Roan Wisse Cognition
Contents

Theoretical research
along the Oude Rijn

The meaning-information relation is our point of entry into the architecture of cognition.

Scroll to explore
Coco, seated on a gallery bench with a blue hat beside him, admiring a framed illustration of a historic library.

Hello. You have found the door behind the bookcase.

The Roan Wisse Cognition Lab is a late-night laboratory of meaning and information, language and mathematics, intelligence and consciousness. Imagine a light still on somewhere along the Oude Rijn.

A grey cat walks past colorful Dutch canal houses beneath a crescent moon; a lamp glows in one upstairs window.

A system reliably responds to cats. What would make one of its internal states mean cat?

Suppose this state usually occurs when a cat is present. Sometimes, in poor light, a skunk causes the same state to occur. Is that a mistake? Or does the state represent a broader category?

These observations alone do not tell us which interpretation is right.

Informational semantics explores the possibility that a representation’s meaning can be understood in terms of the representation’s information-carrying relations to the world. The difficulty is to say which relations determine the representation’s content—and why other relations reflect misrepresentation, accident, or background association.

This issue arises whenever we ask whether a system represents something—and what it represents. It concerns not only what the system responds to but also what we claim when we say that one of its internal states has semantic content.

The Lab is especially interested in what we can learn by studying the relation between meaning and information, even before we know whether meaning can ultimately be reduced to information.

Coco rides a yellow bicycle through wind and rain, its headlight illuminating wet cobblestones and puddles.

The Oude Rijn is not just an old waterway. It is bicycle lights along winter water, wet cobblestones, lamplit leaded windows, university libraries after dark, and the Dutch practice of uitwaaien: walking or riding out into wind and weather until thought loosens and the mind clears. For the Lab, it evokes a way of seeing and thinking that lies just outside the Anglo-American philosophical mainstream.

What do we mean by “meaning”?

On this site, we use “meaning” as an intuitive working term for semantic content. Understanding when two concepts have the same content—and when they do not—is part of the research program.

How are concepts labeled here?

Throughout this site, uppercase labels such as “CAT,” “RABBIT,” and “CAUSE” name concepts. In a claim such as “CAT means cat,” the italicized lowercase expression indicates the semantic content of that concept. In statements about content, we use italicized lowercase expressions to indicate semantic content rather than to mention the corresponding English words.

“CAT” is only a convenient label. The representation itself need not take the form of a wordlike symbol. It may, for example, be an activation pattern or a feature vector in a human or artificial cognitive system.

For readers concerned with the type-token distinction: In this site’s terminology, a concept is a representational type. We also refer to these concepts as “representations.” An occurrence of a concept is a token: a particular instance of that concept in a cognitive system. Thus “CAT” names a concept, while an occurrence of CAT is a token of that concept.

Accordingly, when we use “activation pattern” here without qualification, we mean a type; a particular instance of that pattern is a token. Likewise, “internal state” here refers to a type; a particular instance of that state is a token. We preserve the distinction throughout but make it explicit only where needed.

The examples that follow are schematic. They are meant to make the Lab’s questions concrete and to motivate distinctions worth investigating—not to settle in advance any questions about concept identity, semantic change, or what distinguishes one content from another.

Informational semantics

“Informational semantics” refers to a family of theories developed and applied in cognitive science and artificial intelligence, with roots in analytic philosophy of mind. These theories treat concepts as representations. They seek to understand a concept’s content in terms of the information-carrying relations between occurrences of that concept and conditions in the environment.

Schematic illustration of a grey cat, a neural network with colored nodes, and a glowing output node.

Imagine a cat stepping into view and an activation pattern lighting up in a cognitive system. Think of that particular instance of the activation pattern as an occurrence of CAT.

The basic idea is that CAT means cat because occurrences of CAT vary systematically with the presence of cats or depend causally on the presence of cats—in short, because CAT reliably tracks cats.

The information-carrying relations of interest here include covariation and causal dependence between occurrences of a representation and conditions in the environment.

How do covariation and causal dependence differ?

