The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World
Domingos wants to cover all of machine learning for the layman, but it winds up being a big mess. This is quite possibly the single worst thing I have read in my life about machine learning.
The level of explanation veers wildly from ridiculously oversimplified to technical minutiae. It is more confusing than enlightening as it goes through topics in an almost random order, scattering them all throughout the book. (You would think that Hume’s problem of induction, the underdetermination of data, Occam’s razor, the curse of dimensionality, and overfitting, would all be discussed in one and the same place in order to set the stage for how the various ‘tribes’ work, but you would be wrong.) The manic stream-of-consciousness writing style also drives me nuts, and the little medieval-fantasy passages come off as puerile. (I smiled. Once.)
The explanations are almost uniformly terrible (another reviewer asks if this is the worst explanation of Bayesian inference one has ever read; I would have to say that at least for me, this is competitive for that distinction), and most are explained as briefly as decision trees are endlessly waxed upon. Major premises like there being any really universal algorithm are poorly presented; compare Domingo’s argument for there being a neural algorithm with, say Jacob Cannell’s “The Brain as a Universal Learning Machine” (or my later “On GPT-3”), where Domingos is provoking rather than thought-provoking.
Content-wise, I have to seriously question the inclusion of evolutionary programming as a top-tier paradigm, and analogies hardly seem much more relevant a grouping either, and all that space comes at a huge cost of extreme superficiality about what deep learning is doing right now. Let me remark on how astounding it is to read a book whose self-proclaimed goal is to de-mystify machine learning for the layman, explain recent advances in deep learning that have created such media hype and sparked so much commercial and public and research interest, and which seems to only go from strength to strength to the point where sometimes it feels like one can hardly even skim a fascinating paper before yet another one has been uploaded to Arxiv, and which winds up doing little but explaining what backpropagation is and then passing grandiosely onto other topics and not, y’know, covering anything like solving ImageNet, caption generation, logical inference using reading of passages, etc. Or to read a decent capsule description of the general paradigm of reinforcement learning… and then see deep reinforcement learning described in a few sentences mostly to the effect that learning can be unstable—really? That is what laymen need to know about deep reinforcement learning, that—whatever it is—it can be unstable?
Oh, and he offers us his thoughts on AI risk, the fruit of his decades of experience with machine learning:
Relax. The chances that an AI equipped with the Master Algorithm will take over the world are zero. The reason is simple: unlike humans, computers don’t have a will of their own. They’re products of engineering, not evolution. Even an infinitely powerful computer would still be only an extension of our will and nothing to fear…The optimizer then does everything in its power to maximize the evaluation function—no more and no less—and the evaluation function is determined by us. A more powerful computer will just optimize it better. There’s no risk of it getting out of control, even if it’s a genetic algorithm. A learned system that didn’t do what we want would be severely unfit and soon die out. In fact, it’s the systems that have even a slight edge in serving us better that will, generation after generation, multiply and take over the gene pool. Of course, if we’re so foolish as to deliberately program a computer to put itself above us, then maybe we’ll get what we deserve. The same reasoning applies to all AI systems because they all—explicitly or implicitly—have the same three components. They can vary what they do, even come up with surprising plans, but only in service of the goals we set them. A robot whose programmed goal is “make a good dinner” may decide to cook a steak, a bouillabaisse, or even a delicious new dish of its own creation, but it can’t decide to murder its owner any more than a car can decide to fly away. The purpose of AI systems is to solve NP-complete problems, which, as you may recall from Chapter 2, may take exponential time, but the solutions can always be checked efficiently. We should therefore welcome with open arms computers that are vastly more powerful than our brains, safe in the knowledge that our job is exponentially easier than theirs.
How I wish I was making up these arguments. (Aside from the invocation of complexity theory which is, as stated, not even wrong as many problems we want AI to solve are not expressible as decision problems, the ones which are can fall into anything of ‘much easier than NP-complete’ or ‘much harder’, and a problem falling into a particular complexity class is no guarantee of safety in the first place, this sort of naivete is sad coming from someone so enthusiastic about genetic algorithms—where researchers routinely discover that defining a good reward/fitness/evaluation function is quite difficult and they have to fight their algorithms to get a useful rather than a hilariously perversely correct answer!)
Overall, I would absolutely recommend against this book for any laymen interested in statistics or machine learning. The explanations are so poor and garbled that you will either not learn anything or what you take away will be as likely to be misleading as not. You will be better off with Silver’s The Signal and the Noise, reading random presentations on Schmidhuber’s website, Bostrom’s Superintelligence or even Hutter’s Machine Super Intelligence or Domingos’s own “A Few Useful Things to Know about Machine Learning” (which was really good) or anything really. (Suggestions are welcomed on things I can recommend for laymen instead of this…)