This is a piece I originally wrote on Facebook back in 2020.

Apparently an experiment found that if you paint eyes on a cow’s rump, lions won’t pounce on it. A fun fact. What I find especially fun is that even cows with just a cross painted on their rump were attacked less than cows with nothing painted at all. The moment I read it, I thought of GANs. GAN stands for “generative adversarial network,” a deep learning model that pits two networks against each other in competition to produce a desired output. The two networks are the generator and the discriminator. A common analogy is that the generator is a counterfeiter and the discriminator is an inspector. During training, the counterfeiter network slips its forged bills in among the real ones, and the inspector tries to tell which bills are fake. In the early stages of training, the counterfeiter produces images that are little more than noise from a random seed. But because the inspector is just as undertrained as the counterfeiter, it can’t reliably tell that noise apart from real bills. In an ideal training run, as time goes on the counterfeiter makes increasingly convincing bills, and the inspector gets better at distinguishing real from fake. Generally the inspector learns faster early on. Sometimes the inspector learns so well, so fast, that the counterfeiter can’t learn at all. One thing worth pointing out here: the counterfeiter never once sees a real bill throughout the entire training process. The counterfeiter works only from the inspector’s verdicts, tweaking its drawings little by little to make bills that can fool the inspector. Naturally the counterfeiter has the harder job, so striking the right balance — keeping both learning at a similar pace so training can keep progressing — is a genuinely important challenge in GAN models.

I don’t know much about biology, but it seems to me you can find a similar dynamic between the counterfeiter and the inspector in the relationship between cows and lions. The lion is fooled by the crude eyes painted on the cow’s rump. In fact, even when what’s painted barely looks like eyes, the lion is sometimes fooled. A lion that gets tricked by fake eyes and goes hungry is more likely to die without leaving offspring. As generations pass, the lion’s parameters get updated one step at a time. The next generation of lions — the ones that weren’t fooled by fake eyes, that ate their fill and left offspring — may be a little better at telling fake eyes apart.

The attached images were generated by a GAN I built myself, trained on a dataset of 70,000 face images, after 1, 2, 3, and 50 epochs respectively. At 1 epoch you see images that look more like patterns than faces. And yet, even with such crude images, the discriminator sometimes gets fooled — just like the lion fooled by the cross on the cow’s rump. Look at the results at 2 and 3 epochs and, while they still don’t look like human faces to us, they’re a little closer to one than before. Even a face that looks shoddy to a human is, like the eyes on a cow’s rump, good enough to learn from as long as it can fool the discriminator. Maybe creatures that practice mimicry evolved in this same way?

What about humans? Back in my army days, when I was reading a lot, I read a book called The Mating Mind by Geoffrey Miller. Despite the kindling-for-the-fire whiff the title gives off, it’s a pretty entertaining popular book on evolutionary psychology, and this experiment suddenly brought it to mind. The author argues that the reason humans came to possess such remarkable intelligence is “sex.”

The male peacock has lavish feathers. Looking at those feathers, you can’t help wondering how an animal saddled with such a conspicuous, useless body part managed to evolve and survive all this time. What survival benefit could a peacock’s tail possibly confer that let it survive natural selection? There are many hypotheses, but Darwin said this kind of evolution is the result of sexual selection. Because females preferred peacocks with showier tails, those genes were passed on, and the species evolved in the direction of ever more elaborate male tails. Geoffrey Miller takes Darwin’s argument a step further, marshaling all sorts of principles to argue that human intelligent behaviors — language, art, morality, creativity — are achievements wrought by the process of sexual selection. Just like the peacock’s tail, human intelligent behavior carries sexual appeal, and that is why it could develop so intricately.

Here too you can find the relationship between generator and discriminator. To appear to be acting intelligently, you need intelligence. And to judge whether another human is intelligent or not, you also need a certain level of intelligence. The two interact, advancing each other. Seen this way, I too am ultimately a product — like a GAN — of my ancestors interacting and selecting. Thinking about the GAN that’s so curiously entangled with my own existence is a great deal of fun.

A caveat: I’ve used terms that make it sound as if the networks act with personality and intent, but this is just a metaphor to aid understanding. In reality the generator doesn’t feel joy when it successfully fools the discriminator, and the discriminator likewise doesn’t grow sad or blame itself. The discriminator and the generator are sets of computations.