135 - Jake Quilty-Dunn: The Language of Thought Hypothesis in Cognitive Science

Welcome back to the Stanford Psychology podcast, my name is Joseph and my special guest today is Professor Jake Hultee Dunn. Jake is an assistant professor in the Department of Philosophy and the Center for Cognitive Science at Rutgers University. Jake's research is primarily in philosophy of mind and cognitive science, but he also likes to think about a broad variety of topics including philosophy of language, aesthetics, and early modern philosophy. Today we shall be talking about the language of thought, but first Jake, thank you so much for joining me for this conversation today. Thanks for having me, Joseph. Do you do any empirical research projects or you collect data or are they more like philosophy like where you just write articles? Yeah, I'm definitely primarily just writing theoretical articles. I have collaborated a bit with psychologists, I don't actually have like real experimental skills, so I can't like program or do any of that stuff, but I have worked with psychologists, I've collaborated with Tatiana Emanuel and Michael Epstein, Eric Mandelbaum and I before of us published the paper in the Journal of Experimental Psychology, Human Perception and Performance on how the visual system derives statistical summaries of stimuli, and I also did a postdoc that was a kind of interdisciplinary collaboration that was led by Nick Shea as a philosopher, an empirically mighty philosopher like me, and James Hampton as a cognitive psychologist, and on that project also with Sifira Thorn and Yulia Smoorchkova, we published a couple papers investigating how people kind of metacognitively keep track of their concepts, but other than that, I mean that those are just a small handful of articles that pass majority of my work is kind of more pure philosophy. My hot take is I think even writing a theory paper is empirical in a way, I think it philosophy that some people think of them as thought experiments where you are, you're literally imagining what an experiment would look like, and I think even when it comes to thinking in general, you could sort of somehow start thinking and arrive at a conclusion without gathering any data from the world, which is actually a lovely segue into the language of thought and realizing. So there seems to be a whole buzz currently about this idea, a lot of which I arguably will attribute to a BBS paper that you wrote with Nicholas Porot and Eric Mandelbaum, and it's called The Best Game in Town, and yeah, so let's get right into this. Really briefly, would you like to talk about what the language of thought is about, like is it a kind of theory? What is it trying to describe? Yeah, I guess I would think of it as a hypothesis, it's kind of the language of thought idea in general is to sort of abstract to really constitute a full-blown theory, but it's a kind of general hypothesis about the most basic structural features of human thought, and potentially non-human animal thought. And the idea, as I understand it, it's been defended in very different ways by different theorists, but the kind of core idea that brings all of these theorists together is the idea that our thoughts, when you take a thought like, the thought that it's raining outside right now, that thought breaks down into little bits or constituents, and that those constituents, each correspond to a particular element of that, the content of that thought, like you'll have an element that represents rain, you have an element that represents the current time when you say right now, and that you take those constituents and you combine them into a structured whole and the content of that structured whole depends upon the most basic parts and the way in which they're put together. Now that basic idea, many have argued, so for example, my colleague Liz Camp at Rutgers is a philosopher who's interested in various kinds of representational formats, and she would argue, well, that basic idea is actually compatible with a bunch of different formats. So maps, for example, might have that same property of having little bits that combine in structured holes that depend upon the identity of the bits and the way in which they're combined. So you need more, I think, to get from that core idea of taking bits and combining them to get to the idea of a language of thought as opposed to other kind of rival representational formats, but I do think that when you look at all the different theorists who've defended a language of thought, that's the kind of core thing that they share, is the recombinability of little bits. Wonderful. First of all, it's interesting to me that the language of thought has the word thought attached to it, which makes me think that it's describing the stuff that happens after perception, but before action, I think some philosophers have described this as central cognition, or maybe people just think of it as thinking in general. And it makes sense that the theory is saying that when we think, perhaps, that's formatted in a kind of language-like way. I wonder whether this, I guess, this first way of thinking about it is pretty commonly understood? Yeah, I would say it's definitely, it would be worrisome if central cognition didn't have any language of thought-like structures that would significantly undermine my confidence in the language of thought hypothesis. However, I don't see the view as essentially restricted to kind of higher level thinking or central cognition of whatever you want to call it, reasoning. So for example, in our paper, one of the main areas where we go into the literature in a detailed way and find evidence for structural properties of a