The list does not describe why they are the best books, except for a very short blurb. We read the Deep Learning book by Goodfellow, Bengio, and Courville in our reading group when it came out. Even though it contains useful information, it is written in a very haphazard fashion. It is also very unclear what its target audience is. Some sections start as a foundational description, to suddenly change into something that is only for readers with a strong maths background. No one in the reading group was enthusiastic about the book and most actively recommend against it (some called it 'the deep learning book for people who already know deep learning').
Put differently, a list such as the linked one may attract a lot of visitors. But without critical, in-depth reviews it is not very useful and might set potential learners on the wrong path.
It's a great book if you need references on the basic deep learning stuff for publications or your thesis. However, for getting started it is horrible.
What books did your group find well written?
It would be very helpful for outsiders to know what people who know their stuff consider good learning material.
Deep Learning with Python by Francois Chollet is very accessible as an introductory text. It uses plain language, avoids heavy math, and provides hands-on experience for the reader.
In general, I have found Manning to be one of the best technical publishers in terms of quality of content and updates.
For people who are a little afraid of the word deep learning:
It is not very difficult, in fact it is a borderline easy concept. So, do not be afraid in case you are. The mathematics behind deep learning is a little complicated but not much.
But if you just want to have a working knowledge then its quite approachable even for beginners and you do not need to learn the mathematics at all. In fact even the professionals rarely worry about the mathematics (they should though).
In the fast.ai course, there are two or three amazing videos where Jeremy Howard explains the basic concepts of deep learning (convolutions in this case) using Excel spreadsheets, highly recommended.
I agree with mkagenius, I've played around a lot with deep learning and like physics it suffers a lot from unnecessary jargon that hides relatively simple concepts, for example this random quote:
>ReLU stands for rectified linear unit, and is a type of activation function. Mathematically, it is defined as y = max(0, x).
Now everybody could just use plain English - if you have a negative number, set it to zero, in all other cases, keep the number - I don't know why someone has to use 'rectified linear unit' to describe this simple operation.
> Now everybody could just use plain English - if you have a negative number, set it to zero, in all other cases, keep the number - I don't know why someone has to use 'rectified linear unit' to describe this simple operation.
I agree this is often a problem, but "rectified linear unit" isn't a deep learning term that was made up, as you obviously know from your physics background - its origins are in AC/DC power conversion. Anyone who knows basic signal processing knows what a rectifier is, and something that employs that function is a unit of it.
There's plenty of nonsensical complexity in machine learning, but rectifiers and sigmoids and other 10-cent words describing maths aren't overly complicated, in my opinion.
Lastly, saying "if you have a negative number, set it to zero, in all other cases, keep the number" is a lot more complicated and tedious than just saying "apply a rectifier." Would "ramp function" be acceptably plain?
> I don't know why someone has to use 'rectified linear unit' to describe this simple operation.
To quote some snarky comedian whose name escapes me, because words have meanings. The way I understand it, "rectified" here has an analogous in physics. Beyond saving characters, it also tells that the purpose of this function is similar to the one people use when dealing with current. If I were either a physicist or someone used to deal with wave functions, the title alone would tell me everything I need to know. And I bet the type of people who care about activation functions are the type of people who knows what rectifiers are.
I would argue that learning NNs includes learning the jargon needed to speak about them, the same way learning the meaning of "pointer" is required for programming.
On the other hand, maybe I'm just rationalizing the time I spent learning jargon.
Intuitive descriptions using words have a place, and I do wish there were more of that. However, if you favor words over math notation, at some point you still have to call it something. What would you call it?
For anyone who knows intermediate python and numpy and wants to learn how neural networks (CNNs, RNNs) work through simple pythonic examples, Chollet’s book is excellent. I’ve bought two copies as I can’t bear lending mine out to my students. He’s the creator of the Keras library and it’s beautiful in its symplicity.
The highest-rated Amazon reviews seem to have come to the same conclusion: https://www.amazon.com/Deep-Learning-Adaptive-Computation-Ma...
Put differently, a list such as the linked one may attract a lot of visitors. But without critical, in-depth reviews it is not very useful and might set potential learners on the wrong path.