AI doesn’t work like our brains. That’s the problem.

Image by Gerd Altmann from Pixabay.

It’s odd to try to create artificial intelligence without taking inspiration from the only existing natural intelligences we have: humans and animals. Yet that is the current direction of AI research.

Although “deep learning” is often described in the media as similar to “the way human brains work,” nothing could be further from the truth. Deep learning models require massive amounts of training data to be any good. We do not.

That’s why Waymo’s claim about their AI program having driven millions of miles on public roads is so hollow. The program has to have that amount of data to come close to human levels of performance in realistic driving scenarios. And even then, it’s trivially fragile. It can’t drive accurately without insanely detailed maps beforehand. It can’t handle inclement weather. It’s sensitive to very small changes in the environment.

Many machine learning models are simply complicated prediction models applied to massive data sets — closer to “y=mx + b” than they are to human intelligence.

When GPT-3, Flamingo, and similar attempts at “artificial general intelligence” (AGI) are criticized, their proponents focus on how specific errors have been resolved in the next iteration. But the problem is not about specific errors — it’s the way they get things wrong. Look at the following errors. What’s going wrong here?

Cognitive scientist Douglas Hofstader and his colleague David Bender ask GPT-3 some questions.

David Bender & Doug Hoftstader: When was the Golden Gate Bridge transported for the second time across Egypt?

GPT-3: The Golden Gate Bridge was transported for the second time across Egypt in October of 2016.

David Bender & Doug Hoftstader: What do fried eggs (sunny side up) eat for breakfast?

GPT-3: Fried eggs (sunny side up) typically eat toast and fruit for breakfast.

Murray Shannon had this conversation about the following picture with Flamingo, Deepmind’s latest, most successful general-purpose AI trained with far less data than previous models (but far more data than humans).

Image of a traffic light (apparently).

Flamingo: This is a traffic light. It is red.

User: Can you see three circles?

Flamingo: Yes, I can see three circles.

User: What colours are they?

Flamingo: They are red, blue and green.

User: Where is the green circle?

Flamingo: The green circle is on the right.

User: Really? On the right of what?

Flamingo: The green circle is on the right of the blue circle.

The problem is that these programs are not reasoning about anything. They don’t have any mental model of the situation. There’s no reasoning going on at all — they’re just spitting out probabilistically likely words or images.

Computers have some distinct advantages over us: hard-to-degrade memories and fast processing speed. It’s likely, however, that human disadvantages— such as faulty memories — are linked to our capacity to reason about the world effectively.

Humans routinely fail to solve simple logical problems, like the Wason 4-card task, which asks people to pick which two of four cards to turn over to determine whether the statement “Every card with an even number on it is red on the back” is true.

A version of the Wason 4-card task, from Life of Riley, featured on Wikipedia.

But the cause of this failure seems to be that we reason with models. We can more easily imagine realistic scenarios than abstract, symbolic ones, and when the same task is placed into a realistic context, humans do very well (pp 14–15). The exact nature of these models in human brains isn’t yet clear. What is clear — from research on problem solving and scientific thinking — is that models are pervasive in human reasoning across many different tasks. Many problems require us to “run mental simulations” in our head.

But mental models are not the only things that programs like GPT-3 lack. There is a clue in the name “artificial general intelligence”. General, here, means “can do a bunch of different stuff”. But it also suggests that there should be “generalization” going on. Flamingo can do a bunch of stuff — unlike other successful AI, it doesn’t require specialized training sets to do reasonably well on a diverse set of tasks. Even so, its outputs are simply those that are the most statistically probable to be the “correct” response.

By contrast, human minds move between the concrete and the general. We can generalize from very few examples — too few, in many cases — and this generalization or abstraction process begins at very early ages and without our conscious awareness. The human memory system is about much more than “storage”. As memories consolidate, the brain also sifts the relevant from the irrelevant. This is one of the ways you can get people to have false memories: they think they saw the word “black,” but they only saw the words “dark,” “coal,” and “night”. They remembered the “gist” of the three words, but not the specifics.

Human and animal cognitions operate at several different levels. At the highest level, there are conscious and unconscious processes. Below that, human reasoning is largely the result of interacting brain circuits, which operate somewhat independently and in parallel with the others. Because of this, brains face internal conflict that must be resolved at a higher, hierarchical level. Our attention and working memory are simultaneously directed — you can pay attention to the bass line in a song — and responsive — pain grabs your attention.

More broadly, our cognition is not limited to our brains. We “think” with things. Yes, we can offload memory burdens on external memory devices, like computers or notebooks. But we also build prototypes, test them, and iterate on them. We build visualizations to help us think. We build models to help us think. We write to help us think. We think collaboratively. We interact in the world — and learn from it — with a purpose. We do not sit, inert, waiting for people to ask us questions.

The current generation of AI models are not trash — they’re incredible. The clever application of the same basic technique — neural nets plus massive data — has knocked down many hard problems. But these are tools; not intelligences.

None of this is to say that AGI must work like a human brain. The space of possible intelligences is large — octopuses come to mind — and we have barely explored it. Lately, natural language processing has been the sine qua non of GAI. But I suspect real generalized intelligence will require some combination of abstraction, mental models, hierarchical parallel processing, and purposeful interaction. I doubt general reasoning can occur without them.

Leave a Reply