Again wandering through a Wellington City Library branch, this time I picked up The Ethical Algorithm by Michael Kearns and Aaron Roth, from January 2020. It was an easy read for someone with a PhD in Computer Science and a BSc in Math/CS, and I finished it in about two hours. I didn’t pick up that much that was new to me, but I follow developments in this domain as an interested but technically-educated reader.
I also think all Waterloo Bachelor of Software Engineering grads ought to be conversant with the ideas in this book, which I’ll summarize (extremely briefly) below. Our grads have a solid undergraduate algorithms education, some statistics, and no mandatory machine learning/AI. So they should be able to read this book but they wouldn’t have been exposed to the ideas in our curriculum, which is problematic, given how much of our world is being shaped by AI.
Some of the sources of the cover blurbs are questionable. (I’m not going to elaborate on that point.)
Technical complaint: induced demand
There is an extended discussion of game theory using road navigation as an example. Yes, of course travel time depends on others' navigational choices. But the authors explicitly state that they consider the total amount of traffic to be constant. This is a terrible assumption, because it ignores induced demand. That is, the more lanes you build, the more low-value traffic you’ll attract. There’s nothing wrong with the computer science here, but it’s not properly modelling the real world.
On the other hand, the authors did a good job of giving analogies to describe Generative Adversarial Networks (which intuitively make sense to me now). The concept is a simple extension of fairly standard concepts; I just hadn’t bothered learning what GANs actually are. And of course I still haven’t internalized the technical definition. But I probably won’t need it anytime soon. GANs are really popping up everywhere.
I also appreciated the description (Chapter 4) of using differential privacy to prevent p-hacking. The bit just before that, about short descriptions and p-hacking, needed a couple of reads for me to make sense of.
I’ve had a problem with the modern usage of “algorithm”, and the authors point out that in modern usage we’re usually talking about AI meta-algorithms where we learn parameters and then run the models (which reduces to “algorithm” as we usually understand it).
Again, if you read like I do, then you’ve likely encountered these concepts. Here’s a summary of what we have in this book.
- privacy and differential privacy: how we can publish summary statistics without eroding individuals’ privacy;
- fairness (quite relevant for undergrad admissions, which is an issue for SE!);
- game theory;
- perils of multiple comparisons; and
- open problems eg interpretability of models.
I hadn’t seen the discussion of superintelligence vs diminishing returns previously, but the argument against the Singularity is plausible.
They discuss the unintended problems with optimization (e.g. you didn’t specify what to optimize against properly), which is also illustrated by Universal Paperclips.
I’ve had a discussion this week with a colleague about fact-checking vs the modern Internet and the various dysfunctions (e.g. QAnon). This book points out that human oversight doesn’t scale. (They also don’t discuss how human moderation is problematic for those who are hired to moderate, but that’s mainly off topic here.)
Really, this is a good defense of theoretical computer science, in particular of the value of mathematical formalism and algorithms, and a call to apply these in ethical ways.