‘Neuronal representation’ is not a defective concept: Ambiguity as a sign of science in progress
Lotem Elber-Dorozko and Devin Gouvêa
The term “neural representation” is extremely popular in cognitive neuroscience.
Scientists often claim to have discovered neural properties that represent environmental features, such as specific visual objects or specific sounds.
At the same time, both philosophers and scientists seem increasingly dissatisfied with this term. They often complain that the concept is so ambiguous and imprecise that it breeds confusion and invites equivocation instead of allowing helpful scientific discourse (Favela and Machery 2023; Baker et al. 2022). According to these critics, the term should be eliminated from scientific discourse, or at the very least be precisified, to avoid needless harm to scientific practice.
We present a different interpretation of the current state of affairs: the ambiguity of “neural representation” is an essential feature of the current stage of neuroscience, not a sign of sloppy conceptualization. Thus, such ambiguity cannot be resolved through philosophical analysis alone. To make progress in the understanding of ‘neural representation’, empirical research is also required.
Our argument draws on existing philosophical work on other imprecise scientific concepts, but also goes beyond that work. First, we note that scientific concepts function not merely as labels for settled facts, but also as scaffolding that guides the generation of new knowledge (Brigandt, 2020). Within this framework, a concept is defined in part by its epistemic goal—the specific form of knowledge (e.g., explanation, prediction, classification) that it is intended to enable. Why then should concepts be ambiguous? The literature often focuses on the benefits of ambiguity, suggesting that it allows integration of research methods pursuing a similar goal. This claim may very well be true, but it does not explain or justify ambiguity. When do the proposed benefits outweigh the alleged harms, and how are scientists supposed to tell the difference? We suggest that in the case of “neural representation”, and likely in other cases as well, the justification lies elsewhere. Conceptual ambiguity is ineliminable because it is not yet known how to achieve the epistemic goal associated with the concept, and so it must be tolerated if the research is to proceed at all.
For “neural representation,” the primary epistemic goal is to explain how nervous systems implement cognitive capacities like perception, memory, and decision-making. To quote Cao (2022, p. 150), “We want neural representations to help us bridge the mechanistic ‘wiring-and-connection facts’ that our experimental methods allow us to collect data about, and the ‘interpretation facts’ that defined our explananda in the first place.” To illustrate this point, consider the response patterns of neurons in the retina. The phosphenes that we experience when pressing our eyes indicate that these neurons respond to mechanical pressure as well as to photons. But only neurons’ response to visual stimuli is conventionally related to the cognitive function of visual processing, and thus scientists speak of them as representing light, not mechanical pressure (Elber-Dorozko and Loewenstein, 2023).
The ambiguity of “neural representation” arises because scientists have not yet reached a consensus on the form that successful explanations of cognitive capacities should take. While many agree that explanations should be mechanistic (broadly understood), key properties of said explanations are still unknown. It is not yet known which neural properties will ultimately explain cognitive phenomena, whether these are activity of single neurons, dynamical processes, network properties, or other features. It is not yet known which of the many environmental properties that causally relate to cognitive phenomena will turn out to be essential to the explanation of cognitive functions. Finally, the explananda themselves are still imprecise: working memory, decision-making, language, and visual perception are all paradigmatic cognitive functions which can be understood in various ways. Therefore, it is not surprising that scientists are very liberal when using the term representation. It is still unknown how to determine which properties are viable candidates to play a part in explanations of cognitive functions, and many types of properties cannot be ruled out a priori.
Some may worry that our approach ignores much philosophical writing about the concept of ‘representation’, which suggests that representations must take specific forms. It is widely believed that folk-psychological notions of mental representations are right (see Fodor (1975) on the language of thought hypothesis). This leads many to expect that neural representations should also be explicated in folk-psychological terms. Others argue that scientists are too liberal with the term ‘representation,’ and that mere causal detectors should be excluded from it (Ramsey, 2007). We respond that scientists should not be so convinced by these philosophical arguments as to abandon possible lines of research. Not only are all such philosophical arguments under contention, but new empirical research frequently sheds new light on philosophical debates. For example, the success of deep-learning networks caused many to reconsider the possibility of non-symbolic representations that cannot be explicated with folk-psychological notions. Although philosophical work may assist scientists in examining their assumptions and methods, it cannot introduce precision where the evidence provides none.
We want to emphasize that we do not aim to promote a kind of ‘anything goes’ approach to ‘neural representation’. Certainly, there are cases where ambiguity is unnecessarily harmful and should be avoided when possible. Such cases include the use of ‘representation’ solely for social or professional gains, or the use of ‘representation’ not in service of its epistemic goal. Even as we reject some uses of the term, and some forms of ambiguity, we argue that we must accept that imprecision is unavoidable at this stage of scientific research.
What, then, is the way forward? We suggest that the most promising way to make scientific progress and to clarify the concept of ‘representation’ is by conducting more empirical work, combined with scientific and philosophical debate. That is, in many cases, scientists will not know what they are looking for until they find it. Only by identifying more results, ambiguous and confusing to varying degrees, will scientists make progress. Let us conclude with the perspective of a leading researcher on homology, another imprecise concept: “Giving a definition for a term or concept presumes that we already fully understand our study object,” and premature definition “has the unfortunate effect of suggesting precision where there is none” (Wagner, 2014, p. 242).
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
Brigandt, I. (2020). How are biology concepts used and transformed? In K. Kampourakis & T. Uller (Eds.), Philosophy of science for biologists (pp. 79–101). Cambridge University Press. https://doi.org/10.1017/9781108648981.006
Cao, R. (2022). Putting representations to use. Synthese, 200(2), 151. https://doi.org/10.1007/s11229-022-03522-3
Elber-Dorozko, L., & Loewenstein, Y. (2023). Do retinal neurons also represent somatosensory inputs? On why neuronal responses are not sufficient to determine what neurons do. Cognitive Science, 47(4), e13265. https://doi.org/10.1111/cogs.13265
Favela, L., & Machery, E. (2023). Investigating the concept of representation in the neural and psychological sciences. Frontiers in Psychology, 14, 1165622. https://doi.org/10.3389/fpsyg.2023.1165622
Fodor, J. A. (1975). The language of thought. Harvard University Press.
Ramsey, W. (2007). Representation reconsidered. Cambridge University Press. Wagner, G. (2014). Homology, genes, and evolutionary innovation. Princeton University Press. https://press.princeton.edu/books/hardcover/9780691156460/homology-genes-and-evolutionary-innovation
The authors argue that ambiguity in “neural representation” is ineliminable because scientists do not yet know what form successful explanations of cognitive capacities will take. This is an honest concession, but it raises a question the paper does not address: if the concept is defined by an epistemic goal that remains unspecified, how will scientists recognize success when they encounter it? The appeal to the homology analogy does not help here, because homology, however imprecise, operates within a single descriptive register, namely the biological. “Neural representation,” by contrast, smuggles a semantic notion, namely content or meaning, into a causal-physical domain. That is not imprecision; it is a category error, and more empirical work cannot dissolve it, because no accumulation of firing-rate correlations will ever settle the question of what something means or is about. Neuroscience does not need the concept of representation to do its work. What it needs, and successfully uses, are correlational and causal descriptions. The representational vocabulary adds nothing to the explanatory or predictive power of those descriptions; it only imports normative assumptions about content that cannot be operationalized from within the framework itself.
It certainly sounds promisingly Popperian, with all due respect.
JR