language of thought is visual perception. And so I think even within visual perception, we find that our visual system is constructing language of thought representations. And part of the usefulness of our visual systems doing this is that they can then hand that language of thought representation off to higher level cognitive systems, which can just sort of immediately use it because that's a format that is familiar and native to higher level reasoning. I guess one question is, what exactly counts as thought, if we are restricting it to central cognition and higher level reasoning, then I would want to say the language of thought hypothesis is wider than that. You might view perception as a kind of thought or cognition, and so far as it's, you know, smart, complicated computations that are going on. And so maybe the language of thought is limited to thought in that very broad sense that would capture, you know, almost all of the mind. But if we're using thought in a more restrictive sense, then I would want to say that the language of thought should not be viewed as restricted to just central cognition. Okay, wonderful. The other thing I'd like to also highlight is, language of thought sounds a lot like natural language. So I feel like a lot of people think, you know, when they're thinking, you can sort of hear some words and sentences in your head. And so the theory that our language is thought like, our thought is language like, sorry, it's a claim sort of endorsing this idea that when you're thinking, you sort of are using your language faculties to make sense of the world. It was kind of surprising me to me to learn that there might be a distinction between natural language and language of thought. Do you want to really quickly talk about that distinction if at all you think it exists? Because it would also, I guess it has consequences for whether we think pre-linguistic infants have a language of thought or even animals and so on. Yeah, I don't know that I can talk about it quickly though, because I think the answer to that question, I need to get a little bit into the historical roots of the language of thought. Is that alright if I do that? Yes, please. So I mean, some people would argue that the language of thought I thought this is, you can find roots in the medieval period and blame with Acom famous for, you know, Acom's razor and so on. But to kind of focus on the more contemporary history, Jerry Foder is a philosopher who wrote a book in 1975 called The Language of Thought and a lot of people when they hear that phrase, the language of thought, their minds go right to Foder. And Foder certainly could not have identified the language of thought with the kind of internal use of your spoken natural language or what you might call inner speech, which I think is what kind of what you're referring to before. Like when I think about what I'm going to have for dinner or engage in some kind of, you know, ordinary high-level reasoning like that, I might hear my voice or hear a voice in my head, articulating particular sentences. And that is a phenomenon that we know occurs and it has some utility in facilitating reasoning. There are a bunch of reasons why I don't think we should identify the language of thought idea with the phenomenon of inner speech. But for Foder, that would have been a total non-starter and that's because for Foder, the main reason we want to pause at a language of thought is to explain the interface between natural language and the mind. And so for Foder, the only way that you could learn a natural language is if you already had a language of thought that you could use to map external sounds to concepts in the language of thought. So for Foder, when you learn your lexicon and you map the sound dog to some meaning, what you're doing is pairing that sound to a word in the language of thought. And when you parse a sentence like the dog is barking, you have the concepts dog in barking and you put them together in the right way. And then that representation in the language of thought interfaces directly with your language faculty. And that's what it is to grasp the meaning of the sentence of the dog is barking. And for Foder, if there isn't this kind of separate code in the mind that the language faculty is interacting with, then we couldn't explain how it is that we come to learn our language in the first place. It's almost like we have to be prepared in some way to learn language by being able to parse the world into these discrete structures. This is great because this leads to sort of the next part of the conversation which is actually about the history of the language of thought. So in this paper, you write about how people might think that the language of thought is this old theory, you know, similar to skinnarian behaviorism or Freudian id ego and super ego. And I wonder why you think that's the case. Is this because of sort of the rise of like deep learning methods and in large language models where people think it's something sub symbolic going on or other other reasons to think that sort of like Foderian language of thought had sort of long been forgotten until recently? Yeah, I mean, I think in cognitive science, when the argument between different views sort of degrades a common sort of epithet that people will attach to each other's theories is outdated. So this is outdated. You believe in the language of thought that's an old school idea. And I think there have been a bunch of different kind of alternative perspectives that have risen in the, I guess it's just about 50 years since a photo published a language of thought. And so you mentioned deep learning, that's one that's particularly been particularly