Covariation and causal dependence are related but distinct. Covariation-based theories emphasize how occurrences of a representation vary reliably with conditions in the environment. Causal theories emphasize how those conditions bring about occurrences of the representation.

These approaches can be combined: A theory can appeal both to reliable covariation and to the causal processes that sustain that covariation.

Reductive informational semantics

Reductive informational semantics goes further. It aims to explain facts about meaning in terms of nonsemantic facts about information-carrying relations.

Suppose we say:

CAT means cat because it carries the information that there is a cat.

This does not yet reduce meaning to information if “carries the information that there is a cat” is itself understood as “correctly represents that there is a cat.” The account becomes reductive only if the relevant notion of information can be characterized without presupposing meaning—for example, in covariational, causal, counterfactual, or lawlike terms.

That is why reductive informational theorists have explored notions such as reliable covariation, causal channels, and laws of nature.

But a central problem appears almost immediately: A representation can stand in information-carrying relations to more than one condition.

A skunk in a tree’s shadow, a neural network with colored nodes, and a glowing output node.

Suppose cats usually trigger occurrences of CAT, but skunks seen in poor light sometimes do so as well. What does CAT mean?

  • cat
  • cat-or-skunk-in-poor-light

Or suppose rabbits usually trigger occurrences of RABBIT. Whenever a rabbit is present, its ears, paws, tail, and other parts are also present. Philosophers call such parts “undetached rabbit parts” because they have not been separated from the rabbit. Occurrences of RABBIT therefore covary not only with the presence of rabbits but also with the presence of undetached rabbit parts.

What does RABBIT mean?

  • rabbit
  • undetached-rabbit-part

The difficulty is not that we lack correlations. We have too many.

CAT stands in information-carrying relations to the presence of cats and, in poor light, to the presence of skunks. RABBIT stands in information-carrying relations not only to the presence of rabbits but also to the presence of their undetached parts.

The skunk makes trouble because it sometimes produces an occurrence of CAT. The rabbit’s parts make trouble because they are present whenever the rabbit is.

The problem is to explain which of a representation’s information-carrying relations determine its content—and why.

A grey cat, a puppy in a cat costume, and a skunk in the shade of a tree, sitting together in a meadow.

The CAT case illustrates the disjunction problem. A theory must explain how CAT could mean cat rather than cat-or-puppy-dressed-up-as-a-cat-or-skunk-in-poor-light-or . . .—an open-ended disjunction of conditions that produce occurrences of CAT.

Two white rabbits in a meadow; a golden outline highlights one whole rabbit and a paw of the other.

The RABBIT case illustrates a form of inscrutability of reference. The information-carrying relations described in this example support more than one interpretation of what RABBIT refers to.

Merely listing the conditions that covary with or cause occurrences of a concept does not yet tell us which of those conditions the concept represents.

That is not a reason to abandon the informational project. It tells us where the harder work begins.

Why “inscrutability of reference”?

In Word and Object, W. V. O. Quine (1960) used the “gavagai” example to argue that the same observational evidence can support multiple interpretations of what our expressions refer to without singling out a unique interpretation.

The problem arises here in a different theoretical setting, but the family resemblance matters: Information-carrying relations may fail to determine a unique content.

Misrepresentation

Solving the disjunction problem requires an account of misrepresentation. Return to the skunk in poor light.

A skunk in a tree’s shadow is shown as input to a neural network, with a cat’s face at the glowing output.

Calling the resulting occurrence of CAT a misrepresentation already assumes that CAT has content against which that occurrence can be assessed. What could determine that content without presupposing the distinction between correct and incorrect occurrences?

Misrepresentation is commonplace in cognition, but it introduces a distinction that a list of triggering conditions does not explain. Causal processes simply occur; representations can get things right or wrong.

A simple covariation-based account therefore faces a problem. If both cats and, in poor light, skunks produce occurrences of CAT, why should the skunk case count as a misrepresentation rather than as evidence that CAT means cat-or-skunk-in-poor-light?

