And you don’t even aim for a good definition? For what do you aim then? … I think if I’m doing a priori armchair reasoning on socks, the way you and I do armchair reasoning here, I’m pretty much constrained to conceptual analysis. Which is activity of finding necessary and sufficient conditions for a concept.
The goal of this series is to explain how certain observable facts about the physical universe arise from more basic principles of physics, neuroscience, algorithms, etc. See §1.6.
I’m not sure what you mean by “armchair reasoning”. When Einstein invented the theory of General Relativity, was he doing “armchair reasoning”? Well, yes in the sense that he was reasoning, and for all I know he was literally sitting in an armchair while doing it. :) But what he was doing was not “constrained to conceptual analysis”, right?
As a more specific example, one thing that happens a lot in this series is: I describe some algorithm, and then I talk about what happens when you run that algorithm. Those things that the algorithm winds up doing are often not immediately obvious just from looking at the algorithm pseudocode by itself. But they make sense once you spend some time thinking it through. This is the kind of activity that people frequently do in algorithms classes, and it overlaps with math, and I don’t think of it as being related to philosophy or “conceptual analysis” or “a priori armchair reasoning”.
In this case, the algorithm in question happens to be implemented by neurons and synapses in the human brain (I claim). And thus by understanding the algorithm and what it does when you run it, we wind up with new insights into human behavior and beliefs.
Does that help?
are you even disagreeing with me on this map example here
Yes I am disagreeing. If there’s a perfect map of London made by an astronomically-unlikely coincidence, and someone asks whether it’s a “representation” of London, then your answer is “definitely no” and my answer is “Maybe? I dunno. I don’t understand what you’re asking. Can you please taboo the word ‘representation’ and ask it again?” :-P
Einstein (scientists in general) tried to explain empirical observations. The point of conceptual analysis, in contrast, is to analyze general concepts, to answer “What is X?” questions, where X is a general term. I thought your post fit more in the conceptual analysis direction rather than in an empirical one, since it seems focused on concepts rather than observations.
One way to distinguish the two is by what they consider counterexamples. In science, a counterexample is an observation which contradicts a prediction of the proposed explanation. In conceptual analysis, a counterexample is a thought experiment (like a Gettier case or the string example above) to which the proposed definition (the definiens) intuitively applies but the defined term (the definiendum) doesn’t, or the other way round.
The algorithm analysis method arguably doesn’t really fit here, since it requires access to the algorithm, which isn’t available in case of the brain. (Unless I misunderstood the method and it treats algorithms actually as black boxes while only looking at input/output examples. But then it wouldn’t be so different from conceptual analysis, where a thought experiment is the input, and an intuitive application of a term the output.)
are you even disagreeing with me on this map example here
Yes I am disagreeing. If there’s a perfect map of London made by an astronomically-unlikely coincidence, and someone asks whether it’s a “representation” of London, then your answer is “definitely no” and my answer is “Maybe? I dunno. I don’t understand your question.
But I assume you do agree that random strings don’t refer to anyone, that clouds don’t represent anyone they accidentally resemble, that a fist by itself doesn’t represent anything etc. An accidentally created map seems to be the same kind of case, just vastly less likely. So treating them differently doesn’t seem very coherent.
Can you please taboo the word ‘representation’ and ask it again?” :-P
Well… That’s hardly possible when analysing the concept of representation, since this is just the meaning of the word “represents”. Of course nobody is forcing you to do it when you find it pointless, which is okay.
Of course nobody is forcing you to do it when you find it pointless, which is okay.
Yup! :) :)
The algorithm analysis method arguably doesn’t really fit here, since it requires access to the algorithm, which isn’t available in case of the brain.
Oh I have lots and lots of opinions about what algorithms are running in the brain. See my many dozens of blog posts about neuroscience. Post 1 has some of the core pieces: I think there’s a predictive (a.k.a. self-supervised) learning algorithm, that the trained model (a.k.a. generative model space) for that learning algorithm winds up stored in the cortex, and that the generative model space is continually queried in real time by a process that amounts to probabilistic inference. Those are the most basic things, but there’s a ton of other bits and pieces that I introduce throughout the series as needed, things like how “valence” fits into that algorithm, how “valence” is updated by supervised learning and temporal difference learning, how interoception fits into that algorithm, how certain innate brainstem reactions fit into that algorithm, how various types of attention fit into that algorithm … on and on.
Of course, you don’t have to agree! There is never a neuroscience consensus. Some of my opinions about brain algorithms are close to neuroscience consensus, others much less so. But if I make some claim about brain algorithms that seems false, you’re welcome to question it, and I can explain why I believe it. :)
…Or separately, if you’re suggesting that the only way to learn about what an algorithm will do when you run it, is to actually run it on an actual computer, then I strongly disagree. It’s perfectly possible to just write down pseudocode, think for a bit, and conclude non-obvious things about what that pseudocode would do if you were to run it. Smart people can reach consensus on those kinds of questions, without ever running the code. It’s basically math—not so different from the fact that mathematicians are perfectly capable of reaching consensus about math claims without relying on the computer-verified formal proofs as ground truth. Right?