influential in the past 10 to 12 years. Before that, there were a lot of approaches like embodied cognition and the original rise of connectionism in the 80s and 90s. And I guess broadly you could refer to for E cognition, you know, approaches related to embodied cognition, but also an activist and extended and so on. And so I think given that there are all these other kinds of approaches and then also looking at a lot of cognitive neuroscience, you see descriptions of representations at the neural level and computations appearing at the neural level, which don't always look like they have the properties of a language of thought and perhaps rarely look like they do. And so I think the sense that a lot of people have in cognitive science is that the language of thought idea is really tied to this kind of approach to cognitive science that was one of the first ways of kind of the birth of cognitive science out of behaviorism. I think people think of Tromsky and the rise of generative linguistics. I think they also think of what sometimes referred to as go-fi, good old fashioned artificial intelligence. So symbolic models of artificial intelligence. If you look at what AI people were up to in the 70s, it did look very much like a language of thought model. And so I think there's a bunch of different approaches that have arisen in the decades since Porter originally defended the language of thought. And that's part of the reason why we titled our paper The Best Game in Town because Porter back in the day was fond of making the claim that the language of thought is the only game in town reading. It's the only kind of general approach to mental representation that does any explanatory work that seems to have any hope of accounting for structured human behavior. And in 2024, I think it's harder to make that claim that we just don't even have, we don't even have a single alternative that even begins to explain anything. I don't think that that's correct. And so instead our claim is while there might be lots of processes in the mind that are captured by connectionists, networks, and other kinds of representational features, there are also quite a few that we need a language of thought structure to explain. Okay, this is wonderful because it's not making me think about just the alternative representation formats. We haven't actually explicitly talked about this. I guess you've mentioned some along the way you've talked about maps and you've talked about images. So just so we're clear, let's talk about like what are the different possibilities of different formats that might exist both within perception and cognition. Just in order for it to make sense that that language of thought is the best game, the best is opposed to what? Yeah, I mean, just to be clear, I do think that not only are there other formats that could be rival hypotheses that we think the language of thought is superior to. But in some domains of cognition, there are other formats that I don't want to speak for my co-authors, Nick Perot and Eric Mandelbaum, like we all have exactly the same view about this, but I think we broadly agree. There are other formats in the mind, in fact. So just to take one example, in some visual processing, we argue that there are language of thought structures, but we don't think that that captures all of visual processing. And so like any, we have a section of our paper that concerns perception, but a vision scientist that read that section of the paper would notice that there's quite a few visual phenomena that we're not discussing. And for some of those visual phenomena, I think other formats are more explanatoryly successful, including, for example, iconic representational formats, or image-like or picture-like representational formats. And that's also something that we see in the literature on mental imagery, where some aspects of mental imagery seem like they might be best explained by an iconic representational format that doesn't have the features of a language of thought. Another kind of format that I think can do some explanatory work in the mind is an associative structure, where you just learn an association between doctor and nurse, for example. So take the phenomenon of semantic priming. One of the core effects of cognitive psychology that has successfully survived the replication crisis is semantic priming. You read the word doctor then, you know, a couple hundred milliseconds later, you're quicker at reading the word nurse. Whatever structure is present in memory that explains semantic priming might be purely associative. There might be some forms of semantic priming that are not relying on associative structures. So, for example, when you read the word sister and that primes the word cousin, that might not be because of a pure associative structure between sister and cousin. And perhaps you don't even hear those words. There hasn't been much associative continuity between those two words in your environment in the past. That might rely on a more hierarchical structured representation that ties them both together as instances of the more general category family member or something like that. And there's even some evidence to think that there are different structures here in memory despite they're both being tested by semantic priming. So, for example, all timeers patients often lose the more structured form of priming. So, that they won't show priming from sister to cousin, but they will still show priming from cat to dog or salt to pepper. So, the associative priming is still there, but whenever neural damage is being caused by Alzheimer's disease is disrupting the more structured form of representation. So, iconic representations, associative