Representations are thought to be among the essential ingredients for explaining cognition. In cognitive psychology, behaviour and decision making are modelled in terms of computations performed over representations. In neuroscience, responses of single cells, neural populations, and structural features are claimed to function as representations for the agent in question. In AI, what machines learn is often characterized in terms of the kinds of representations those machines acquire and how they manipulate them.
Although the term ‘representation’ is common to each of these domains, there is a marked lack of agreement about what that term means, or how it should be used in scientific discourse. Our grasp on ‘representation’ is rooted in our everyday experience of external, conventional representations – e.g. written or spoken language, diagrams, maps, symbols, signs. A host of difficulties spring up if one attempts to export assumptions about external, conventional forms of representations to representations inside the brain or representations inside modern AI systems.
The aim of this Special Issue of Philosophy and the Mind Sciences is to shed light on the concept of representation and its correct use in neuroscience and AI. Contributors to the Special Issue were encouraged to stake out ideas that are bold, constructive, or opinionated. The objective was to provide a space in which to share approaches, that while they may not be fully developed, nevertheless show a direction of travel for how one should understand representation in the brain and AI systems.
The papers for the Special Issue have fallen into roughly four clusters.
First, there are papers on large language models (LLMs). Raphael Millière and Dimitri Coelho Mollo argue that the representations manipulated by LLMs do not just have derived content (representational content that depends on external interpreters and social conventions). LLMs also have intrinsic content. Moreover, they don’t require additional sensorimotor grounding or embodied contact with the world in order to have this intrinsic content. Colin Klein argues that although current LLMs capable of representing complex structures – for example, causal structures in the world or linguistic structures – the structures of their own internal representational formats are remarkably simple compared to our own. Fintan Mallory claims that we should be pluralists about the representations inside LLMs: the vehicles of LLM representation should not be identified with any one functional component of the LLM architecture (e.g. single neurons, patterns of activation, or regions in activation space); the format of LLM representations should also not restricted to a single type (e.g. analogue, structural, or symbolic representation). Patrick Butlin examines whether LLMs have higher-order representations. Higher-order representations have been suggested as means to explain the ability of LLMs to ‘introspect’ on their own processes – to describe, predict, and explain their own behaviour. Butlin argues that there is prima facie evidence that LLMs do possess these higher-order representations, although the evidence is not conclusive.
Second, there are papers on types of representation. Megan Peters and Hojjat Azimi Asrari examine how the brain represents perceptual uncertainty (e.g. one’s estimation of whether one has gotten a particular perceptual discrimination right or wrong). They argue that measures of perceptual uncertainty are higher-order representations that are governed by the same Bayesian logic that controls first-order perceptual representations. Just as the brain attempts to build a Bayesian posterior concerning first-order perceptual properties based on its available information and prior expectations, so it also attempts to build a Bayesian posterior concerning how uncertain it should be about its first-order perceptual judgement based on its available information and prior expectations about uncertainty. Nick Shea explores how attributions of agency constrain attribution of representations for biological and AI systems. Realists about representation tend to ascribe representations to systems in an piecemeal fashion – whether a system contains some representation A is regarded as a largely independent matter from whether it contains a distinct representation B. Shea argues that if we were to grant that a system is an agent, this would place constraints on the representations it can contain: it needs to have representations that are in some sense coherent and a mechanism for ensuring their continued coherence. Realists should perhaps take a lesson from instrumentalists and begin ascribing representations to agents en bloc, rather than one by one. Corey Maley argues that two apparently distinct types of representation – structural representation and analogue representation – are actually the same kind of representation. The relationship between them becomes evident once one generalises from one-dimensional analogue representations (e.g. liquid thermometers) to their multi-dimensional counterparts (e.g. photographs): these analogue representations function by preserving a structure between a vehicle and a target, exactly as is assumed of structural representations. Mazviita Chirimuuta provides an historical analysis of the concept of structural representation. She explores connections between ideas currently associated with structural representation in AI and cognitive neuroscience and prior structuralist movements in mathematics, physics, and philosophy from the early twentieth century.
