Quilty-Dunn et al. offer six properties of language-of-thoughts (LoTs). Setting aside whether those are the right ones, all proponents of the language-of-thought hypothesis (LoTH) must specify the status of candidate properties: Are they diagnostic or constitutive? On the diagnostic view, these properties are indicators. Their presence is evidence for an underlying LoT-like representational format, but the format itself would be a distinct natural kind, causally prior to the properties. On the constitutive view, what it is to be an LoT is to exhibit some (or all) of these properties; a system just is “LoT-like” to the extent that it implements them. Both views require methods for determining whether the behavior of a given system reflects properties of LoTs. On the diagnostic view, a further question remains regarding whether observed properties really reflect an underlying LoT.
The authors do not make this distinction and different parts of their argument suggest different interpretations. Here we show that one can't have both, because the consequences of these interpretations are incompatible. For one thing, the choice determines whether neural networks are competitors to the LoTH or candidate implementations.
The authors' central argument seems to place them in the diagnostic camp. They suggest that their six core properties comprise a homeostatic property cluster (Boyd, Reference Boyd1991, Reference Boyd1999). As opposed to prototype concepts, in which features need not be related, a crucial question for homeostatic property clusters is: What maintains the clustering? In the standard case, it is maintained by an underlying homeostatic mechanism (or set of mechanisms), which causes the properties to co-occur to an unexpected degree (Boyd, Reference Boyd and Savage1990). For example, the characteristic properties of a biological species – the paradigm case of a homeostatic property cluster – are diagnostic of that species, and tend to co-occur because of the shared genotype of species members, which is maintained by evolutionary forces. Proposing that the LoT is a homeostatic property cluster evokes an analogy: Some underlying homeostatic mechanism causes the properties of LoTs to cooccur. This mechanism is the extra constitutive component; even if two systems exhibit identical indicators, the presence or absence of the mechanism determines which ones really are LoTs. This accords with the authors' treatment of “non-LoT-like architectures such as DNNs” (target article, sect. 3, fn. 5) as a priori incompatible alternatives to LoTs through much of the paper. Although the core properties may emerge in these systems, a difference in (or lack of) the underlying LoT-like mechanism would mean that deep neural networks (DNNs) could never count as LoTs.
If the properties are just diagnostic, what more is needed for an LoT? The authors do not provide a specific proposal for an underlying homeostatic mechanism, but without one it is unclear what rates of co-occurrence of properties the LoTH predicts. In lieu of a specific prediction, the authors suggest that properties should at least co-occur more frequently than one would expect “from a theory-neutral point of view” (target article, sect. 2, para. 13). But what would one expect? The baseline cannot be a “chance” rate of co-occurrence, because the properties that the authors specify are not, in principle, independent. Predicate–argument structure seems to presuppose both role-filler independence and discrete constituents, while having logical operators should enable inferential promiscuity. Co-occurrence of properties can only be evidence for the LoTH if it is co-occurrence over and above the rate implied by their mutual dependence. Without an estimate of this baseline, it is unclear whether the evidence the authors review actually provides a compelling “abductive, empirical argument for LoTH” (target article, sect. 2, para. 13). On a diagnostic view, both what the diagnostic properties are and their expected rate of co-occurrence should ultimately be causally determined by the underlying homeostatic mechanism. The challenge, then, is to characterize that mechanism.
The constitutive view sidesteps this explanatory challenge. On this view, all there is to being LoT-like is exhibiting the relevant properties. There is no further prediction about above-baseline cooccurrence and no need to posit any underlying mechanisms that maintain homeostatic unity. Nor are the LoTH and DNNs incompatible explanatory paradigms competing to account for the same experimental data. Rather, the LoTH highlights important, multiply realizable properties that stand as targets for any representational format to instantiate, and which might emerge in neural networks that are not explicitly augmented with other, more LoT-like mechanisms. This amounts to a form of compatibilism about DNNs and the LoTH. At one point, the authors explicitly endorse this position. They write, “neural-network architectures might be able to implement an LoT architecture… Our six core LoT properties help specify a cluster of features that such an implementation should aim for” (target article, sect. 4.3, para. 6), and they are unwilling to suggest any “in-principle limitations of DNNs” (target article, sect. 4.3, para. 6). However, this clashes with their central argument. On a constitutive view, there is no unifying homeostatic mechanism, so the homeostatic cluster collapses into a prototype concept. Moreover, the pieces of evidence presented by the authors that particular current DNNs fail to manifest LoT properties or explain some phenomenon (e.g., abstract object representations) cannot weigh in favor of the LoTH and against DNNs as cognitive models in principle. They just suggest that those DNNs do not implement LoTs. This is a critical choice point for proponents of the LoTH: You can embrace compatibilism and the constitutive view or endorse the more robust commitments of the diagnostic property cluster account. You cannot have both.
An unresolved question may influence the choice between these two options. Will neural networks need to be augmented with rule-like operations to account for human competences, as the authors suggest? If so, this would favor the diagnostic view, with LoTs as underlying mechanisms that play a strong explanatory role in cognitive architecture. If, on the contrary, DNNs turn out to be able to explain human cognitive capacities without being augmented with other kinds of architectures (as in neuro-symbolic hybrids), that would support the constitutive view and a weaker, guiding role for the LoTH. Far from abandoning it, this result would allow the LoTH to provide its traditional explanatory benefits without requiring implementation in a rule-based system.