structures, maplich representations and navigation, although there's a really fascinating paper that recently came out in Trends and Cognitive Sciences by Nina Kazanina and David Popol to people working in linguistics that argue that actually if you look at cognitive maps in the hippocampus, you see core properties of the language of thought and they're thinking about the language of thought in very similar ways that Nick and Eric and I are in our BBS paper. So, in all these cases, I think we have a fairly clear idea of what other formats will look like. And in some cases, the language of thought is the best explanation and in other cases, iconic formats or associative structures or maybe mental maps or maybe analogue magnitudes in some aspects of the representation of number. For example, something you see in Susan Carey's work, for example. In other cases, one of these for mental do a better job. It's all got to be, we've got to ask very specific empirical questions, I think, about particular kinds of cognitive systems. Good. So, it almost sounds like at least from the paper that you're sort of like representational pluralists, if you will, that there are different formats and sometimes some formats do the best job and sometimes it's others. And okay, one quick question is about whether you think that each format is orthogonal to the other or in some sense, like do you think like an iconic representation can simultaneously be formatted in a language like way or or not? And I guess this maybe goes into like the properties of them because I was thinking that some representations might also be discrete and have some properties that we think language of thought has, but maybe they don't have others. And so how we to make of them versus we just think like if something is iconic, it just is entirely different from like if it's language like versus with map like. So, so yeah, I guess that's like one question I posed and as I was posing it, I just realized there's a whole conversation to be had about the properties of them. Sure, I guess I'll just say without getting too into detail about the specific properties of language of thought, just kind of very generally as a methodological point, do we want to pause it representations that can sort of be a member of multiple formats at once? I guess would be the question. And I think as I'm thinking about it, there's going to be a fact about which particular formative representation is. And so if it's an icon, then it's not a language of thought representation. However, that said, I don't want to rule out that you might have a representation that has let's let's take for example, an analog magnitude representation of a particular color like red. I don't want to rule out that you might be able to take that representation and combine it with other representations of different formats, maybe into a larger structure that could kind of have multiple formats as its constituents. I don't have an a prior argument for why that couldn't happen. So for example, and some philosophers like to talk about cases like this, you might think that shirt is and then instead of using your concept red, you could have an analog magnitude representation of the particular color red. And maybe you could think a thought like that shirt is red, where the red constituent is actually just an analog magnitude. And maybe there are lots of cases like this where we think thoughts that have hybrid kinds of constituents. So that's a way in which is that representational language of thought representation? Well, maybe in respect of the compositional principles that are being used to take bits and put them together into a more complex hole, the answer would be yes. But maybe in respect of the particular constituents that are being put together, the answer is more complicated. So I don't, again, I don't think we have a priori reasons to roll this out. And this is part of why we were taking the core claim and maybe this can segue us to talking about the properties of a language of thought. Our claim was you will tend to find that if you've got one language of thought property, you've got this other bunch too. And so these properties are to cluster together. Wonderful. So yeah, let's talk about the properties. I'm not sure whether you think of them as like as a static list, you know, it's these six or five or so properties or whether, you know, it's subject to change. What are the core properties that would identify a language of thought? Yeah. And this is where we get into things that theorists who are sympathetic to a language of thought don't necessarily agree with each other or say the exact same thing. But what we were trying to do in writing this paper was think about what are what properties seem to really get at the core idea that's shared by people like photo and also a chomsky, volition, mandigalistel, Susan Carrey, and many others. And our thought was it's that you've got little bits and that you put them together to get a structured hole, but also which little bits matters and what kinds of structured holes you're putting them together into matters. And so that's what we were trying to capture with our list of six properties. So the properties are number one, discrete constituents. And that just means that you've got different little bits rather than a kind of holistic representation that kind of combines the two into one, but without having to discrete things that are getting combined. The second property is role-filler independence. And that property is if you take a representation like John Loves Mary, that has a verb loves, which is a two-place predicate. And so there's an agent and a patient. There's something