Third, there are papers on the empirical evidence for neural representation. John Krakauer and William Ramsey examine approaches to empirically testing for representation exemplified by, among others, Pohl et al. (2024). They argue that these authors overstate the explanatory and predictive benefits of positing neural representations; equivalent benefits can be gained by non-representational ‘causal mediator’ explanations. The standard of evidence for positing neural representations needs to be set higher. Krakauer and Ramsey argue it should, for example, depend on how the state gets ‘consumed’ in higher-level computational operations. Kenneth Aizawa suggests that authors like Pohl et al. (2024) ignore an important alternative empirical source of evidence for neural representations. Justification for positing neural representations need not rely on evidence of covariation. It may instead fall out of general epistemic practices already adopted across the brain sciences, namely, abduction that posits unseen and unmeasured states to explain observed behaviour. Along similar lines, Daniel Burnston and Tomás Ryan suggest that the debate about the empirical evidence for neural representation has gone astray because of its reliance on the assumption that single neurons ‘encode’ information in their spiking and that this is somehow fundamental to neural representation. According to Burnston and Ryan, talk of neural representation should instead be understood as a commitment to the existence of ‘latent structures’ that govern behavioural responses of the entire organism. These latent structures may take any number of physical forms, some of which may render issues about the responsivity of isolated single neurons largely irrelevant. Kevin Mitchell develops a parallel line of attack against the encoding conception of neural representation. He proposes that neuroscientists, instead of attempting to identity spiking activity with represented content, should rather identify certain relations between the organism and environment as pragmatically meaningful for the organism (e.g. as ‘good’ or ‘bad’ for its continued existence). Reparsing talk of neural representation in this way would allow the search for empirical evidence for neural representation to be recast as a set of more familiar and tractable questions about evidence for an organism’s adaptive behaviour, learning, and evolution.
Fourth, there are papers on wider metaphysical and issues about representation. Zina Ward raises the ‘job description challenge’ for directive representations (e.g. motor commands, imperatives). She asks what it takes for a directive state to be a representation rather than just a causal trigger that makes something happen. Her answer is that a directive representation is a state that brings about its outcome in a way that is, in a certain sense, decouplable from any one specific behaviour of the system. Eric Hochstein asks what it means to naturalise mental representation. Attempts to answer this often draw a distinction between two forms of naturalism: methodological naturalism (which aims to show that mental representations are a useful posit in natural science) and ontological naturalism (which aims to show that mental representations are somehow built out of natural, non-spooky ingredients). Hochstein argues that these two forms of naturalism, while conceptually distinct, are not independent; one naturalisation project cannot be pursued without the other. Nina Poth and Annika Schuster examine how the concept of representation varies across AI, cognitive science, and philosophy of science. They argue that while it is common to assume that the representations found in modern AI are of the same kind as those posited in cognitive science or in scientific modelling, the concept of representation employed in AI has unique characteristics – regarding semantic content, ability to misrepresent, and use – that mark it out as fundamentally different from those in the other two domains. Lotem Elber-Dorozko and Devin Gouvêa close the Special Issue with some potentially cheering news. The lack of clarity and ambiguity surrounding the concept of representation in neuroscience is often treated as a problem. Elber-Dorozko and Gouvêa argue that it should be viewed as a feature, not a bug. Lack of clarity about precisely which commitments neural representation talk entails offers neuroscientists a range of options to draw on when theory building. It would be precipitate for philosophers to agree on a single, fully articulated notion of representation in advance of knowledge of what neuroscientific explanations of cognition will ultimately look like.
These four groupings of papers emerged organically from the submissions we received. The clusters named above are far from a perfect way of grouping the papers – as one might expect, the papers crosscut these boundaries in various ways. There are also numerous contributions to debates about representation beyond those summarised here. For the full story, you have to read the papers.
A few papers still remain to join the Special Issue. We will not comment on them here, but please look out for them – they are good!