Informational theorists have proposed several ways of answering this question. These proposals appeal to factors such as learning history, biological function, asymmetric dependence, or characteristic use.

The reductive hope is that such factors can distinguish the relations that determine content from those that reflect misrepresentation, accident, or background association, without presupposing the semantic distinction the theory is meant to explain.

One enduring contribution of informational semantics is to show that a theory of meaning must explain not only successful representation but also the possibility of getting the world wrong.

The reductive picture has two explanatory levels.

At the nonsemantic informational level are relations such as covariation and causal dependence, which the reductive program seeks to characterize without presupposing semantic content.

At the semantic level are representations with content—representations whose occurrences can be correct or incorrect and whose content can remain stable or change.

The reductive task is to show how facts at the nonsemantic informational level determine facts at the semantic level.

That is an ambitious research program. The reductive picture also gives the explanation a direction: from information to meaning.

Can these additional resources themselves be characterized nonsemantically?

That is part of the dispute. Critics of reductive accounts have argued that appeals to function, learning history, normal conditions, or characteristic use presuppose the very semantic or normative distinctions the theory is supposed to explain.

The Lab does not settle this question in advance. Even when attempts at reduction fail, it is instructive to analyze why they fail.

Relational informational semantics

One question is whether—and how—information-carrying relations determine what a representation means. But we can also ask how meaning and information vary together.

Here and below, when discussing the meaning-information relation, we use “information” as shorthand for a representation’s informational profile: the structured configuration of its information-carrying relations.

Some information-carrying relations within this profile may be covariational. Some of these covariational relations can be characterized using Shannon-theoretic measures. The profile as a whole, however, is not captured by any single scalar quantity.

Which changes in a concept’s information-carrying relations leave its meaning intact? Which changes would be accompanied by semantic change? And how might a system’s existing concepts shape the information-carrying relations it goes on to develop?

We use the label “relational informational semantics” for this research program.

The shift is methodological; it does not commit us to either accepting or rejecting reduction. We do not assume that the reduction of meaning to information is impossible. We set the question of reduction aside while we study the meaning-information relation itself.

The label therefore names a research program about the relation between semantic content and informational structure; we do not assume that a representation’s informational profile, by itself, uniquely determines its content.

Coco stands near a demonstration table during a thermodynamics lecture in a crowded historic lecture hall.

Thermodynamics offers a useful analogy. We can ask how to explain thermodynamic quantities in lower-level terms, but we can also study systematic relations among distinct state variables such as temperature, pressure, and volume.

Temperature, pressure, and volume are distinct thermodynamic variables; describing the relations among them does not amount to reducing one variable to another. Nevertheless, for a fixed amount of an ideal gas, these variables stand in stable mathematical relations. Holding one variable constant constrains how the other two can vary together. Those relations can be studied at the thermodynamic level without first explaining them in lower-level terms.

At the same time, the relations among these thermodynamic variables can themselves be explained at a deeper level. Statistical mechanics helps explain how thermodynamic regularities arise from lower-level molecular and statistical properties. The deeper explanation therefore need not take the form of reducing one thermodynamic variable to another. It can instead explain why the regularities hold.

This suggests a corresponding possibility for meaning and information.

A reductive informational theory asks whether meaning can ultimately be reduced to information. In that project, explanation runs from informational facts to semantic facts: Meaning is the phenomenon to be explained, and information provides the nonsemantic explanatory resources.

Relational informational semantics asks a different question. For the purpose of this analysis, we treat a concept’s meaning and its informational profile as distinct theoretical variables. We then investigate whether these variables stand in a relatively stable and systematic relation. We examine which changes in a concept’s information-carrying relations leave its meaning intact, which changes are associated with semantic change, and how a system’s existing concepts shape the information-carrying relations it goes on to develop.

The thermodynamic analogy suggests that the meaning-information relation may itself have a deeper explanation. Underlying neural, computational, or physical mechanisms may explain any regularities we find linking meaning and information.