doing the loving and there's something being loved. And if you swap John and Mary, then you get Mary loves John. Now those roles of agent and patent are exactly the same in those two representations, but the which representation fills those roles changes. So John Loves Mary and Mary loves John. There's an independence of the role of being agent or patient and then there's also an independence of the filler. So John is the same John when he's the one being loved as when he's the one loving the other person. That I think is a key property of language-like representations that you don't see cropping up in many other representational formats. The third is predicate argument structure, which is the claim that you have argument structure of the kind that I just mentioned as an example for a Loves where you have a two-place argument structure. And more generally the claim is when you represent a fact like that thing is brown, which you might think that that's something you could encode in a bunch of representational formats. You could also draw a picture of the thing being brown, right? And maybe that image would encode that thing is brown. But in that image you won't have a separate representation for the thing and another representation for the property where there's a predicate argument structure explicit in that image. The image kind of just represents that information as a package without... So in a way here you have a claim that there are discrete constituents but more specifically one constituent corresponds to a predicate and another constituent corresponds to an argument. And that I think is a key to the idea that you have a little sentence. The claim is really at bottom that you've got a representation of a predicate and you're applying that predicate to a particular argument or set of arguments. And then the fourth property is the presence of logical operators. And this one, the idea is that in the literature on representational format one kind of key role for language of thought architectures is that they're supposed to explain the way in which thoughts seem to have logical form. And logical form is something that would include features that I've just mentioned like predicate argument structure. But it also includes the fact that you can represent that there's a bird outside my window but you could also represent there's not a bird outside my window. So it's not true that there's a bird outside my window. So you can take the very same thought and then negate it so you can add in this logical operator that applies to that representation and says no it's not true. That's very hard to do in other representational formats. That's hard to do in an image for example. Right you can draw an image of the scene outside your window and fail to add a bird into that scene. But that doesn't express exactly the content that there's not a bird outside my window. Someone you'd need to do some interpreting of the image to get to that particular representation. Whereas in a language like format you can just say that explicitly. You could have a logical operator and combine it in that way. The fifth property is inferential promiscuity which says that if you have a language of thought then you ought to be able to do some general reasoning with it. And in particular you ought to be able to do reasoning that sensitive to logical form. Sensitive to predicate argument structure sensitive to logical operators. And the final property is abstract conceptual content. And that gets to as I said you have to say in order to defend the language of thought hypothesis you don't just want to say that there are bits that combine you want to say something about the bits being language like it's something about the modes of combination and being language like. And when we're talking about the bits being language like the idea of abstract conceptual content is that the primitive representations in your language of thought don't just represent perceptible features like red and you know shape properties and size and so on. But also can represent more abstract categorical properties like the property of being a bottle or the property of being a giraffe. And in a language like representation you can have a word like giraffe that just corresponds to that property of being that type of animal. And that representation need not carry any information about the particular shape or color or or anything else. It just represents giraffe. So those are the six properties that we argue will tend to cluster together. So maybe none of them is necessary or sufficient for being a language of thought. But if there's a natural what philosophers would call a natural kind. And if we're if we're really getting the structure of the world right and saying there's a particular kind of representational format in animal cognition then we ought to see these whenever you have one or two of these properties should tend to have the others. Start to that very long model. No this was absolutely wonderful realizing now that it would probably have been a good idea to talk about them at the beginning but all the same. So I actually I have a few questions about some of these properties and I don't want to take too much of our time. Okay so one of them is about the role-filler independence property. So the roles here we're talking about seem to be like syntactic roles. So a role like an agent or a patient like in a sentence you could you could imagine a subject being an agent and an object being a patient and so on. And I'm wondering whether whether the roles are restricted to these kinds of roles like the role of a word in a sentence or it's sort of the more broad interpretation that I have of roles which is like