Hacker News new | past | comments | ask | show | jobs | submit login
Foundations of ML and AI: Book recommendations (dragan.rocks)
134 points by tosh 4 months ago | hide | past | web | favorite | 17 comments



Original title: Programmer, Teach Yourself Foundations of ML and AI with these 6 Books

Books:

- Ian Hacking - Introduction to probability and Inductive Logic

- John Kruschke - Doing Bayesian Data Analysis

- Gareth Williams - Linear Algebra with Applications, Alternate Edition

- Matthew Scarpino - OpenCL in Action

- Michael Nielsen - Neural Networks and Deep Learning

- Goodfellow, Bengio, and Courville - Deep Learning


How to teach yourself ML in two easy steps

1) Have a strong knowledge of undergraduate mathematics, probability, statistics, numerics

2) read a book about machine learning


Piggybacking on this, I instead recommend an introductory ML book like Bishop or Murphy, a statistical ML book like Mohri or Shai Shalev-Schwartz, and a textbook on nonlinear optimization, convex or otherwise.

The jump from classical machine learning to deep learning is not far if you have a good understanding of first principles.


Any reason to pick Mohri over Shai Shalev-Schwartz, or the other way around?


I think that SSS is a little easier to follow, while Mohri is more complete.

It was helpful to me to have both, as details skipped over in one proof were often highlighted or better explained in its corresponding description. For someone whose training was not theoretical computer science, SSS left me with a better understanding most of the time.


Great, I'll start with SSS. Any tips for a good optimization book to read afterwards?


I may be all wrong about this (and welcome any dissenting opinions) but it seems to me that:

(1) Bayesian data analysis isn't actually necessary for most AI/ML work. Bayes theorem is fundamental of course, but Bayesian data analysis (BDA) seems to rarely come up in practice. ML algorithms like Naive Bayes don't really require knowledge of BDA, plus PGMs aren't really that common (and not within BDA's scope anyway).

The computational methods behind BDA (mostly MCMC-based) are also fairly heavy and I don't know too many ML shops that actually use MCMC-based Bayesian ML.

To me, BDA's primary value is to "upgrade" the type of NHST-based data analysis done in science and social science, and ground it on what I think is more solid epistemology. I don't know if BDA is really that practical for machine learning, where the goal is not analysis but prediction.

(2) Kruschke might not be the best book for learning BDA. I own a copy of Krushke (puppies on cover and all), and found the first few chapters interesting, but it then quickly got tedious. It seemed to me that Kruschke, in an effort to make things accessible to social scientists, belabors the subject a little in later chapters without adding pedagogical value (I realize this is controversial statement to Kruschke fans, and I am prepared to change my mind).

Gelman's BDA on the other hand (I also own a copy) is less accessible to beginners, but ultimately rewards the reader in a more consistent manner.


What's the legality of making your own pdf of Goodfellow's book from the html?


I’m not sure; I personally saved copies for myself, which I assume is okay. I believe (but could be mistaken) that it’s more about redistribution than consumption.


Nice List!


Pattern Recogntion and Machine Learning by Bishop

I’m highly skeptical of lists that do not include this standard text.


PRML was my introductory textbook, and it’s excellent. (The neural network chapters are a quite good introduction.) I also hear good things about Kevin Murphy’s text, which brings a more Bayesian approach and is preferred by some instructors I know, but I’ve never used it.


I think that before going into ML and AI it's worth to understand how to perform the quantitative analysis end-to-end. ML/AI would be just one of the process steps there, but without the bigger picture, you miss the value of it.

One of the good books to understand the quantitative analysis is "Keeping Up with Quants" https://www.amazon.com.au/Keeping-Up-Quants-Understanding-An...


Whatever happened to Norvig's AIMA book [1], did it fall out of fashion? I don't ever see it mentioned in these kind of posts nowadays.

Perhaps is because it doesn't cover deep learning?

1: http://aima.cs.berkeley.edu/


I believe there’s a new edition in prep.



Looks cool!




Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: