EXCERPT: Concepts without categories

Download MP3

Welcome to Oddly Influenced, a podcast for people who want to apply ideas from *outside* software *to* software. EXCERPT: Concepts without categories

I grew up – and you probably did too – in a tradition that makes a big deal about hard-edged categories. An item either *is* in the category or it isn’t. That plant over there either is a tree, or it isn’t. Each category has a set of yes-or-no questions associated with it. When applied to an item, if the answer is “yes” for all of them, the item is in the category. If the answer to *a single one* of the questions is “no”, it’s not.

That in itself is not a big deal. Hard-edged categories do exist. But I blame Plato for saying that *ideas* – or concepts – are, all of them, examples of such categories. The problem is that’s just not true. It seems like a lot of philosophy is an argument between people who point out it’s not true, people who say it is so true, and people who say that maybe it’s not true but it *ought* to be true so let’s talk about that instead.

This episode isn’t really about any of those things. The question I want to answer is: if it’s not true, what is true?

In episode 40, I first described a neurologically-based theory of concepts and then used it to justify the behavior of collaborative circles. This episode excerpts the theory part for people who couldn’t care less about collaborative circles.

I’m making use of two books published in 2011: Louise Barrett’s /Beyond the Brain: How Body and Environment Shape Animal and Human Minds/ and Anthony Chemero’s /Radical Embodied Cognitive Science/.

≤ music ≥

I’m going to start with some introspection about a teensy-tiny creative breakthrough of my own. A question you may have heard before is: “what’s the difference between a chair and a stool?” You might answer that a stool doesn’t have arms or a stool doesn’t have a back. But you can easily buy something labeled “stool” that does have arms and a back. And bean-bag chairs are called chairs, not bean-bag stools, even though they don’t have arms or backs. Sort of? Until you sit down in them? Anyway, the conclusion you’re supposed to draw is that the boundary between the two concepts is fuzzy. Fine.

But: while I was writing the script for this episode, I found myself wondering what makes me choose one of those words over the other in boundary cases, and I flashed on the word “comfort”. I realized that I associate chairs with situations where comfort is at least somewhat important, and stools with situations where comfort takes a definite back seat.

Consider the classic bar stool. More important than comfort is that you can squeeze a lot of them in the linear space in front of a bar, and that they’re the right height for both sitting patrons and a bartender who’s standing and moving around. Or: I inherited a carpenter’s stool from my dad: it’s quite short, so comfort is clearly secondary to portability. Or: If you do an image search for “dunce cap”, you’ll find many pictures of people sitting in corners on… stools. Because they’re being punished, so of *course* they’re not sitting in a comfy lounge chair.

≤ short music ≥

OK. Let’s say that my conception of “stool” has now changed. How did that happen? What things mattered, specifically?

1. I had an idea and I sanity-checked it against three examples: bar stool, carpenter’s stool, dunce sitting in a corner. Examples seem important.

It’s interesting to note that I treated my memories as if they were raw perceptions. That is, I didn’t actually go look at any stools. You could think of that two ways. The first is that memory is just a stored perception. But we know that’s not true. Memories are incomplete and malleable; eye-witness testimony, for example, is notorious fallible (except, alas, to the keepers of the US legal system). Memories just *aren’t* verbatim storage of sense perceptions. They’re more like an awfully lossy form of compression. Like in JPEG compression, they’re compressed so they can be expanded into something good enough for our purposes, such as using them as examples of concepts like “stool”. Features of a memory unlikely to be useful to animals in our ecological niche are discarded, just like MP3 encoding might discard very high or low frequency sounds.

Scientists like Barrett go further: they think perception itself is very lossy and purpose-dependent. We don’t come anywhere near constructing a model of the world from our sense perceptions: we perceive only those features in the world that are likely to be useful.

So, the claim I’m going to go with here is that there’s no fundamental difference between a memory and a perception, not when it comes to evaluating concepts.

2. Repetition strengthens and adjusts concepts. The three examples were better than one. And I’ve now written four versions of this “stool” story for the script you’re listening to me read. If I’d just flashed on the idea of “comfort” and not worried it like a dog worrying a bone, I’d probably have forgotten it by now. Instead, it’s stuck in my head.

3. Definitions include debris. We tend to think of concepts like “stool” or “tree” or even “concept” itself as being labelled by words. Barrett, following Andy Clark, suggests that words come first. It’s not absurd to think that what learning is, is attaching (lossily-compressed) experiences to preexisting words.

Consider this everyday scene: A child is walking with his mother. “Look at the doggie!” she says, pointing. The child reacts with some sort of interest and gets a rewarding smile. After several similar episodes, spread out over time, the child starts pointing and saying “doggie!” himself. Each time he’s rewarded with the pleasure of his mother. Once in a while, he gets it wrong, illustrating that counterexamples are important too. He might point at my daughter’s mostly-Chihuahua rescue dog and say, “rat!” His mother will correct him kindly and say, “No, dear, that’s a *little* doggie”. Over time, the child will be able to reliably identify dogs, and we can say he’s learned the concept of “dog”.

I say “concepts include debris” because concept formation is not exceptionally scrupulous about *what* compressed experience it attaches to words. Like a logjam in a river catching logs, twigs, empty beer cans, and fishing line, a concept will have associations that don’t make a lot of sense.

I apologize to my wife, who listens to these podcasts, but the sight of umbrella clotheslines – or even hearing them mentioned, as she just now has – provokes in her the same feeling as in Emily Dickinson’s poem about encountering a snake, which ends:

But never met this Fellow,
Attended, or alone
Without a tighter Breathing
And Zero at the Bone –

Dawn’s ability to successfully navigate a world with clotheslines, despite her visceral reaction or association, means that we might as well say that she has the same “concept” of “clothesline” as I do, even though hers clearly differs in a way that’s important to her.