a variable anything that can take on different values. So I would imagine a representation of something like animal that could take the value of like cat or dog or elephant or giraffe and so it's almost like it's a container for a lot of different entities inside of it. Yeah so the current thing you're just talking about variable binding I would think that as being a very good example of role-filler independence. If what's going on is really in the computational architecture if it's really variable binding right so if you really have a representation of the variable animal and then you have a representation of a particular zebra and then there's some operation that's performed by the system that binds the representation of this particular zebra in front of me to this variable animal. Then I would end that it's the very same variable when I bind the particular giraffe to it and it's the very same representation of this particular zebra if I'm binding it to animal or if I'm binding it to a striped thing or something else. That's more demanding I think than just saying I wouldn't want to say that anytime you can represent two different animals as being an animal that therefore you necessarily have role-filler independence right because like you could be an opponent to the language of thought hypothesis and not want to buy into this role-filler independence idea but still think that you can you know represent things as being an animal. You just want to you would not posit that kind of explicit variable binding structure to explain how it is that people are capable of categorizing things as animals for example. Interesting okay so I'll just do this more rapid fire now okay so truth of value ability so I think the property here is predicate argument structure and I think you end the sentence and say something like they're almost like propositions that have truth value and I wonder whether that's specific to a language of thought or whether you could say that if I have an iconic representation and a visual representation of a house it's possible to ask whether or not it's true or it's actually what I'm seeing and so it does fulfill that property of being sort of truth valuable and yeah or does it have to be sort of something more like a proposition which feels a bit different from just like oh this is this is just some perception I'm experiencing that I could evaluate for truth. Does that make sense? Yeah it definitely does and I would want to say that just being a valuable for truth is that that says very little about what the represent what the underlying representational form is and I mean I guess some philosophers would be sticklers and say it's awkward to talk about a perception or an image as being true but it may be better to talk about it as being accurate and so they'd say accuracy valuable but I'm not that much of a stickler about it so I'm happy to talk about truth or accuracy and yeah you can say that your visual representation that let's say it's a very low-level visual representation that does not have a language of thought structure it could then the less be accurate or inaccurate it could be right or wrong you know even you could talk about a smoke detector that way right if it goes off and they're smoke then it's true it's accurate if it's going off and there's no smoke then that's inaccurate so the mere presence of this ability to kind of get things right or get things wrong I mean a lot of people would say that that's really the core of the notion of representation is being the capacity to misrepresent which presupposes some notion of truth of valuability or accuracy of valuability so it's not the claim of a predicate argument structure is not just that the representation has some kind of import about what's out there in the world that can be correct or incorrect it's the claim that the manner in which you get to the representation that's correct or incorrect it's true or false is by taking representations of arguments and representations of predicates and combining them in such a way that the predicate applies to the argument and then the that determines that the the whole complex representation is true or accurate if the thing represented by the argument has the property that's represented by the predicate so if I take to use an example from a loan half-free who is a post-account chestnut at Hopkins then if I represent something like the pen is in the cup they argue that in in vision or in visual cognition you kind of automatically can extract an abstract relation like that now you have a representation that's accurate if and only if the pen is in the cup but it's the claim is more than that the claim is that you have a representation of the relation in and that's a two-place relation and so that takes two arguments and then you have a representation of the pen you have a representation of the cup and you take those representations and you combine them with the representation of in so that they occupy those argument places and then you get a representation that's accurate if there's a pen in the cup and inaccurate if the pen is not in the cup and so something that they show for example is that actually people's inaccurate judgments are super informative for the kind of representation that they have there so they have an experiment that I think has just come out in open mind maybe I think that's the journal and it's called a float in the basket or a knife in the cup or something like that and what they show is that if you have people do a visual search task or the pen is in the cup then they'll get false alarms for otherwise very visually dissimilar scenes that happen to instantiate that same relation so if you see a phone