"One useful question is whether neurons and circuits carry out identifiable transformations on their inputs while representing something relevant to the world they inhabit. At the level of single neurons, local circuits, neural populations, and large-scale networks, the answer is almost certainly yes."
I am curious. Are there folks in the neuroscience community who would disagree that neurons and circuits carry out identifiable transformations and their functional organization not only correlates with aspects of the external world but has downstream causal relevance? If so, what explains the disagreements? And if not, then what's the whole "representation" theoretical debate really about? It seems to me that once we concede that functional organization captures some property of the external world and is causally relevant, debating whether this is "representation" or not seems rather superflous.
I agree with you. The debate is mostly within the philosophy of mind/science, and I’m not sure it has much impact at the level of systems or cognitive neuroscience. The working assumption that circuits carry out identifiable transformations with downstream causal relevance is pretty much the foundation the field operates on. The actual debates are around how causality is established between particular circuit configurations and behavioral outcomes, and to what degree of specificity. Naturally, the fly community have much better tools to do so!
Michael, lovely piece and unusually honest. “What Might Cognition Be, If Not Computation?” is exactly the kind of question cognitive science has been circling for decades, and it still matters. I really liked that you named the “describable by math ≠ computing” trap yourself; that move saves the argument from the cheap version of this debate. And the fly example is beautiful, one of those rare cases where the levels really seem to close down to the wiring. “The synaptic wiring is the computation” is a great line.
Two thoughts/questions.
First, the definition of computation still feels broad enough that I start wondering about the pancreas: it also transforms inputs into outputs according to rules over variables that matter to the organism. So what exactly disqualifies it? I’m sure you have an answer, but making the cutoff explicit would strengthen the conclusion.
Second, regardinig representation: the fan-shaped body lesion indicates that the signal is necessary for function, but necessity alone may not be sufficient. A power supply is necessary for navigation too, and it doesn’t represent heading. The harder comparison might be van Gelder’s Watt governor rather than the pendulum: a system that causally organizes downstream behavior in a content-sensitive way yet which many would still resist calling representational.
The thing I’d most like to see you chase is that your strongest examples, the ring attractor and the integrator, are not merely input-output mappings. They hold state: the bump persists in darkness, the eye stays put. That sounds like memory/maintenance or more generally, state dynamics, rather than just an input >> output transformation. So maybe the intro framing slightly undersells your own best cases.
Looking forward to Part Two. I’m curious whether predictive processing will come in as support for the thesis or as a competing way of telling the story :)
Great points! I don’t object to the pancreas performing computations but I think the spatiotemporal scales may not be optimized for what we consider ones relevant for thinking. I think time makes a big difference here, but will need to think about that. Thanks so much for this terrific input/pushback. It’s great
Interesting article Michael. I think any brief that gets us to thoroughly explore a phenomena and opens our mind to new ways of looking at it, potentially tying together activity we may have previously overlooked, is a useful tool.
However, I worry that the computer metaphor has outlived its usefulness and may now be giving rise to misconceptions. I wonder whether it isn't worth reaching for a fresh metaphor that better reflects the situation and doesn't lead to the misnomer we're similar to AI.
An example of the problems with metaphors is that when you find that parts of the brain involved in navigating spatially may also be involved in navigating concepts and memories, precedence leads us to imagine we explore ideas in a spatial way, when if we'd made these discoveries the other way round we may imagine we explore spaces as ideas. Do you see what I mean?
Well, I think it's hard to come at it from the other direction because we tend to start from physics. For example, kids (and non-human animals) have a very good sense of intuitive physics well before they start developing metacognitive capacity. So I don't know, just my immediate reaction.
Yes I didn't mean it literally. Merely that when we apply a metaphor to something this tends to be how we frame it. In this case what is going on maybe more primitive and biological than spatial mapping (hence why it is implicated in other activities). So it might be more fundamental abstract mapping of some other kind. I'm not saying it is. I'm just talking hypothetically. So if we thought it was doing spatial mapping we'd assume not to bother testing if it's involved in something like taste. When actually, if our thinking is more agile, and it's more loosely to do with abstract mapping in general, we might think to test for this.
Again, I don't mean this literally. You're in a much better position to know the particulars of what you're studying. I'm merely expressing a concern at how neuroscience might be riddled with metaphors based on the precedence of discoveries and what was first attributed to them, as well as the overriding metaphor of a computer, and the effort to make that metaphor fit. Such artefacts might be cluttering the intellectual landscape somewhat.
It's certainly a wonderful puzzle and articles like yours only make it more intriguing. Thanks 😊👍
This is a thoughtful, serious article. I was glad to read it.
I need models for the brain. I don’t know how else anyone can even begin. The brain is too much. Too many cells, too many pathways, too many things happening at the same time. So we reach for something smaller. A computer. A prediction machine. A network. Something we can hold in our hands, at least for a minute.
