Luis H. Favela and Edouard Machery
We thank Dan Burnston and Nick Byrd for the opportunity to present and discuss our paper, “Investigating the concept of representation in the neural and psychological sciences” (Favela & Machery, 2023) with The Brains Blog community.
Our project was sparked by the following facts: First, the concept of representation is frequently used in the mind and brain sciences and philosophies of those sciences. These facts are evident from research in basic neuroanatomy—e.g., on single neurons responding to specific shape orientation—to research on highly abstract cognitive capacities—e.g., on creative problem solving. Second, many seem to take it as obvious that neural activity and mental capacities are representational in nature (see the computational theory of cognition/mind).
But what does the concept of representation even mean in empirical and theoretical brain and mind research? While there have been some attempts to provide clarity on the wide and varied application of the concept of representation (e.g., Baker et al., 2022), there have been very few studies attempting to provide empirical evidence about its use (but see Vilarroya, 2017).
Thus, the aim of our project was, first, to provide evidence about how neuroscientists, psychologists, and philosophers use the concept of representation. By addressing the lack of systematic evidence about the ways the concept of representation is applied, we also hoped to bring new considerations to bear on whether the concept of representation should be used in the brain and mind sciences.
With these goals in mind, we conducted four studies directed at identifying neuroscientists’, psychologists’, and philosophers’ implicit commitments that guide their application of three kinds of concepts to describe brain states: intentional (“representation,” being “about,” “identifying”), causal (“responding,” “processing”), and information theoretic (“carrying information”) concepts. In each study, participants were presented with a fictional study involving the perception of faces inspired by the kind of neuroimaging work commonly conducted for the past 30 years (e.g., Kanwisher et al., 1997; Harpaintner et al., 2020). After reading a short description of the study and examining figures (e.g., magnetic resonance image) and a time series, participants were asked to rate whether the brain state recorded in the study could be described by means of the six terms mentioned above (1: “Strongly agree;” 7: “Strongly disagree”). This methodology is inspired by elicitation techniques (e.g., Gillioz & Zufferey, 2020), which do not ask participants to reflect on how they use terms, but instead provide them with realistic scenarios or situations and ask them to use the target concept.
Details of the studies can be found in our paper (Favela & Machery, 2023). As an overview, Study 1 examined at which scale neuroscientists expect representations to be found (e.g., single neuron or population of neurons)? Study 2 examined whether the specificity of the relationship specific brain activity and particular stimuli matters for the application of the concept of representation. For example, high specificity would obtain if the fusiform face area (FFA) were activated only by faces; low specificity if it were activated by faces and by houses. Study 3 examined whether the functional role of brain activity (Cummins, 1975) influences how brain activity is described by contrasting being correlated with a stimulus and being part of a neural network. Study 4 examined whether participants were willing to describe brain activity as misrepresenting when the FFA is activated by a house instead of a face.
In brief, our main findings were as follows: First, neuroscientists, psychologists, and philosophers distinguish between the different ways of describing the brain’s reaction to stimuli. They find causal (“responds” and “processes”) and information-theoretic descriptions (“carries information”) more applicable than intentional descriptions (“about” and “identify”), including the concept of representation. On average, neuroscientists, psychologists, and philosophers were uncertain about the use of these latter descriptions, neither accepting nor rejecting them. Second, neuroscientists, psychologists, and philosophers appear to have no expectation at the scale at which representations are to be found. Third, whether the brain state elicited by some stimulus was explicitly described as having a function made no difference to their willingness to describe it as a representation. Fourth, whether the brain activation was perfectly specific or not mattered for psychologists’ willingness to describe it as a representation, but we found no evidence it does for neuroscientists. Fifth, neuroscientists, psychologists, and philosophers were reluctant to describe the brain as failing to do what it is meant to do, including as misrepresenting.
We draw two main conclusions from these findings. First, the concept of representation is unclear. Participants are uncertain in how to apply the concept of representation in contrast to causal and information-theoretic concepts. Participants demonstrated that they are noncommittal regarding: one, what brain scale and structure could be a representation; two; if a representation must have a function (Cummins or otherwise); and three, the role of sensitivity (especially neuroscientists). Second, is that the concept of representation is confused. Participants are unwilling to state that a brain is misrepresenting, which reflects a confusion between representations and natural signs.