in a basket right that looks very different from a knife in a cup but that will in a way that's special to that kind of stimuli that instantiate that same relation it'll it'll initiate a false alarm and they argue that and and we endorse the argument in the BBS paper the reason why you get that false alarm is precisely because you have this predicate in with these two argument places and that predicate satisfies role filler independence such that it's the very same predicate with the two-place structure in that's instantiated for when you see knife in the cup and when you see phone in the basket and that's what causes the false alarm so all of that I think in no way just falls out of the idea that you have a representation in your mind that can be accurate or inaccurate you have a much more precise claim being made about how those what philosophers would call accuracy conditions or truth conditions the conditions under which the representation is accurate or inaccurate how those are constructed and the claim is they're constructed in this specific way using predicates and arguments I love your answers because you're covering a lot of ground that I'm hoping that I have time to cover like talking about empirical findings of forelanguage of thought and perception and in other cognitive domains so this is absolutely wonderful okay so one other question was about sort of the abstract content property I wonder whether it entails any commitments about the origins of these abstract concepts so someone might read it as a sort of more like nativist or maybe a minimal nativist sort of position where because the abstract content does not come from like experience it's something that was unlearned in a kind of like a spelky and way you might say yeah we were we were sort of conspicuously non-committal on the topic in the paper where the language of thought comes from except to say that we don't think it comes from natural language and that's kind of what importantly what separates the language of thought hypothesis from positions that a lot of empiricists might be totally fine with which is somehow you learn natural language and then you are able to reuse that in thought to go back to something that we talked about earlier and we want to say that it's important for the language of thought hypothesis that you can represent these language of thought structures without having the reuse resources from natural language so that's why infants and animals are such important test cases for the language of thought hypothesis but now for those creatures that have language of thought representations and did not get them from natural language where did they get them from that's such a tough question I mean the nativeest position as I understand it is basically a negation of the empiricist position right so the empiricist says this stuff is all learned and the the nativeest says it's not learned now all of that debate hangs on us having a notion of a learning that can because if we don't have a coherent notion of learning then neither of these conditions has been coherently stated and so I think that that's just such a tough question about what exactly counts as learning so something that I have wondered about is can there be learning at a kind of lower level of the system that can generate these new primitives at a higher level of the system and this is something that I've been influenced by my former postdoc supervisor Nick Shay unlike me Nick Shay is very much sympathetic to connectionist approaches and he thinks that well no networks actually obviously they're very good at learning right that's the why they're turning out to be so useful in artificial intelligence and so you know he's sympathetic to the idea at least as I understand him that there can be learning at the kind of connectionist level that gives you new symbols at the symbolic level the language of thought level and so the symbol of the language of thought level might just represent bottle and not represent any particular information about bottles but it might have you might have gotten the ability to have that symbol at the higher psychological level through some deep learning in a neural network like structure now I know a lot of developmental psychologist like you mentioned spelled E perhaps Susan Kerry would not find that story satisfying and that of course brings up this whole issue about how do you can you and if so how get a language of thought symbols out of the operations of a connectionist network at a lower level so this is as you can see there's there are several cans of worms right here and so that's that's why we really didn't get into this in the paper but my own view is I'm very open-minded about this I really do not have a strong view about nativism versus empiricism I do want to deny the format claim there's a version of empiricism that says that your representation of bottle is built up out of a bunch of sensory representations and I do think that the evidence undermines that that view but the question of where we get it from and you know this is a reason why some people don't like the language of thought hypothesis is because that if you if you have just all these primitive words in your language of thought then the question of how you got them in the first place it arises and that's a painful problem to have in cognitive science and does not being able to explain where the representations come from I guess just one last thing on this point you can be a language of thought theorist and deny that there are all of these primitives in the language of thought you could think that there are only a small number of primitives in the language of thought maybe those are those are