|Deep Learning (No Starch Press)|
Author: Andrew Glassner
Is a really good idea but, I think it only makes sense if you already know the equations. This is a very good book and don't let any of my comments put you off. It has clearly benefited from a great deal of work, care and attention but I don't think it will serve the complete beginner particularly well. For one thing it is just too long. If you are trying to learn something as complex as deep learning then you need a shorthand, i.e. math, to enable you to keep it all in your head. Without the math you have a lot of learning about individual cases to master and remember.
The first thing to say is that this book starts from the basics. You essentially get a statistics 101 course and lots of ideas from calculus, linear algebra and geometry. The geometry suits the style because everything is explained in color diagrams. The diagrams are mostly excellent and will help you understand the concepts. This first part is 150 pages long and it is a lot to take in if you really have never encountered the ideas before.
I'm a little distressed by the explanation of frequentist versus Bayesian statistics. The explanations are all good. but the idea that frequentists don't use Bayes-style reasoning when it is appropriate is wrong. Even the hardest-line frequentist would use Bayes theorm to work out conditional probabilities in situations where decisions were needed. The division between them is that frequentists don't think that probabilities exist unless they can be measured, if only in theory, by repeated experiements. Bayesians are generally happy with probability interpreted as a measure of belief. I'm glad to say that this is the only part of the book I take issue with - the rest is very mainstream.
Part II is all about basic machine learning and it goes into far too much detail on techniques, principal components, clustering, support vector machines, etc that aren't really essential to understanding deep learning, but they are good background. The sections on confusion matrices and what goes wrong with machine learning is very good.
Parts III and IV are all about the subject suggested by the title and you have to read and follow 300 pages to reach it. When you do this isn't a gentle introduction to the basics, it deals with advanced techniques. Amazingly, Part IV deals with convolutional networks, recurrent networks, reinforcement learning, attentions and adverserial networks. This is cutting-edge stuff and all explained without a program or an equation. Instead there are lots of pictures.
This is an amazing book. It is lavishly illustrated throughout with full color diagrams, charts and samples, but it is very big and very demanding book. I don't think you could create a more approachable introduction to machine learning and deep learning, but it is still 750 pages of difficult ideas. As I've already said, without math to reduce it to general principles it's going to be tough to keep in your head. What is slightly sad is that if the equations were in the book the illustrations would be an excellent way of understanding what they mean. I would strongly suggest that the author produces another version of this book complete with equations, it would be shorter but more valuable.
My guess is that the ideal reader for this book is someone who knows the math, but wants to understand its implications at a more intuitive level. If you are such a reader you will make rapid progress though the book and repeatedly have "ah ha" movements as you understand what you thought you already knew.
If you have any interest in machine learning then this is worth trying. You will either find it too long or simply brilliant.
|Last Updated ( Saturday, 02 July 2022 )|