Learning about (machine) learning

Last week I trotted off for my first Davos experience with four books in my bag and managed to read only one – no doubt old hands could have warned me what a full-on (and rather weird) experience it is. The one (and that read mainly on the journeys) was The Master Algorithm by Pedro Domingos. I was impressed when I heard him speak last year & have been meaning to read it ever since.

The book is a very useful overview of how machine learning algorithms work, and if you’ve been wondering, I highly recommend it. On the whole it stays non-technical, although with lapses – and I could have done without the lame jokes, no doubt inserted for accessibility. The book also has an argument to make: that there is an ultimate ‘master algorithm’, a sort of Grand Unified Theory of learning. This was a bit of a distraction, especially as there’s an early chapter doing the advocacy before the later chapters explaining what Domingos hopes will eventually be unified.

However, the flaws are minor. I learned a lot about both the history of the field and its recent practice, along with some insights as to how quickly it’s progressing in different domains and therefore what we might expect to be possible soon. Successive chapters set out the currently competing algorithmic approaches (the book identifies five), explains their history within the discipline and how they relate to each other, how they work and what they are used for. There is an early section on the importance of data.

As a by the by, I agree wholeheartedly with this observation: “To make progress, every field of science needs to have data commensurate with the complexity of the phenomena it studies.” This in my view is why macroeconomics is in such a weak state compared to applied microeconomics: the latter has large data sets, and ever more of them, but macro data is sparse. It doesn’t need more theories but more data. Nate Silver made a simliar point in his book The Signal and the Noise – he pointed out that weather forecasts improved by gathering much more data, in contrast to macro forecasting.

Another interesting point Domingos makes en passant is how much more energy machines need than do brains: “Your brain uses only as much power as a small lightbulb.” As the bitcoin environmental disaster makes plain, energy consumption may be the achilles heel of the next phase of the digital revolution.

I don’t know whether or not one day all the algorithmic approaches will be combined into one master algorithm – I couldn’t work out why unification was a better option than horses for courses. But never mind. This is a terrific book to learn about machine learning.