unlearned but you could think that many representations like bottle and giraffe and so on are constructed out of a smaller number of language of thought representations that are just composed so you could think that bottle for example is a definition the concept of a bottle is a definition for the necessary insufficient conditions of being a bottle that are built up out of little parts I sort of read some of the contemporary Bayesian proponents of the language of thought this way like Stephen P. Entedocie and I also read Randy Gallisnell this way and photo was very strongly opposed to this view and he thought looking at linguistics and cycle linguistics it's very implausible that people even have these definitions in their mind for what the necessary insufficient conditions are for being a bottle it's you know famously extremely hard to spell them out and so you know this was the project of analytic philosophy in the 20th century that's where the word analytic comes from the idea that we can analyze our concepts into these necessary insufficient conditions and what photo thought was that project failed so spectacularly that it should really undermine our confidence that there are these definitions in our minds and so what photo thought is we need to go for the troubling view that we do have a ton of primitives in the language of thought this position he called concept atomism that the concept bottle the concept dog the concept table that these are just atoms in the language of thought that don't decompose into anything simpler and so where do they come from he was led to the conclusion they're all innate so the concept of a table is innate that strikes a lot of people as absurd I don't really have intuitions about what's absurd in this neck of cognitive science because I just see it as this extremely tricky area where every view has to be highly non-intuitive in some respects and so as I said I kind of try to be open-minded about it and I recognize that it's a huge outstanding problem for the theory and I don't really have much to say about it that's fair but yeah thank you thank you for your wonderful thoughts and again realizing they're really running low in time but this has been absolutely wonderful and in fact it's a good thing that we didn't really cover everything because then this is a reason for people to check out the really really awesome BBS paper that the Jake and his colleagues wrote okay so as I was like a way to wrap things up we ask our academics to provide advice to young academics and yeah I wonder from your experiences your philosophy professor most of our listeners are psychology students but also their academics and yeah they were just like to know like what what's your journey been like do you have any things you've learned along the way that are super helpful yeah well I mean I am a young academic so I'm reluctant to give out too much advice like I have it all figured out I very much don't I guess one thing that has been useful for me and this is a bit of my like philosophy privilege so I'm trying to check my philosophy privilege right now but one thing that we have the opportunity to do that I think makes us potentially useful is read super widely because we don't have to run labs and you know master experimental techniques and that's something that I have found just super useful just to like follow your curiosity more than anything else more than just kind of trying to limit yourself to the specific things that you think will help you accomplish this project to get this line on your CV or whatever just to allow yourself to read some you know something written in the 1800s or something that's pretty far outside your area and I think you won't see the connections ahead of time and so trying to gain trying to maximize your efficiency by only reading the things that you know for sure will help your current research projects you can't actually predict where you'll be five or six years from now intellectually and so I think the best thing to do in terms of maximizing your satisfaction and is also equally efficient in terms of maximizing your research output is to just read things that you're curious about and rather than getting locked into the particular research projects that are at the forefront for you in the current moment I guess that would be one thing that I would recommend especially to graduate students. Awesome. Would you say you read more than your rights or vice versa like which of the two do you think you do more? As a grad student I definitely read more than I wrote and then you get into you know there are there are of course real career pressures forcing you to write at this point I am reading more than I write but yeah I guess I go in cycles I definitely when I was opposed that I was doing way more writing than reading and when I was working on this thing I just thought paper I was doing way more reading than writing but yeah for it's a bit like a sign way for me I go into different phases. Cool wonderful all right so it's 3 p.m. which is just on time thank you so so much for this amazing discussion I've learned so much from you and I really really appreciate your time I hope you had a great time chatting with me as well. Thanks so much Joseph who's wonderful to get the chance to talk to you. Thank you so much for listening we would love to hear what you think of this episode or our podcast in general if you have any suggestions for future guests or topics of the podcast please reach us at [email protected] you can also connect to us on twitter at stanfordpsychpod and finally if you enjoyed this podcast please consider living as a review on apple podcasts or Spotify or elsewhere so people can find us.