Tutorials about machine learning are, it seems two a penny. However, guides to machine learning by machine learning experts, however solid they may be as textbooks, frequently seem to lack ways in which the tools can be applied to a particular domain – in my case, to publishing.
In this respect the book Algorithms to Live By: The Computer Science of Human Decisions, by Brian Christian and Tom Griffiths is a marvel. Someone had the brilliant idea of combining a journalist with a cognitive scientist to examine widely used computer algorithms and to compare them with everyday life. In a very wide-ranging set of topics, the authors consider common computing topics such as randomness, caching, sorting, and scheduling, and describe firstly how IT approaches such topics, and then how such IT techniques could be applied to everyday life.
I was looking specifically for coverage of widely used tools in AI for textual analysis such as Bayes Theorem, but I was fascinated to discover the much wider scope of this book, since the combination of journalism and serious subject knowledge produces a result that is better, I think, than either author could have achieved alone. The book has some startling recommendations – I’ve already noted one idea they have on second-hand bookshops grouping together new acquisitions, and the suggestion that academic libraries should follow a similar system and display recently returned items as the user enters the building, since these are the titles most likely to be borrowed.
But there are lovely insights throughout the book. For example, I’ve been using email for years, but I never thought of email, in the chapter on networking and latency, viewed in light of protocols and transfers. Email, the authors point out, has an eternal memory, as does much social media. They point out that it would take decades for singer Katy Perry to respond to tweets from her 81.2 million followers. Email follows the same principle (I don’t have quite as many as 81.2 million people emailing me, but it feels like it sometimes). They describe the phenomenon of “bufferbloat”, by which messages are never lost, but simply pile up. As the authors state, “we used to reject; now we defer.” If you applied this principle to catching up on all the TV shows you have missed, or all the books you haven’t read, you would soon drown in uncompleted tasks. Technology, in other words, has introduced the challenge of bufferbloat, the idea of infinite storage, that we have to deal with, frequently just by discarding stuff.
Of course, the book is not perfect. In fact, they unwittingly reveal the sheer power of algorithms. Although we all live by algorithms in a rather unconscious way (for example, the way we simplify available information to solve a problem in order to reach a solution in a reasonable time for everyday living), this does not give a free rein to machine learning experts to implement ML-based approaches to solve all major societal problems. For example, the authors optimistically recommend using the technique of “Experimental Backoff” as used for managing networks to courts reviewing the cases of prisoners on parole violating their probation terms – instead of initial leniency, the authors recommend immediate imprisonment for a day. While it provides a perfect analogy with providing network access, the situation is completely different – there is a host of other factors involved in the decision to use imprisonment, and one would hope that the glib application of an algorithm would be as rapidly dismissed as anyone proposing to reinstate Jeremy Bentham’s ideas about dealing with prisoners; interesting in theory but invalidated by experience.
Nonetheless, an inspiring book that makes you want to reassess your daily processes, from the order in which you deal with your to-do list each day to the way you should find matching socks from the pile of clean washing. If anyone needed to be convinced of the power of algorithms to transform work and household processes, this is it. The authors even give a satisfying justification why a Nobel prize-winning economist ignored his own research when choosing what to invest in for his pension: “I should have computed the historical covariances of the asset classes and drawn an efficient frontier… Instead .. my intention was to minimise my future regret. So I split my contributions fifty-fifty between bonds and equities.” In other words, simpler might sometimes be better - and they have Nobel prize-winners to prove it.