There are two versions of the language-of-thought hypothesis. The representational language-of-thought hypothesis (r-LOT), introduced by William of Ockham and defended by Quilty-Dunn et al., concerns the structure of mental representation. The computational language-of-thought hypothesis (c-LOT), introduced by Jerry Fodor, concerns computation over mental representations. r-LOT is much weaker than c-LOT and is more widely accepted. I accept the former but reject the latter. As a result, I agree with many of Quilty-Dunn et al.'s conclusions while finding that they have not really defended the most controversial form of LOT.
In more detail: r-LOT (I use “LOT” for both the language and the hypothesis) says roughly that thought involves sententially structured mental representations. At minimum, there are nominal representations (e.g., Biden) and predicative representations (e.g., president) that combine into structured representations (e.g., Biden is president) with propositional content. Structured representations may also involve connectives (e.g., and), quantifiers (e.g., all), operators (e.g., always), and other types familiar from the linguistic case.
r-LOT is not trivially true, but it is plausible and hard to deny. It follows naturally from the claims that (1) people make judgments such as Biden is president, (2) these judgments involve combining nominal and predicative representations (or concepts, in the sense where concepts are mental representations) such as Biden and president, and (3) these representations can be recombined in judgments such as Biden is from Delaware. My sense is that most contemporary cognitive scientists and philosophers of mind accept these fairly weak claims. Importantly, these claims do not have immediate consequences regarding computation or cognitive architecture.
The c-LOT adds to r-LOT the key claim that thought involves computation over these sententially structured representations. The classical version of this hypothesis says that r-LOT representations are the medium through which all cognitive computation takes place. That is, the basic vehicles of representation in the r-LOT system (atomic words in the representational language-of-thought) also serve as the basic vehicles of computation (atomic computational states to which cognitive algorithms apply).
The c-LOT hypothesis was canonically formulated by Jerry Fodor's book The Language of Thought (Reference Fodor1975). Computation plays a central role throughout the book, from the main argument for LOT at the start of chapter 1 (“Computation presupposes a medium of computation: a representational system,” p. 27) to the conclusion (“More exactly: Mental states are relations between organisms and internal representations, and causally interrelated mental states succeed one another according to computational principles which apply formally to the representations,” p. 198). There are other works (e.g., “Propositional Attitudes”) in which Fodor focuses mainly on r-LOT, but computation is central in the canonical statement.
(Related distinctions: Fodor himself [Reference Fodor1980] distinguishes the “representational theory of mind” and the “computational theory of mind” [though neither requires a language-of-thought]. Rescorla [Reference Rescorla and Zalta2017] distinguishes a core version of LOT that involves “representational theory of thought” plus “compositionality of thought” and perhaps “logical structure” [in my terms, a version of r-LOT] from a stronger version that adds “the classical computational theory of mind” [yielding a version of c-LOT, though I understand the computational constraint differently from Rescorla].)
Most work in symbolic artificial intelligence (AI) uses a version of c-LOT. Both involve computation over atomic symbols: Entities that are both representationally atomic and computationally atomic. Atomic symbols have no computationally relevant internal structure (if they did, they would not be computationally atomic). Instead, their internal form is arbitrary.
The most significant opposition to LOT, in the classical-connectionist debate, has been opposition to c-LOT. In most neural network models there are no computationally atomic symbols. Representations are distributed over multiple quasi-neural units. As a result, in these models computation is subsymbolic computation: Computation takes place among units below the level of representation. Because computational primitives (units) are not representational primitives in these models, representation is not the medium of computation. Subsymbolic computation is incompatible with c-LOT.
At the same time, subsymbolic computation is quite compatible with r-LOT. This is clearest in the work of structured connectionists (e.g., Chalmers, Reference Chalmers1990; Smolensky, Reference Smolensky1988), where distributed representations (e.g., of Biden and president) can combine with each other systematically to yield new distributed representations such as Biden is president. This is naturally seen as a structured representational system involving subsymbolic computational: r-LOT without c-LOT. The structured connectionist research program is still a work in progress, but it is arguable that contemporary large language models also combine structured representation (of facts such as Biden is president) with subsymbolic computation. A second and third way of combining r-LOT with subsymbolic computation are provided by the framework of vector symbolic architectures (Kleyko et al., Reference Kleyko, Davies, Frady, Kanerva, Kent, Olshausen and Rabaey2022), where representations are vectors, and Piantadosi's combinator framework (Reference Piantadosi2021), where the computational primitives S and K fall below the level of representation.
