Well, I'm biased but also experienced (former academic, now industrial practitioner doing a startup). My advice to the people who think that industrial ML/AI is "just applying some base cunning to 1970s problems" is to try to trade equities and generate durable above-benchmark profits over 2-3 years (where the industrial state of the art controls hundreds of billions per year and the competition can pay $300k+/year for fresh-out-of-grad-school talent). Algorithmic equities trading is sort of the "UFC" of ML; money talks and, er, bovine byproducts walk. Having said that, I'm now out of equities trading because I figured having a startup was more than enough risk for me; so I'm somewhere between talking and walking, I guess :P.

In terms of books: I recommend grabbing as many domain-specific books as possible rather than general-purpose ML books. Look at bioinformatics, speech processing, text processing, image processing, algorithmic trading, epidemiology, system identification, adaptive filtering, etc.; each of these disciplines has its own approach to signal/feature extraction, and ML gives you a unified way to fuse multiple signals into an estimate/decision. In my experience the tricks of the trade arise from learning lots of domain-specific "hacks" and thinking about how they generalize to other problems (one example: look at the Viola-Jones feature extractor for images and think about how you might apply that in, e.g., equities trading).

Just like with programming, practical ML is best learned by just solving a bunch of problems and learning what works (informed by a theoretical framework about what can't possibly work :-)

Parent is talking about Agent systems and just about everything that isn't regular statistics. Basically, in the past 20 years, computers got 1000 times smarter and people didn't, so old statistical models became tractable to apply to terabytes of data, and the schools of "invent a thinking algorithm" stopped being relevant.

/slightly bitter former "academic AI" student.

It's not really Academic vs Industry, though. It is Agents and Logic vs Statistics.

The standard text is Elements of Statistical Learning. It is a grad-level and mostly theory. For goofing around in Python, Programming Collective Intelligence

I agree that Rusell and Norvig AI doesn't have much penetration yet. As for Elements of Statistical Learning... That's the canonical textbook for ML. If industry relies on splines, boosting, and support vector machines, then it is really not that far from modern academic ML research.

There was a neat tangent on this subject when one of the more recent free online ML classes came up. Basically the free course is "ML for practitioners" and the Stanford-only course is "ML for researchers". The former group is less interested in advancing the state of the art and more interested in using known ML techniques to solve business problems.

Turns out if you have oodles and oodles of data, and oodles and oodles of computing power, a lot of the simplier older solutions work quite well.