[khuss]

>> it replaces very simple explanations of concepts with complicated paragraphs that I can't make sense of

It is good to see critical views but it will be even better if you could give concrete examples for statements like the above. Also, what other books do you recommend?

[j2kun]

> Also, what other books do you recommend?

This has been my frustration. I've been wanting to grok deep learning, but haven't found any source that can explain it in a way that doesn't overcomplicate simple things (which I can only tell it's doing when I already understand the topic). I also don't have the time or incentive to really dig in and do the math myself from scratch, since so much of it is wide open research directions.

I've also had this experience with much much simpler areas of statistics and statistical ML (cf. Markov chain monte carlo), so this is a sort of recurring theme for me. Considering how simple MCMC is now that I do understand it, it's difficult to dispel the nagging feeling that all the statistical ML literature is (likely unintentionally) obfuscated.

I could give concrete examples of excerpts and entire sections of the book that don't make sense to me, but I don't think it's all that productive because a lot of it boils down to organizational disagreements, cultural assumptions behind the math, and differences in priorities. Individually it seems like nitpicking, but they add up quickly to general muddled confusion. This is especially true when, for example, almost always the right answer to the question "Why does this particular technique work?" is, "We have no clue, but here is some anecdotal evidence and half-substantiated oversimplified theories." Instead these answers are passed as well-known fact.

[kevinskii]

Have you read the acclaimed "Elements of Statistical Learning"[1] or its lighter weight counterpart, "Intro to Statistical Learning"[2]?

I think both explain statistical ML clearly at just the right level of detail. But I'm not an advanced math person, so it's certainly possible that I've failed to notice any flaws in their approach.

[1] https://web.stanford.edu/~hastie/local.ftp/Springer/OLD/ESLI...

[2] http://www-bcf.usc.edu/~gareth/ISL/

[gambler]

I haven't looked too deeply at this particular book yet, but your overall sentiment is sounds very familiar.

Too many people are willing to accept (and defend!) sub-par explanations and overcome them through titanic mental grind. The scary part of it is that often this leaves you with a broken mental model that continues to require tons of effort in application. Meanwhile, much better explanations exist. They make learning easy and fun, while also giving you intuition that is easy to apply and even expand to other areas.

I am pretty good at spotting bad mental models, but it's really hard to prove that they are bad, until you find one clearly superior. Recently I stumbled upon MIT's linear algebra class at Open Courseware and was blown away by how much easier it was to follow than the stuff I had in college, while covering the same material in much greater detail.

[mattmcknight]

The new O'Reilly book "Fundamentals of Deep Learning" by Nikhil Buduma (available on Safari for a while now) is good at the fundamentals- very clearly explained, nice diagrams. It it relatively close to the path of my Neural Networks classes (although those were 20 years ago).

It necessarily shorter on detail in terms of the tricks of implementation that have radically improved performance of these techniques over the past 5 years, and might serve as a good read before diving into this via Goodfellow, Bengio, and Courville.

[j2kun]

I haven't seen this one before. Looks like it's still in development.

[mattmcknight]

Yes, only the first three chapters are released.

[bronxbomber92]

It is valuable to cite examples because it will help others determine if they have the same organizational disagreements and cultural assumptions that would make the book burdensome to read.

On the other hand, if we find the cited examples clear due to e.g. different backgrounds, expectations, perspectives, etc, then that might be a good signal that we should look into the book.

[slagfart]

You know, if you wanted to write a book, you could set up a Patreon with a per-chapter benefit. Perhaps write the introduction for free, setting out your goals for the book. I'd chip in.

Even if the material is already 'covered' elsewhere, simply re-explaining the same concepts would do a world of good for those who think in the same way as you.