4 ms·
Original title: Programmer, Teach Yourself Foundations of ML and AI with these 6 Books Books: - Ian Hacking - Introduction to probability and Inductive Logic
by dragandj 8y ago
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
- s3f0 8y agoNice List!
- internet555 8y agoHow 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
- stochastic_monk 8y agoPiggybacking 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.
- basjacobs 8y agoAny reason to pick Mohri over Shai Shalev-Schwartz, or the other way around?
- stochastic_monk 8y agoI 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.
- basjacobs 8y agoGreat, I'll start with SSS. Any tips for a good optimization book to read afterwards?
- GlenTheMachine 8y agoWhat's the legality of making your own pdf of Goodfellow's book from the html?
- stochastic_monk 8y agoI’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.
- wenc 8y agoI 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.