My pet pig won't fly and I want a refund

Abstract

Pigs can't fly. Any person buying a pig should understand this – it would be absurd to be upset that they can't fly or play poker. But pigs are amazing creatures and can do many interesting and useful things.

Since I retired to Florida, I have been a bit at loose ends. So, I got a pet pig. I was pretty excited. On the billboard they looked real cute, flying around on those little wings. When I picked up the piglet they told me he would get bigger fast. Sure enough, he ate like a pig. Named him PeteyPig, made a pen for him in the front yard and he rolled around in the mud and ate. But he wasn't really that cute and he sure wasn't flying. I thought, he's gotta be defective – the brochure showed pigs doing all sorts of neat things in addition to flying: carrying golf clubs, playing poker. So, I went down to the pig emporium (Fig. 1).

Figure 1. Pigs can't fly. This image was created with the assistance of DALL⋅E 2 from open.ai (https://labs.openai.com).

I said, “Hey you sold me a bad pig. He won't play poker and I am pretty sure he is never going to fly.”

The salesman stared at me, then said, “Sir, you realize that those billboards and brochures are just to catch your eye? It's a pig. Pigs don't fly.”

I wasn't having it. I replied, “Look here, your brochure shows poker-playing pigs, what does that mean? Petey just wallows in the mud and expects me to feed him all the time. How is that fun?”

That salesman looked worried. “Sir, you understand that is what pigs do? They don't have wings and can't play golf. That's just marketing.”

Nope I thought, I am not going to be taken for a fool. “Why are you selling these pigs as pets at all? Who would want a nonflying, fat, muddy pig? You need to fix your pigs, clean them up, give em some wings, and teach them to play poker!”

The salesman gave me another look. “Look sir, pigs are great. They can do amazing things. A miracle of nature. They're playful and smarter than dogs. But pigs will do what pigs will do. Complain all you want, but they won't fly. If you want a golf club-toting poker buddy, hire someone. Flying is out.”

Well, I had heard enough. I walked straight back to the Villages. Petey was in the front yard, covered in mud. I tossed him some carrots and sat down. Ever hopeful, I said, “Petey old buddy, let me show you the queen of hearts….”

Bowers et al. build a straw house by motivating their arguments through quotes that are more marketing than scientific claims. Much like our protagonist, we need to be smart consumers of science. I don't think there is much actual confusion that deep neural networks (DNNs) are “models of the human visual system.” Rather, like the computer vision models that preceded DNNs, they serve as “proxy models” that surface the role(s) of assumptions and constraints in complex systems (Leeds, Seibert, Pyles, & Tarr, 2013).

As proxy models, DNNs are remarkable because of what they can do in comparison with prior models. DNNs learn task-relevant representations that are often well aligned with representations in neural systems that support a common task (Yamins & DiCarlo, 2016). This level of alignment is a dramatic shift from the mostly much poorer attempts to account for neural data that preceded DNNs. Even so, it should be obvious that DNNs, in and of themselves, don't have many of the characteristics that define intelligence in biological systems.

As a field we should have a productive discussion about what inferences we can draw from DNNs and other computational models (Guest & Martin, 2023). However, such discussions should involve less hyperbole (“Deep problems…”) and less handwringing about what current models can't do; instead, they should focus on what DNNs can do. They might be pigs, they will never fly, but they can do some pretty cool stuff. We should figure out how and why.