The predominance of machine learning based techniques in cognitive neuroscience raises a host of philosophical and methodological concerns. Given the messiness of neural activity, modellers must make choices about how to structure their raw data to make inferences about encoded representations. This leads to a set of standard methodological assumptions about when abstraction is appropriate in neuroscientific practice. Yet, when made uncritically these choices threaten to bias conclusions about phenomena drawn from data. Contact between the practices of multivariate pattern analysis (...) and philosophy of science can help to illuminate the conditions under which we can use artificial neural networks to better understand neural mechanisms. This paper considers a specific technique for MVPA called representational similarity analysis. I develop a theoretically-informed account of RSA that draws on early connectionist research and work on idealization in the philosophy of science. By bringing a philosophical account of cognitive modelling in conversation with RSA, this paper clarifies the practices of neuroscientists and provides a generalizable framework for using artificial neural networks to study neural mechanisms in the brain. (shrink)

Neural structural representations are cerebral map- or model-like structures that structurally resemble what they represent. These representations are absolutely central to the “cognitive neuroscience revolution”, as they are the only type of representation compatible with the revolutionaries’ mechanistic commitments. Crucially, however, these very same commitments entail that structural representations can be observed in the swirl of neuronal activity. Here, I argue that no structural representations have been observed being present in our neuronal activity, no matter the spatiotemporal scale of observation. (...) My argument begins by introducing the “cognitive neuroscience revolution” (Sect. 1 ) and sketching a prominent, widely adopted account of structural representations (Sect. 2 ). Then, I will consult various reports that describe our neuronal activity at various spatiotemporal scales, arguing that none of them reports the presence of structural representations (Sect. 3 ). After having deflected certain intuitive objections to my analysis (Sect. 4 ), I will conclude that, in the absence of neural structural representations, representationalism and mechanism can’t go together, and so the “cognitive neuroscience revolution” is forced to abandon one of its commitments (Sect. 5 ). (shrink)

Are there representations in the brain? It depends on what you mean by representations, and it depends on what you want them to do for you—both in terms of the causal role they play in the system, and in terms of their explanatory value. But ideally, we would like an account of representation that allows us to assign a representational role and content to the appropriate mechanistic precursors of behavior that in fact play that role and conversely, search for the (...) mechanistic realizers of representational roles that are posited by our models of behavior and cognition. Such an account would be methodologically valuable in neuroscience. Lately, people have started asking similar questions about deep neural networks. What representations do they learn and use, and what is the relation between those representations and the sometimes impressive capacities these networks exhibit? More philosophically puzzling, perhaps, is the question of why the internal activities or dispositions labeled as such in neuroscience and AI research should count as representations at all. I think we can give a unified answer to both sets of questions, in the form of a kind of representational pragmatism. What makes something a representation just is that we can identify and re-identify it as such, and moreover, that we can manipulate it effectively to do whatever it is that we think representations ought to do for us in that particular context. For the first condition, the idea is that so long as we have some probe that allows us to pick out the relevant causally effective candidate, we should consider that to be a candidate representation-relative-to-that-probe. The second condition can be understood as a kind of anti-gerrymandering constraint: one ought to be able to intervene on the candidate, and see effects of that intervention consistent with the functional role the representation is supposed to play. Naturally the details need to be spelled out for different contexts—and I will articulate them for a few illustrative cases—which effectively allows for a kind of pluralism about representation depending on the setting and the kinds of questions being asked. With the extra components filled in, representational pragmatism can make sense of existing practices in neuroscience and AI, as well as their relationship to naturalistic theories of representation in philosophy. (shrink)

“The physics of representation” aims to define the word “representation” as used in the neurosciences, argue that such representations as described in neuroscience are related to and usefully illuminated by the representations generated by modern neural networks, and establish that these entities are “representations in good standing”. We suggest that Poldrack succeeds in, exposes some tensions between the broad use of the term in neuroscience and the narrower class of entities that he identifies in the end, and between the meaning (...) of “representation” in neuroscience and in psychology in, and fails in. This results in some hard choices: give up on the broad scope of the term in neuroscience or continue to embrace the broad, psychologically inflected sense of the term, and deny the entities generated by neural nets are representations in the relevant sense. (shrink)