And the better the model is, the more dangerous it gets.
Because after a while, it starts to feel like we understand the thing itself. Not just the model. The brain. We read a careful explanation, follow the parts, see how one system talks to another, and some part of us says, “Ah. Now I get it.”
I don’t think we do.
I have read a ridiculous amount about the brain over the years. Papers, studies, books, theories, all of it. And the more I read, the less comfortable I am pretending we understand very much about how the brain actually works. We know pieces. Some pieces are astonishing. But still, pieces.
That was the thought I had while reading this article. It gives the reader a real feeling of understanding. Not in a cheap way. It is careful and deep. But still, I kept thinking about what was not in the frame.
Neurotransmitters, for one.
The whole article can move through an impressive explanation of the brain as a kind of computational system, and somehow the chemical mess of the brain is mostly off to the side. That is not a small omission. We don’t even know if we have identified all the neurotransmitters. And we certainly do not know, in any complete way, what each one does in each part of the brain, in each state, in each person, at each moment.
That is where psychiatry gets humbled very quickly.
A person takes a medication. It helps. Or it does nothing. Or it helps the thing it was supposed to help, but now sleep is strange, or desire changes, or the body feels wrong, or a side effect appears that nobody quite predicted. Then we adjust, switch, add, subtract, wait, guess again.
That is not because psychiatry is useless. It is because the brain is not giving up its secrets that easily.
So yes, I am grateful for models like the one in this article. We need them. Without them, we are staring into fog. But I also don’t want to mistake the flashlight for daylight.
Thanks: I definitely agree with your high-level point about confusing bits and pieces with a grand theory. On the neurotransmitter bit, they are included in the models. I haven’t discussed them explicitly but I don’t know that doing so would provide more conceptual clarity. Regardless, your points are well taken.
“Describing the brain computationally also lets us see deep links between things that otherwise appear unrelated. Who would have guessed that the machinery we use to navigate physical space might also help organize knowledge more generally?”
Interestingly, I interviewed someone for my book who has aphantasia, and this is exactly how he remembers things, not in images, but spatially, like rooms in a building.
Illuminating discussion!
"One useful question is whether neurons and circuits carry out identifiable transformations on their inputs while representing something relevant to the world they inhabit. At the level of single neurons, local circuits, neural populations, and large-scale networks, the answer is almost certainly yes."
I am curious. Are there folks in the neuroscience community who would disagree that neurons and circuits carry out identifiable transformations and their functional organization not only correlates with aspects of the external world but has downstream causal relevance? If so, what explains the disagreements? And if not, then what's the whole "representation" theoretical debate really about? It seems to me that once we concede that functional organization captures some property of the external world and is causally relevant, debating whether this is "representation" or not seems rather superflous.
I agree with you. The debate is mostly within the philosophy of mind/science, and I’m not sure it has much impact at the level of systems or cognitive neuroscience. The working assumption that circuits carry out identifiable transformations with downstream causal relevance is pretty much the foundation the field operates on. The actual debates are around how causality is established between particular circuit configurations and behavioral outcomes, and to what degree of specificity. Naturally, the fly community have much better tools to do so!
Michael, lovely piece and unusually honest. “What Might Cognition Be, If Not Computation?” is exactly the kind of question cognitive science has been circling for decades, and it still matters. I really liked that you named the “describable by math ≠ computing” trap yourself; that move saves the argument from the cheap version of this debate. And the fly example is beautiful, one of those rare cases where the levels really seem to close down to the wiring. “The synaptic wiring is the computation” is a great line.
Two thoughts/questions.
First, the definition of computation still feels broad enough that I start wondering about the pancreas: it also transforms inputs into outputs according to rules over variables that matter to the organism. So what exactly disqualifies it? I’m sure you have an answer, but making the cutoff explicit would strengthen the conclusion.
Second, regardinig representation: the fan-shaped body lesion indicates that the signal is necessary for function, but necessity alone may not be sufficient. A power supply is necessary for navigation too, and it doesn’t represent heading. The harder comparison might be van Gelder’s Watt governor rather than the pendulum: a system that causally organizes downstream behavior in a content-sensitive way yet which many would still resist calling representational.
The thing I’d most like to see you chase is that your strongest examples, the ring attractor and the integrator, are not merely input-output mappings. They hold state: the bump persists in darkness, the eye stays put. That sounds like memory/maintenance or more generally, state dynamics, rather than just an input >> output transformation. So maybe the intro framing slightly undersells your own best cases.
Looking forward to Part Two. I’m curious whether predictive processing will come in as support for the thesis or as a competing way of telling the story :)
Great points! I don’t object to the pancreas performing computations but I think the spatiotemporal scales may not be optimized for what we consider ones relevant for thinking. I think time makes a big difference here, but will need to think about that. Thanks so much for this terrific input/pushback. It’s great
a beautifully polite takedown, if you ask me 😁
Can you tell if a car has a driver only by looking outside its doors?