As an example, “the locker problem” is basically describing an algorithm, and asking what happens when you run that algorithm. That question is readily solvable without running any code on a computer, and indeed it would be perfectly reasonable to find that problem on a math test where you don’t even have computer access. Does that help? Or sorry if I’m misunderstanding your point.
The goal of this series is to explain how certain observable facts about the physical universe arise from more basic principles of physics, neuroscience, algorithms, etc. See §1.6.
I’m not sure what you mean by “armchair reasoning”. When Einstein invented the theory of General Relativity, was he doing “armchair reasoning”? Well, yes in the sense that he was reasoning, and for all I know he was literally sitting in an armchair while doing it. :) But what he was doing was not “constrained to conceptual analysis”, right?
As a more specific example, one thing that happens a lot in this series is: I describe some algorithm, and then I talk about what happens when you run that algorithm. Those things that the algorithm winds up doing are often not immediately obvious just from looking at the algorithm pseudocode by itself. But they make sense once you spend some time thinking it through. This is the kind of activity that people frequently do in algorithms classes, and it overlaps with math, and I don’t think of it as being related to philosophy or “conceptual analysis” or “a priori armchair reasoning”.
In this case, the algorithm in question happens to be implemented by neurons and synapses in the human brain (I claim). And thus by understanding the algorithm and what it does when you run it, we wind up with new insights into human behavior and beliefs.
Does that help?
Yes I am disagreeing. If there’s a perfect map of London made by an astronomically-unlikely coincidence, and someone asks whether it’s a “representation” of London, then your answer is “definitely no” and my answer is “Maybe? I dunno. I don’t understand what you’re asking. Can you please taboo the word ‘representation’ and ask it again?” :-P
Einstein (scientists in general) tried to explain empirical observations. The point of conceptual analysis, in contrast, is to analyze general concepts, to answer “What is X?” questions, where X is a general term. I thought your post fit more in the conceptual analysis direction rather than in an empirical one, since it seems focused on concepts rather than observations.
One way to distinguish the two is by what they consider counterexamples. In science, a counterexample is an observation which contradicts a prediction of the proposed explanation. In conceptual analysis, a counterexample is a thought experiment (like a Gettier case or the string example above) to which the proposed definition (the definiens) intuitively applies but the defined term (the definiendum) doesn’t, or the other way round.
The algorithm analysis method arguably doesn’t really fit here, since it requires access to the algorithm, which isn’t available in case of the brain. (Unless I misunderstood the method and it treats algorithms actually as black boxes while only looking at input/output examples. But then it wouldn’t be so different from conceptual analysis, where a thought experiment is the input, and an intuitive application of a term the output.)
But I assume you do agree that random strings don’t refer to anyone, that clouds don’t represent anyone they accidentally resemble, that a fist by itself doesn’t represent anything etc. An accidentally created map seems to be the same kind of case, just vastly less likely. So treating them differently doesn’t seem very coherent.
Well… That’s hardly possible when analysing the concept of representation, since this is just the meaning of the word “represents”. Of course nobody is forcing you to do it when you find it pointless, which is okay.
Yup! :) :)
Oh I have lots and lots of opinions about what algorithms are running in the brain. See my many dozens of blog posts about neuroscience. Post 1 has some of the core pieces: I think there’s a predictive (a.k.a. self-supervised) learning algorithm, that the trained model (a.k.a. generative model space) for that learning algorithm winds up stored in the cortex, and that the generative model space is continually queried in real time by a process that amounts to probabilistic inference. Those are the most basic things, but there’s a ton of other bits and pieces that I introduce throughout the series as needed, things like how “valence” fits into that algorithm, how “valence” is updated by supervised learning and temporal difference learning, how interoception fits into that algorithm, how certain innate brainstem reactions fit into that algorithm, how various types of attention fit into that algorithm … on and on.
Of course, you don’t have to agree! There is never a neuroscience consensus. Some of my opinions about brain algorithms are close to neuroscience consensus, others much less so. But if I make some claim about brain algorithms that seems false, you’re welcome to question it, and I can explain why I believe it. :)
…Or separately, if you’re suggesting that the only way to learn about what an algorithm will do when you run it, is to actually run it on an actual computer, then I strongly disagree. It’s perfectly possible to just write down pseudocode, think for a bit, and conclude non-obvious things about what that pseudocode would do if you were to run it. Smart people can reach consensus on those kinds of questions, without ever running the code. It’s basically math—not so different from the fact that mathematicians are perfectly capable of reaching consensus about math claims without relying on the computer-verified formal proofs as ground truth. Right?
As an example, “the locker problem” is basically describing an algorithm, and asking what happens when you run that algorithm. That question is readily solvable without running any code on a computer, and indeed it would be perfectly reasonable to find that problem on a math test where you don’t even have computer access. Does that help? Or sorry if I’m misunderstanding your point.