Quilty-Dunn et al. offer six properties of language-of-thoughts (LoTs). Setting aside whether those are the right ones, all proponents of the language-of-thought hypothesis (LoTH) must specify the status of candidate properties: Are they diagnostic or constitutive? On the diagnostic view, these properties are indicators. Their presence is evidence for an underlying LoT-like representational format, but the format itself would be a distinct natural kind, causally prior to the properties. On the constitutive view, what it is to be an LoT is to exhibit some (or all) of these properties; a system just is “LoT-like” to the extent that it implements them. Both views require methods for determining whether the behavior of a given system reflects properties of LoTs. On the diagnostic view, a further question remains regarding whether observed properties really reflect an underlying LoT.
The authors do not make this distinction and different parts of their argument suggest different interpretations. Here we show that one can't have both, because the consequences of these interpretations are incompatible. For one thing, the choice determines whether neural networks are competitors to the LoTH or candidate implementations.
The authors' central argument seems to place them in the diagnostic camp. They suggest that their six core properties comprise a homeostatic property cluster (Boyd, Reference Boyd1991, Reference Boyd1999). As opposed to prototype concepts, in which features need not be related, a crucial question for homeostatic property clusters is: What maintains the clustering? In the standard case, it is maintained by an underlying homeostatic mechanism (or set of mechanisms), which causes the properties to co-occur to an unexpected degree (Boyd, Reference Boyd and Savage1990). For example, the characteristic properties of a biological species – the paradigm case of a homeostatic property cluster – are diagnostic of that species, and tend to co-occur because of the shared genotype of species members, which is maintained by evolutionary forces. Proposing that the LoT is a homeostatic property cluster evokes an analogy: Some underlying homeostatic mechanism causes the properties of LoTs to cooccur. This mechanism is the extra constitutive component; even if two systems exhibit identical indicators, the presence or absence of the mechanism determines which ones really are LoTs. This accords with the authors' treatment of “non-LoT-like architectures such as DNNs” (target article, sect. 3, fn. 5) as a priori incompatible alternatives to LoTs through much of the paper. Although the core properties may emerge in these systems, a difference in (or lack of) the underlying LoT-like mechanism would mean that deep neural networks (DNNs) could never count as LoTs.
If the properties are just diagnostic, what more is needed for an LoT? The authors do not provide a specific proposal for an underlying homeostatic mechanism, but without one it is unclear what rates of co-occurrence of properties the LoTH predicts. In lieu of a specific prediction, the authors suggest that properties should at least co-occur more frequently than one would expect “from a theory-neutral point of view” (target article, sect. 2, para. 13). But what would one expect? The baseline cannot be a “chance” rate of co-occurrence, because the properties that the authors specify are not, in principle, independent. Predicate–argument structure seems to presuppose both role-filler independence and discrete constituents, while having logical operators should enable inferential promiscuity. Co-occurrence of properties can only be evidence for the LoTH if it is co-occurrence over and above the rate implied by their mutual dependence. Without an estimate of this baseline, it is unclear whether the evidence the authors review actually provides a compelling “abductive, empirical argument for LoTH” (target article, sect. 2, para. 13). On a diagnostic view, both what the diagnostic properties are and their expected rate of co-occurrence should ultimately be causally determined by the underlying homeostatic mechanism. The challenge, then, is to characterize that mechanism.
The constitutive view sidesteps this explanatory challenge. On this view, all there is to being LoT-like is exhibiting the relevant properties. There is no further prediction about above-baseline cooccurrence and no need to posit any underlying mechanisms that maintain homeostatic unity. Nor are the LoTH and DNNs incompatible explanatory paradigms competing to account for the same experimental data. Rather, the LoTH highlights important, multiply realizable properties that stand as targets for any representational format to instantiate, and which might emerge in neural networks that are not explicitly augmented with other, more LoT-like mechanisms. This amounts to a form of compatibilism about DNNs and the LoTH. At one point, the authors explicitly endorse this position. They write, “neural-network architectures might be able to implement an LoT architecture… Our six core LoT properties help specify a cluster of features that such an implementation should aim for” (target article, sect. 4.3, para. 6), and they are unwilling to suggest any “in-principle limitations of DNNs” (target article, sect. 4.3, para. 6). However, this clashes with their central argument. On a constitutive view, there is no unifying homeostatic mechanism, so the homeostatic cluster collapses into a prototype concept. Moreover, the pieces of evidence presented by the authors that particular current DNNs fail to manifest LoT properties or explain some phenomenon (e.g., abstract object representations) cannot weigh in favor of the LoTH and against DNNs as cognitive models in principle. They just suggest that those DNNs do not implement LoTs. This is a critical choice point for proponents of the LoTH: You can embrace compatibilism and the constitutive view or endorse the more robust commitments of the diagnostic property cluster account. You cannot have both.
An unresolved question may influence the choice between these two options. Will neural networks need to be augmented with rule-like operations to account for human competences, as the authors suggest? If so, this would favor the diagnostic view, with LoTs as underlying mechanisms that play a strong explanatory role in cognitive architecture. If, on the contrary, DNNs turn out to be able to explain human cognitive capacities without being augmented with other kinds of architectures (as in neuro-symbolic hybrids), that would support the constitutive view and a weaker, guiding role for the LoTH. Far from abandoning it, this result would allow the LoTH to provide its traditional explanatory benefits without requiring implementation in a rule-based system.
Acknowledgments
We are grateful for helpful discussions with Joshua Schechter, Richard Kimberley Heck, Stavros Orfeas Zormpalas, and Randall O'Reilly.
Financial support
This work was supported in part by an SSNAP Fellowship (award no. 383-001202) to S. M. and J. R., funded by the John Templeton Foundation, and by a Jacobs Foundation Fellowship, awarded to R. F.
Competing interest
None.