The resulting explanatory picture is layered. At the level of the meaning-information relation, we characterize the relation between semantic content and informational profiles. At the mechanistic level, we investigate the mechanisms and organizing principles that give rise to that relation. The mechanistic explanation need not reduce meaning to information or information to meaning. It may instead explain why the meaning-information relation has the form it does.

There are two claims here. First, a systematic relation between XX and YY does not imply that XX reduces to YY. Second, the relation itself may call for explanation. We treat meaning and information as distinct variables at this level of analysis. Any systematic relation between them could still be explained by mechanisms and organizing principles of the underlying cognitive system as it interacts with its environment.

Nor do we need that mechanistic explanation before studying the meaning-information relation. Thermodynamic regularities could be discovered, measured, and modeled before they were explained by statistical mechanics. In the same way, we may be able to identify regularities connecting meaning and information before we know which underlying neural or computational mechanisms explain them.

This leaves open the possibility of eventually explaining meaning, information, or both in lower-level terms. Relational informational semantics is compatible with that possibility. Its methodological claim is narrower: Explaining a relation between two variables does not require reducing one of those variables to the other.

The contrast, then, is not between accepting and rejecting reduction. It is between two distinct but compatible explanatory projects. One asks whether meaning can be reduced to information. The other asks what systematic relation holds between meaning and information—and what underlying mechanisms and organizing principles explain why that relation has the form it does.

Coco, wearing a blue hat and red coat, walks to the right.

Semantic stability, change, and reciprocity

The layered picture gives us a way to frame the next questions. If we treat meaning and information as distinct theoretical variables, what kind of systematic relation might hold between them?

We can begin by looking for regularities in how they vary together. Some changes in a concept’s information-carrying relations may leave its meaning unchanged. Others may be associated with changes in meaning. The relation may also be reciprocal: A system’s existing concepts may help shape the information-carrying relations it goes on to develop.

Semantic stability, semantic change, and reciprocity provide three starting points for investigating such regularities.

Semantic stability

One phenomenon to investigate is semantic stability: A concept’s information-carrying relations can change while its meaning remains the same.

Suppose someone has the concept MAMMAL but does not recognize whales as mammals. Learning that whales are mammals may change which encounters trigger occurrences of MAMMAL without changing what MAMMAL means.

Coco sits on the back of a blue whale, looking up at yellow stars above the sea.

When this person begins to classify whales as mammals, what would make this a case of learning without a change in MAMMAL’s meaning?

That is the distinction we want to understand: What distinguishes learning more about a category from changing the meaning of the concept used to represent that category?

Studying semantic stability helps us understand how a concept can retain its meaning through learning, disagreement, error, scientific revision, and cognitive development.

A river’s course can stay the same even as the water flowing through it changes. The harder question is when the course itself has changed.

Semantic change

A second phenomenon to investigate is semantic change: Some changes in a concept’s information-carrying relations may be accompanied by changes in what that concept means.

Categories may split or merge. Concepts may be reorganized as their relations to other concepts change. A cognitive system may discover that a property once treated as essential is incidental or that phenomena once treated as unrelated share a common structure.

Which changes alter what a concept means rather than merely what the system believes, how it classifies cases, or what it infers?

Semantic stability and semantic change pose a joint problem. We want to understand which changes in information-carrying relations leave a concept’s meaning unchanged and which are accompanied by a change in meaning. We also ask whether some changes are better understood as giving rise to a different concept altogether.

Characterizing these distinctions would be an important step in mapping the systematic relation between meaning and information.

Reciprocity

So far, we have approached stability and change from one direction: Information-carrying relations change, and we ask what happens to meaning.

But the meaning-information relation may not be one-way.

Coco at his desk, drawing a neural network with colored pencils in front of bookshelves.

Our working hypothesis is that the relation between meaning and information is reciprocal: Information constrains meaning, and meaning, in turn, organizes information.

Saying that meaning organizes information does not imply that meaning is a separate causal force. Rather, a system’s cognitive mechanisms and organizing principles—including those involved in using concepts—help shape how the system processes input from its environment. This processing, in turn, influences the information-carrying relations the system develops.

