even though I personally find the Einstein notation [1] most convenient.

[1]: https://en.wikipedia.org/wiki/Einstein_notation

I had not caught that note from the author. Thanks for pointing it out.