(Terminology: All three of these are computational versions of r-LOT in a broad sense. In an alternative phraseology, one might call the Fodorian version the classical computational LOT (cc-LOT), while calling subsymbolic versions nonclassical computational LOT (nc-LOT). But I will reserve “c-LOT” for the classical Fodorian version.)
Proponents of LOT often argue that structured connectionism is merely an implementation of LOT. We can now see that this claim is false or at best misleading. Implementation is standardly a computational relation between algorithms, requiring the implementing algorithm to be a more fine-grained version of the implemented algorithm with the same input/output behavior. The most interesting subsymbolic algorithms (e.g., in artificial neural networks) are never implementations of symbolic algorithms in this sense. The success of the deep-learning paradigm has provided strong evidence that the behavior of these systems (especially their success in learning and generalizing, but also their post-learning success) is not the result of implementing a more coarse-grained symbolic algorithm and cannot be duplicated by such algorithms. These systems may realize an r-LOT, but they do not implement a c-LOT. The quasi-symbolic operations of composition, decomposition, and quasi-logical inference may be available, but they are a tiny subset of the operations one can perform on the relevant distributed representations. As I argued in Chalmers (Reference Chalmers1990), one can also perform all sorts of holistic operations on distributed representations that do not proceed via these symbolic operations. It is plausibly subsymbolic operations like this that are largely responsible for the remarkable capacities of neural network systems.
Quilty-Dunn et al. don't make the distinction between r-LOT and c-LOT in their article, but their LOT appears to be a version of r-LOT. Their six core claims defining LOT do not mention computation (except in one case, incidentally). Four of the key claims (role-filler independence, predicate–argument structure, logical operators, abstract conceptual content) clearly pertain to representation but not computation. A fifth (inferential promiscuity) mentions computational theories of logical inference as versions of LOT, but computation does not play a defining role, and inferential promiscuity can equally be present in r-LOT without c-LOT (e.g., Ockham-style or subsymbolic systems).
The requirement of “discrete constituents” may suggest c-LOT, though it doesn't mention computation explicitly. Distributed representations in a structured connectionist systems arguably aren't discrete in the authors sense, in that representation of Biden and of president (say) can be intertwined nondiscretely in a representation of Biden is president. On the contrary, many subsymbolic computational systems involve discrete constituents without c-LOT. Piantadosi's system is one. Another is provided by the word embedding format for representing words that is ubiquitous in current language models. Here words are represented by multidimensional vectors where individual units often lack any clear semantic significance. “Biden is president” may be represented as a sequence of vectors for the individual words, so the constituents are discrete, but representations remain distributed and processing remains subsymbolic. So the discrete representational constituents do not require c-LOT.
Now, perhaps the absence of a computational constraint is an easily correctable omission. Quilty-Dunn et al. discuss computational approaches at some length in other sections of their article. They could easily enough add a seventh constraint connecting computation to representation, holding that the representational primitives are computationally primitive and serve as the medium of computation. The trouble is that strong evidence for this seventh claim is much harder to find.
The target article does argue that many Bayesian theorists provide computational accounts involving a “probabilistic LOT” associated with sententially structured representations. This suggests r-LOT, but it does not obviously lead to c-LOT, as Bayesian accounts are usually not cast at the algorithmic level (rather, at Marr's higher “computational” level). These accounts have many algorithmic implementations, including subsymbolic implementations in deep-learning systems. So there is no obvious strong evidence for c-LOT here, and any evidence would need to be stacked against the counterevidence provided by deep-learning models.
Overall: If Quilty-Dunn et al. are defending c-LOT, then more work is needed to make the defense explicit. If they are defending only r-LOT, then their conclusion is plausible, and my only objection is one of relative unambition.
