Three-Toed Sloth

Books to Read While the Algae Grow in Your Fur, April 2021

Attention conservation notice: I have no taste, and no qualifications to opine on ethics of any sort.

Michael J. Kearns and Aaron Roth, The Ethical Algorithm: The Science of Socially Aware Algorithm Design

There are, roughly speaking, three schools of thought when it comes to "fairness" and "ethics" in artificial intelligence machine learning predictive statistical modeling and data mining. I will caricature them as follows:

1. "Problem? I don't see any problem": maximize accuracy (or utility, etc.), and let the results take care of themselves.
2. "Everything is problematic": the data sets are biased, the objective functions to be maximized are biased (in some more obscure way), the very maximization algorithms are biased (in some yet more obscure way), and the only hope is to appoint duly-certified ethicists as censors trust that can all somehow be re-imagined after the arrival of the millennium / revolution.
3. "Problems? I'm good at solving problems! what penalty term should we add to the Lagrangian?"

This book is the best presentation I have encountered, and indeed about the best I can imagine, for this third, temporizing school of thought. (It is, in case that's not clear, the tendency with which I have the most sympathy.) That is, this book tends to regard ethical and political desiderata as constraints which should be imposed on algorithms that are otherwise seeking to optimize some well-defined objective function (such as travel-time for mapping software, or "probability that the user will watch the recommended movie" for recommender systems, etc.). There is a strong analogy here to a certain kind of technocratic, American-sense liberal approach to public policy, in which private firms maximize profit, subject (ideally) to constraints imposed by regulation ("don't dump too much dioxin into the water supply"). (I don't recall the book making this analogy explicit.)

I used this book quite successfully in my data mining class, but my students there found the most technical parts (like "possibility frontiers") the most congenial, and the more rhetorical-argumentative bits about fairness more preplexing. I strongly suspect this reflects having a very unusual audience. I would cheerfully teach from it again, and strongly recommend it to readers interested in these subjects, perhaps especially if they're new to this area. §

Disclaimers: I have been an admirer of Kearns's work since the 1990s, I know him a bit from conferences &c., and I requested an examination copy of this book before assigning it to my class.

Walter Jon Williams, Fleet Elements

Military space opera mind candy of the very highest grade. For one thing, it earns operatic levels of emotion. --- Previously. §

Ausma Zehanat Khan, The Black Khan

Continuing an epic fantasy saga where a lot of the details are the recent history of Afghanistan and environs with the serial numbers filed off. Only in this installment we spend a lot of time at the court of Isfahan Ashfall (via the ruins of Nishapur Nightshaper), complete with a scheming Nizam al-Mulk Nizam al-Mulk. Also, there is even more angsty romance than in the first volume. (Fortunately, AZK writes angsty romance well.) There are at least two more volumes to the saga, which I intend to devour as soon as I can arrange suitably long stretches of un-interrupted time. §

Dennis Culver and Justin Greenwood, Crone

Comic book swords-and-sorcery mind candy, in which the former Red Sonja Bloody Bliss, now the titular crone, is dragged out of retirement to re-confront a Dark Lord she knows she killed...

Tony Cliff, Delilah Dirk and the Pillars of Hercules

Comic book historical-fantasy mind candy. (Previously.)

K. C. Constantine, Joey's Case

Similar remarks to last month's entry.

Books to Read While the Algae Grow in Your Fur; Scientifiction and Fantastica; Pleasures of Detection, Portraits of Crime; Enigmas of Chance; Automata and Calculating Machines