Interesting article Michael. I think any brief that gets us to thoroughly explore a phenomena and opens our mind to new ways of looking at it, potentially tying together activity we may have previously overlooked, is a useful tool.
However, I worry that the computer metaphor has outlived its usefulness and may now be giving rise to misconceptions. I wonder whether it isn't worth reaching for a fresh metaphor that better reflects the situation and doesn't lead to the misnomer we're similar to AI.
An example of the problems with metaphors is that when you find that parts of the brain involved in navigating spatially may also be involved in navigating concepts and memories, precedence leads us to imagine we explore ideas in a spatial way, when if we'd made these discoveries the other way round we may imagine we explore spaces as ideas. Do you see what I mean?
Well, I think it's hard to come at it from the other direction because we tend to start from physics. For example, kids (and non-human animals) have a very good sense of intuitive physics well before they start developing metacognitive capacity. So I don't know, just my immediate reaction.
Yes I didn't mean it literally. Merely that when we apply a metaphor to something this tends to be how we frame it. In this case what is going on maybe more primitive and biological than spatial mapping (hence why it is implicated in other activities). So it might be more fundamental abstract mapping of some other kind. I'm not saying it is. I'm just talking hypothetically. So if we thought it was doing spatial mapping we'd assume not to bother testing if it's involved in something like taste. When actually, if our thinking is more agile, and it's more loosely to do with abstract mapping in general, we might think to test for this.
Again, I don't mean this literally. You're in a much better position to know the particulars of what you're studying. I'm merely expressing a concern at how neuroscience might be riddled with metaphors based on the precedence of discoveries and what was first attributed to them, as well as the overriding metaphor of a computer, and the effort to make that metaphor fit. Such artefacts might be cluttering the intellectual landscape somewhat.
It's certainly a wonderful puzzle and articles like yours only make it more intriguing. Thanks 😊👍
You bring up very good points. I def agree with their spirit and yes, we are riddled with metaphors. I hope the post didn’t do that!
Haha no it didn't. It just gave me an excuse to bring this point up, seeing as you were asking about the computer metaphor 😊
This is a thoughtful, serious article. I was glad to read it.
I need models for the brain. I don’t know how else anyone can even begin. The brain is too much. Too many cells, too many pathways, too many things happening at the same time. So we reach for something smaller. A computer. A prediction machine. A network. Something we can hold in our hands, at least for a minute.
And the better the model is, the more dangerous it gets.
Because after a while, it starts to feel like we understand the thing itself. Not just the model. The brain. We read a careful explanation, follow the parts, see how one system talks to another, and some part of us says, “Ah. Now I get it.”
I don’t think we do.
I have read a ridiculous amount about the brain over the years. Papers, studies, books, theories, all of it. And the more I read, the less comfortable I am pretending we understand very much about how the brain actually works. We know pieces. Some pieces are astonishing. But still, pieces.
That was the thought I had while reading this article. It gives the reader a real feeling of understanding. Not in a cheap way. It is careful and deep. But still, I kept thinking about what was not in the frame.
Neurotransmitters, for one.
The whole article can move through an impressive explanation of the brain as a kind of computational system, and somehow the chemical mess of the brain is mostly off to the side. That is not a small omission. We don’t even know if we have identified all the neurotransmitters. And we certainly do not know, in any complete way, what each one does in each part of the brain, in each state, in each person, at each moment.
That is where psychiatry gets humbled very quickly.
A person takes a medication. It helps. Or it does nothing. Or it helps the thing it was supposed to help, but now sleep is strange, or desire changes, or the body feels wrong, or a side effect appears that nobody quite predicted. Then we adjust, switch, add, subtract, wait, guess again.
That is not because psychiatry is useless. It is because the brain is not giving up its secrets that easily.
So yes, I am grateful for models like the one in this article. We need them. Without them, we are staring into fog. But I also don’t want to mistake the flashlight for daylight.
Thanks: I definitely agree with your high-level point about confusing bits and pieces with a grand theory. On the neurotransmitter bit, they are included in the models. I haven’t discussed them explicitly but I don’t know that doing so would provide more conceptual clarity. Regardless, your points are well taken.
“Describing the brain computationally also lets us see deep links between things that otherwise appear unrelated. Who would have guessed that the machinery we use to navigate physical space might also help organize knowledge more generally?”
Interestingly, I interviewed someone for my book who has aphantasia, and this is exactly how he remembers things, not in images, but spatially, like rooms in a building.
Thanks for sharing Trysa. This is cool. I'm very curious to learn more!
I think more accurately, the computer is modeled after the brain or at least takes its que from it.