If the concept of representation is indeed unclear and confused, then we suggest three possible ways forward. One is to reform the concept, that is, to explicate it (Carnap) or to engineer it conceptually (e.g., Cappelen, 2018; Machery, 2017). This approach, however, seems unlikely to succeed given the little influence that philosophical work about representations has on neuroscience and psychological research, and the apparent disinterest of scientists to clarify that concept. Another option is to eliminate the concept (e.g., Churchlands, Dennett). This approach too seems unlikely to succeed: given the widespread use of the concept (i.e., Pandora’s box is already open), it may be impractical to eliminate that term, and we would have to offer a non-representational alternative that would be accepted across neuroscience and psychology as much as representations already are (for an attempt see Favela, 2024). A third option is to embrace an “epistemology of the imprecise” (Rheinberger, 2000). A number of concepts central to various disciplines continue to be used in spite of being “unclear” and “confused” (e.g., the concept of gene). Such concepts stick around for various reasons, including the fact that their lack of clarity facilitates cross-disciplinary employment: If the concept were too precise, then it would not be possible to use it widely across disciplines.
We hope this work contributes to and helps make progress on foundational issues in the mind and brain sciences and the philosophy of those sciences. We look forward to engaging with our commentators and The Brains Blog community.
References
Baker, B., Lansdell, B., & Kording, K. P. (2022). Three aspects of representation in neuroscience. Trends in Cognitive Sciences, 26(11), 942-958. https://doi.org/10.1016/j.tics.2022.08.014
Cappelen, H. (2018). Fixing language: An essay on conceptual engineering. Oxford, UK: Oxford University Press.
Cummins, R. (1975). Functional analysis. Journal of Philosophy, 72(20), 741-765. https://doi.org/10.2307/2024640
Favela, L. H. (2024). The ecological brain: Unifying the sciences of brain, body, and environment. New York, NY: Routledge.
Favela, L. H., & Machery, E. (2023). Investigating the concept of representation in the neural and psychological sciences. Frontiers in Psychology: Cognition, 14:1165622, 1-13. https://doi.org/10.3389/fpsyg.2023.1165622
Gillioz, C., & Zufferey, S. (2020). Introduction to experimental linguistics. Hoboken, NJ: John Wiley & Sons.
Harpaintner, M., Sim, E.-J., Trumpp, N. M., Ulrich, M., & Kiefer, M. (2020). The grounding of abstract concepts in the motor and visual system: An fMRI study. Cortex, 124, 1-22. https://doi.org/10.1016/j.cortex.2019.10.014
Kanwisher, N., McDermott, J., & Chun, M. M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. Journal of Neuroscience, 17(11), 4302-4311. https://doi.org/10.1523/JNEUROSCI.17-11-04302.1997
Machery, E. (2017). Philosophy within its proper bounds. Oxford, UK: Oxford University Press.
Rheinberger, H.-J. (2000). Gene concepts: Fragments from the perspective of molecular biology. In P. J. Beurton, R. Falk, & H.-J. Rheinberger (Eds.), The concept of the gene in development and evolution: Historical and epistemological perspectives (pp. 219-239). Cambridge, MA: Cambridge University Press. https://doi.org/10.1017/CBO9780511527296.012
Vilarroya, O. (2017). Neural representation. A survey-based analysis of the notion. Frontiers in Psychology: Theoretical and Philosophical Psychology, 8(1458), 1-15. https://doi.org/10.3389/fpsyg.2017.01458
I welcome strongly the F&M paper (Favela and Machery, 2023) as summarised in this précis. Having argued for decades (e.g. Harvey, 1996; 2008; 2021) that the concept of ‘representations’ is confused and unclear in the disciplines covered here, it is gratifying to see this opinion supported by the rigorous elicitation studies presented by F&M. Two brief comments:
Firstly, I was for a long time puzzled by scientists and philosophers who maintained representation was a key concept in studies of cognition, yet felt no need to define the concept in operational terms. Then I realised that they were often treating representation as an explanans — something familiar from everyday life that itself needs no further explanation — that could be uncontroversially used in explaining their theories. Much as physicists use billiard balls to explain (or illustrate) aspects of particle physics, representations may be the billiard balls of many cognitive scientists. In contrast other scientists (… or sometimes the same ones on different days of the week…) treat representations as explananda, concepts that their theories should explain. So a key question asked of respondents in F&M (Section 3.1): “Does cognition involve representations? Yes or no,” can easily be answered Yes by one respondent (meaning representations as explanans) and Yes by another respondent (meaning representations as explananda); without either realising how fundamentally they disagree.
A second comment is that F&M’s body of respondents does not obviously include anyone from artificial intelligence (AI); though possibly they may have been amongst those self-identified as ‘cognitive scientists’ who were excluded from F&M’s study. But I am most familiar with these conflicts on the roles and nature of representations in the fields of cognitive science and artificial intelligence, where for some 30 years it has been known as the Representation Wars (Williams, 2018). Much of the flavour of this has leaked across to the neuroscience, psychology and philosophical fields that the F&M respondents do reflect. There is a specific reason why practitioners of AI should in practice typically treat representations as explanans; the core technology of computer programming has the use of representations at its very centre, and this dependence makes their usage familiar and unquestioned; as a fish takes water for granted.