4. It matters how meaningful experiences are. An interesting fact about rabbits is that there’s a strong sense in which a rabbit can’t smell something that doesn’t yet matter to it. That is, if a rabbit has never smelled a carrot before and you waft carrot smell into its hutch, nothing much happens in its brain’s olfactory bulb, where smells are processed, and there are no resulting signals sent to its cortex. If you sent it carrot scent 100 days in a row at 10 a.m. each day, *still* nothing would happen. If, however, you deliver sugar water at the same time as the smell, the rabbit will begin a process called Hebbian learning, in which certain neurons will start firing together as a self-reinforcing cluster – that is, the output of the cluster feeds back into the input. Interestingly, the feedback loop is tuned to the frequency of a rabbit’s sniffs, so each new sniff of carrot gives the cluster a little bit of a push, like pumping your legs in a swing. When the cluster reaches a certain threshold, a circular activation pattern develops across the entire olfactory bulb, and that (or a derivative of that) is what’s delivered to the cortex, where it provokes some sort of action. That circular activation pattern is called an “attractor”, in the same sense it’s used in chaos theory: a not-exactly repetitive pattern that a system tends to settle into.

As a side note, while it might be fair to say that the smell of carrot is represented in the olfactory bulb by a collection of neurons that, when stimulated, fall into an attractor, it’s probably not fair to say that the rabbit has a concept of “carrot” that pulls together the smell and taste and color and meaning of a carrot into a single “place” that we could say “represents” the idea of “carrot”. The reason is that neurons are fantastically expensive, as cells go. Evolution can be haphazard, but there’s strong selection pressure toward thriftiness with the body’s energy supply. It does not appear a rabbit *needs* an internal representation of “carrot” to get on with its rabbit life, so it doesn’t have one.

5. Words are good for learning. Barrett and Clark agree that you and I have a concept for carrot, because we have language that allows concepts to accrete around words. For the types of lives we humans live, concepts are worth the cost.

“Clark literally thinks of language as a tool. That is, it hasn’t changed anything about the basic structure of our brains, or how they work, but rather it complements their functioning. It allows a parallel-processing pattern-recognizer to process things in serial fashion according to a set of precise rules. In other words, it allows the brain to operate “as if” it were a modern digital computer even though, as we’ve seen, brains don’t seem to work in that way at all. Language gives human brains a new way of dealing with the world.”

Barrett doesn’t quite say so, and I haven’t read Clark’s book yet, but it seems to me part of what language is *for* must be to allow us to form more concepts – and form them more quickly – than if we relied on experience alone. Language allows me to juxtapose an abstract feeling like “comfort” to an abstraction of a set of physical objects like “stool”. Rabbits, Barrett believes, can’t do anything like that.

Here’s another example. I’m going to ask you a question and assume you’ve never heard it before. Here goes: is the Pope a bachelor?

Let’s assume that concepts are something like the smell of a carrot. A “latent” concept is a set of nerve cells, wired in a particular way. A concept is *activated* when those cells are stimulated and fall into a self-reproducing attractor pattern. Because you’ve never heard the question before, let’s say the concepts of “Pope” and “bachelor” have never both been activated at the same time, at least not in conjunction with a motivation to link the two. Now they are. A quick victory for the power of words, together with a possible change to your concept of “bachelor” (represented by possibly ad-hoc changes to neurons).

6. Argument is effective when it’s affective. I’m being intolerably cutesy with the words, but I want to extend point 4 – that it matters how meaningful experiences are – into the realm of sequences of words. That is, we saw that rabbits learn smells if they’re meaningful: they are associated with rewards or, for that matter, dangers. I implied that a child learns “doggie” faster because of the emotional rewards his mother gives him. Here, I want to claim the same is true of *arguments*, using the word both in the sense of “a sequence of statements that justify a conclusion” and “an at-least-slightly emotionally charged discussion”.

As an example, I offer myself. I’ve had enough bad experience with arguments that start by appealing to dictionary definitions that an argument with that structure has very little effect on my understanding of a concept. You could say that it would, at best, add a tiny twig or a discarded fishing lure to my logjam. Or, hearing an argument structured that way for the ten thousandth time, I might – in a fit of pique – rip the fishing-lure-sized definition out of my logjam and throw it away.

(I may be getting somewhat carried away with this metaphor.)

But for other people, dictionary definitions carry a weight of authority. Let’s say they can add a big log to the logjam or reinforce one already there.

The point here is that different arguments “hit” different people differently, and the same argument may hit the same person differently at different times. David Byrne sang “Say something once, why say it again?”, but he sings it in a song titled “Psycho Killer”. I’m not a psycho killer, and I’ll assume you aren’t either. For people like us, whatever “it” is *has* to be said again: repetition and variation and some degree of tailoring is required.

≤ music ≥

The most important point in this episode is that concepts have to be *effective* more than they need to be *true*. Now, there are zillions of situations in which truth is required for effectiveness, but there are other situations where it isn’t, especially since we are social animals operating in an environment whose most salient feature is *other humans*.

How well you and I can work together depends most of all on whether our concepts are compatible enough with each other *for the task at hand*. If it never matters to us whether the Pope is a bachelor, you can say “yes, technically”, I can ostentatiously refuse to have an opinion, and we can move on. Unless you whip out a dictionary and insist on discussing the definition of “bachelor”, in which case I will seize the book from your hands and run off, cackling like a fool.

Thank you for listening.

EXCERPT: Concepts without categories
Broadcast by