Machine learning needs to read books _ Learning materials

If you only want to read a book, then recommend Bishop's Prml, full name pattern recognition and Machine Learning. This book is a machine learning Bible, especially for the Bayesian method, the introduction is very perfect. The book is also a textbook for postgraduate courses in machine learning in universities, such as deep Learning's pedigree Hilton, who also chose the book at the University of Toronto's Faculty of Machine Studies (CSC2515). The electronic version of the book is available for direct downloading on Bishop personal pages. But in advance, this book if read for half a year, read three times can understand thoroughly, also belong to normal. The first time does not understand very normal, so need to persevere.

http://research.microsoft.com/en-us/um/people/cmbishop/prml/

The second book is ESL (the Elements of statistical Learning). Chinese translation is called statistical learning basis, in fact, this translation is not accurate, statistical learning is not statistics, called machine learning base more accurate. The book Mathematical Deduction, the theoretical system is very complete, combined with the later exercise with the R language of their own contact, for understanding the basic methods of machine learning is very helpful, such as: Logistic,ridge regression. The book can also be downloaded directly to the electronic version on the author's website.

http://statweb.stanford.edu/~tibs/ElemStatLearn/

With a theoretical basis, combined with a number of professors of the curriculum to study, the effect is better. The current popular Stanford University machine learning Open Class, in the NetEase open class even have Chinese subtitles out, is a very good introductory tutorial. But individuals prefer the Hilton machine learning program. Because you read the above two books, basic knowledge is OK, and then see Stanford Open Class or foundation. The Hilton curriculum is closer to current academic research hotspots, such as the introduction of neural network,deep belief nets, whose curriculum includes the earliest RBM implementation http://www.cs.toronto.edu/~ Hinton/matlabforsciencepaper.html. It's worth a visit. His old man's course address: http://www.cs.toronto.edu/~hinton/csc2515/lectures.html specially recommended to do one of the assignments:http:// Www.cs.toronto.edu/~hinton/csc2515/assignments.html

These three books have been brushed some, recommend Mlapp.

1. PRML and Mlapp a bit like, are listed ml various models, but PRML than mlapp more partial probability interpretation, some for probability and probability. Mlapp is more neutral, the content is newer, and the attachment material is sufficient (have code)
2. ESL content and "PRML and Mlapp" are different, the details are often not detailed, need to see the corresponding paper to understand. Not a very suitable introductory textbook.
3. In addition, there is a copy of the foundations of Machine Learning

Recommend Mlapp. The reasons are as follows:
(1) Mlapp The basic Theory of machine learning (probability, distribution, inference) in place, and the previous chapters devoted to this part of the Knowledge Supplement, review and training, very helpful to the understanding of the specific model, when the specific model needs to be derived from the formula is also valuable.
(2) The Mlapp model is complete, what kind of model has, not only has the most common machine learning model based on Bayesian statistics, but also includes some non probabilistic models or models with difficult probability interpretation. Sometimes these models are also useful, even if you don't need to be in the same group as you do, it is necessary to understand these models (the PRML of this part of the model is relatively scarce).
(3) for the model based on probability, there are a number of models Mlapp talk about enough, after all, is written after the book, there are some more in line with the trend of the times, and also more detailed, some methods if only to do not in-depth understanding of the words, do not look at the paper, directly to see the Mlapp is enough, which in itself is Some methods even prml, I think Mlapp speak better than PRML, such as kernel and sparse kernel. As someone said above, prml a bit for probability.