F&M rigorously expose the confusion and lack of clarity across these disciplines. They suggest either an eliminative approach to resolving the confusion, or a ‘precisification’ approach (that I would phrase as the need for operational definitions). But they should be prepared for those who take representations as unanalysed givens to offer resistance.
Harvey, I. (1996): Untimed and misrepresented: connectionism and the computer metaphor
AISB Quarterly, 1996 no. 96, pp. 20-27.
Harvey, I. (2008). Misrepresentations.
In S. Bullock, J. Noble, R. A. Watson, and M. A. Bedau (Eds.) Proceedings of the Eleventh International Conference on Artificial Life, pp.227-233, MIT Press, Cambridge, MA. ISBN: 978-0-262-75017-2g
Harvey I. (2021), Neurath’s boat and the Sally-Anne test: Life, Cognition, Matter and Stuff. Adaptive Behavior. 2021;29(5):459-470. doi:10.1177/1059712319856882
Williams, D. (2018). Predictive Processing and the Representation Wars. Minds and Machines 28 (1):141-172.
Inman
Many thanks for this comment.
I agree that the notion of representation is used both to describe the phenomena to be explained, and also as part of explanation. This is an insightful observation. Either way, however, the unclarity of the notion in neuroscience is remarkable, and might be problematic (but see the discussion of the epistemology of the imprecise).
I do not think however that calling the notion of representation a primitive will be of much help. First, I am not sure what it means to say that a notion is primitive. Perhaps it means that the notion cannot be defined because the notion is not made up of other notions. This might well be true, but of course this does not mean we cannot explicate that notion in some sense: for instance, by characterizing its role in the theory, by identifying converging operationalizations. Perhaps it means that the notion cannot be defined within the theory itself or at the same level. But again this would be compatible with some kind of explication, operationalization, and it could be defined from the outside (e.g., from a reducing theory).
Consider other scientific notions: fitness, field, species, organism, chemical bond. Whether or not they are primitive, and whatever that means, we can cast light on them, and we can usually find ways to operationalize them.
Your second point about AI researchers is interesting. We initially planned to have a broader, more diverse sample of participants, but getting scientists to take part to the survey was really challenging, and this is definitely one of the limitations of the project. What do you think is happening among AI researchers working on DNNs or ANNs? DNNs and ANNs are very different from digital computers and traditional algorithms, but nonetheless they cannot stop using representation talk.
thanks again for the comments.
Cool stuff, thank you.
It does seem to me that the “about faces” and “identifies faces” options are somewhat ambiguous.
If I see and recognize a zebra, for instance, and think, “That’s a zebra,” my thought was not about zebras, plural, nor was my percept. It was about that zebra.
A representation could similarly play some role in face processing, without playing the role of identifying faces.
The misrepresentation study is interesting. I do wonder whether scientists need a higher bar for attributing, as it were, a mistake to the brain. You might want to know a lot more about what it’s doing before you’re confident saying it’s doing something wrong. (I think I would feel that way.)
Thanks Lizzie for these comments!
I will start with the last point. It is indeed possible that participants’ reluctance to characterize what the brain does as a mistake, including as a misrepresentation, is epistemic rather than reflecting an uncertainty about what it would meant for the brain to do something incorrectly. Our data really cannot exclude this possibility. The only thing that speaks against it is the fact that we are talking about the FFA, a structure that is well known for its role in face processing (I know this is not uncontroversial – see expertise hypothesis), and that reacting to a single house does not seem to be what the structure should be doing.
Concerning the first two points: yes this is correct too, but here too, the literature on the FFA does not shy away from assigning face representations. The same is true of much assignment of content in neuroscience. Why were participants so shy in our study?
Thanks again
Thanks, Edouard, for your response to my comment. Just to avoid misunderstanding, I myself am in full agreement with you that treating the notion of representation as a primitive, or an unanalysed given, is of no help at all for the philosophical project of analysing representations. Yet many others involved in these debates seem regrettably unwilling to offer operational definitions. I suspect this is largely due to the seductive appeal of everyday functional explanations of complex systems to homuncular metaphors, where subsystems are framed as homunculi ‘communicating’ with each other via ‘representations’. For most such projects, representations work just fine as explanans; computer programming explicitly works like this. But this cannot and should not carry over to the project of analysing representations themselves, as explananda — that would be begging the question. I attribute much of the confusion to misunderstanding this.
You ask about AI researchers working on ANNs. A great majority of them do very explicitly ‘use representation talk’, which in my view has unnaturally biased their NN architectures to pipeline feedforward style. A minority (including e.g. Beer at Indiana, and the group I have been associated with at Sussex) eschew such representational-biased architectures for e.g. dynamical recurrent real-time networks.
Inman