There are two versions of the language-of-thought hypothesis. The representational language-of-thought hypothesis (r-LOT), introduced by William of Ockham and defended by Quilty-Dunn et al., concerns the structure of mental representation. The computational language-of-thought hypothesis (c-LOT), introduced by Jerry Fodor, concerns computation over mental representations. r-LOT is much weaker than c-LOT and is more widely accepted. I accept the former but reject the latter. As a result, I agree with many of Quilty-Dunn et al.'s conclusions while finding that they have not really defended the most controversial form of LOT.
In more detail: r-LOT (I use “LOT” for both the language and the hypothesis) says roughly that thought involves sententially structured mental representations. At minimum, there are nominal representations (e.g., Biden) and predicative representations (e.g., president) that combine into structured representations (e.g., Biden is president) with propositional content. Structured representations may also involve connectives (e.g., and), quantifiers (e.g., all), operators (e.g., always), and other types familiar from the linguistic case.
r-LOT is not trivially true, but it is plausible and hard to deny. It follows naturally from the claims that (1) people make judgments such as Biden is president, (2) these judgments involve combining nominal and predicative representations (or concepts, in the sense where concepts are mental representations) such as Biden and president, and (3) these representations can be recombined in judgments such as Biden is from Delaware. My sense is that most contemporary cognitive scientists and philosophers of mind accept these fairly weak claims. Importantly, these claims do not have immediate consequences regarding computation or cognitive architecture.
The c-LOT adds to r-LOT the key claim that thought involves computation over these sententially structured representations. The classical version of this hypothesis says that r-LOT representations are the medium through which all cognitive computation takes place. That is, the basic vehicles of representation in the r-LOT system (atomic words in the representational language-of-thought) also serve as the basic vehicles of computation (atomic computational states to which cognitive algorithms apply).
The c-LOT hypothesis was canonically formulated by Jerry Fodor's book The Language of Thought (Reference Fodor1975). Computation plays a central role throughout the book, from the main argument for LOT at the start of chapter 1 (“Computation presupposes a medium of computation: a representational system,” p. 27) to the conclusion (“More exactly: Mental states are relations between organisms and internal representations, and causally interrelated mental states succeed one another according to computational principles which apply formally to the representations,” p. 198). There are other works (e.g., “Propositional Attitudes”) in which Fodor focuses mainly on r-LOT, but computation is central in the canonical statement.
(Related distinctions: Fodor himself [Reference Fodor1980] distinguishes the “representational theory of mind” and the “computational theory of mind” [though neither requires a language-of-thought]. Rescorla [Reference Rescorla and Zalta2017] distinguishes a core version of LOT that involves “representational theory of thought” plus “compositionality of thought” and perhaps “logical structure” [in my terms, a version of r-LOT] from a stronger version that adds “the classical computational theory of mind” [yielding a version of c-LOT, though I understand the computational constraint differently from Rescorla].)
Most work in symbolic artificial intelligence (AI) uses a version of c-LOT. Both involve computation over atomic symbols: Entities that are both representationally atomic and computationally atomic. Atomic symbols have no computationally relevant internal structure (if they did, they would not be computationally atomic). Instead, their internal form is arbitrary.
The most significant opposition to LOT, in the classical-connectionist debate, has been opposition to c-LOT. In most neural network models there are no computationally atomic symbols. Representations are distributed over multiple quasi-neural units. As a result, in these models computation is subsymbolic computation: Computation takes place among units below the level of representation. Because computational primitives (units) are not representational primitives in these models, representation is not the medium of computation. Subsymbolic computation is incompatible with c-LOT.
At the same time, subsymbolic computation is quite compatible with r-LOT. This is clearest in the work of structured connectionists (e.g., Chalmers, Reference Chalmers1990; Smolensky, Reference Smolensky1988), where distributed representations (e.g., of Biden and president) can combine with each other systematically to yield new distributed representations such as Biden is president. This is naturally seen as a structured representational system involving subsymbolic computational: r-LOT without c-LOT. The structured connectionist research program is still a work in progress, but it is arguable that contemporary large language models also combine structured representation (of facts such as Biden is president) with subsymbolic computation. A second and third way of combining r-LOT with subsymbolic computation are provided by the framework of vector symbolic architectures (Kleyko et al., Reference Kleyko, Davies, Frady, Kanerva, Kent, Olshausen and Rabaey2022), where representations are vectors, and Piantadosi's combinator framework (Reference Piantadosi2021), where the computational primitives S and K fall below the level of representation.
(Terminology: All three of these are computational versions of r-LOT in a broad sense. In an alternative phraseology, one might call the Fodorian version the classical computational LOT (cc-LOT), while calling subsymbolic versions nonclassical computational LOT (nc-LOT). But I will reserve “c-LOT” for the classical Fodorian version.)
Proponents of LOT often argue that structured connectionism is merely an implementation of LOT. We can now see that this claim is false or at best misleading. Implementation is standardly a computational relation between algorithms, requiring the implementing algorithm to be a more fine-grained version of the implemented algorithm with the same input/output behavior. The most interesting subsymbolic algorithms (e.g., in artificial neural networks) are never implementations of symbolic algorithms in this sense. The success of the deep-learning paradigm has provided strong evidence that the behavior of these systems (especially their success in learning and generalizing, but also their post-learning success) is not the result of implementing a more coarse-grained symbolic algorithm and cannot be duplicated by such algorithms. These systems may realize an r-LOT, but they do not implement a c-LOT. The quasi-symbolic operations of composition, decomposition, and quasi-logical inference may be available, but they are a tiny subset of the operations one can perform on the relevant distributed representations. As I argued in Chalmers (Reference Chalmers1990), one can also perform all sorts of holistic operations on distributed representations that do not proceed via these symbolic operations. It is plausibly subsymbolic operations like this that are largely responsible for the remarkable capacities of neural network systems.
Quilty-Dunn et al. don't make the distinction between r-LOT and c-LOT in their article, but their LOT appears to be a version of r-LOT. Their six core claims defining LOT do not mention computation (except in one case, incidentally). Four of the key claims (role-filler independence, predicate–argument structure, logical operators, abstract conceptual content) clearly pertain to representation but not computation. A fifth (inferential promiscuity) mentions computational theories of logical inference as versions of LOT, but computation does not play a defining role, and inferential promiscuity can equally be present in r-LOT without c-LOT (e.g., Ockham-style or subsymbolic systems).
The requirement of “discrete constituents” may suggest c-LOT, though it doesn't mention computation explicitly. Distributed representations in a structured connectionist systems arguably aren't discrete in the authors sense, in that representation of Biden and of president (say) can be intertwined nondiscretely in a representation of Biden is president. On the contrary, many subsymbolic computational systems involve discrete constituents without c-LOT. Piantadosi's system is one. Another is provided by the word embedding format for representing words that is ubiquitous in current language models. Here words are represented by multidimensional vectors where individual units often lack any clear semantic significance. “Biden is president” may be represented as a sequence of vectors for the individual words, so the constituents are discrete, but representations remain distributed and processing remains subsymbolic. So the discrete representational constituents do not require c-LOT.
Now, perhaps the absence of a computational constraint is an easily correctable omission. Quilty-Dunn et al. discuss computational approaches at some length in other sections of their article. They could easily enough add a seventh constraint connecting computation to representation, holding that the representational primitives are computationally primitive and serve as the medium of computation. The trouble is that strong evidence for this seventh claim is much harder to find.
The target article does argue that many Bayesian theorists provide computational accounts involving a “probabilistic LOT” associated with sententially structured representations. This suggests r-LOT, but it does not obviously lead to c-LOT, as Bayesian accounts are usually not cast at the algorithmic level (rather, at Marr's higher “computational” level). These accounts have many algorithmic implementations, including subsymbolic implementations in deep-learning systems. So there is no obvious strong evidence for c-LOT here, and any evidence would need to be stacked against the counterevidence provided by deep-learning models.
Overall: If Quilty-Dunn et al. are defending c-LOT, then more work is needed to make the defense explicit. If they are defending only r-LOT, then their conclusion is plausible, and my only objection is one of relative unambition.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Competing interest
None.