A system’s existing concepts can shape which conditions it registers in its environment and which of those conditions it encodes in its representations. These concepts can also influence the information-carrying relations the system develops and exploits in subsequent cognition. The cognitive relevance of those relations depends on both the system’s organization and its environment. How the system represents what it encounters can therefore affect what it attends to, what similarities and differences it detects, what it finds surprising, and what it goes on to learn.

In a given task, a system may use the concept PREDATOR rather than PREY to categorize an animal that both hunts and is hunted. Using one categorization rather than the other may affect what the system attends to, expects, and learns next. How might such differences reshape the information-carrying relations the system develops?

On this hypothesis, meaning does not simply sit at the end of an informational pipeline. The concepts available to a system help shape which information-carrying relations it develops, how those relations are structured within the informational profiles of its representations, and which relations it exploits.

Information constrains meaning; meaning organizes information.

Learning from idealized cases

Thought experiments can function as simple conceptual models: They let us explore ideas about meaning and information in deliberately minimal settings.

A girl peeks through a door concealed in a bookcase into a reading room with two white rabbits, one on the floor and one on a green armchair.

We can then compare the models’ implications with data from real-world cases, observations of research practices in cognitive science and AI, and findings from empirical studies of human and artificial cognition.

The aim is to isolate candidate regularities in the meaning-information relation. We vary one aspect of the informational structure while holding other aspects fixed and ask what happens to the representation’s content.

In artificial systems, some of these variations may be implemented directly by changing training environments, sensory channels, representational formats, or task demands. We can then examine what would count as evidence of semantic stability or change in the resulting representations.

One approach we are exploring is to construct ideal concepts as comparison objects for studying concepts in real cognitive systems.

Here, “ideal” means idealized. An ideal concept is a theoretical idealization constructed by abstracting away from selected complications to make the relations of interest easier to investigate, much as an ideal gas model omits certain properties of real gases. The point is to learn from both similarities and differences between ideal concepts and real concepts, not to treat ideal concepts as standards that real concepts ought to meet.

One natural starting point is noise in the channels connecting a cognitive system with its environment. Imagine an artificial neural network operating in a virtual environment. All rabbits in the environment share the same relevant properties. These properties are encoded in a long binary feature vector that is identical for every rabbit. We call this sequence of 0s and 1s the “distal vector”—the message before transmission through a noisy channel. The network encounters at most one rabbit at a time. Its task is to detect whether a rabbit is present.

The neural network receives either a vector resulting from noisy transmission of the distal vector (in the rabbit-present condition) or a uniformly random binary vector of the same length (in the rabbit-absent condition). In both conditions, we call the received vector the “proximal input.”

For this thought experiment, we use features that encode properties of undetached rabbit parts, such as their edge structure, two-dimensional shape, and texture. We leave aside color, depth, and motion.

To generate a rabbit-present input, we model channel noise by flipping individual feature values independently (from 0 to 1 or from 1 to 0) during transmission of the distal vector. Each value is flipped with the same fixed probability, called the “corruption rate.” We set this rate below 0.5.

To generate a rabbit-absent input, we set each feature value independently to 0 or 1, with probability 0.5 for either value.

Under these assumptions, rabbit-present inputs tend to retain a nonaccidental resemblance to the distal vector. By contrast, rabbit-absent inputs do not result from transmission of that vector, so any resemblance to it is purely coincidental.

The behavior described below could result from supervised training on examples from both input conditions, with labels indicating which condition generated each input. Here we specify and analyze a simple triggering rule that a network trained in this way could approximate. We do not model the learning process itself.

In this simplified model, the network produces an occurrence of the internal state labeled RABBIT when—and only when—at least a specified proportion of the proximal input’s feature values match the corresponding values in the distal vector. The distal vector is the fixed reference pattern for this rule, not an additional input to the network during detection.

We set this threshold above the expected proportion of matches in rabbit-absent inputs (0.5) but below the expected proportion of matches in rabbit-present inputs (1 minus the corruption rate).

Different combinations of matching feature values can trigger an occurrence of RABBIT, so detection can succeed even when some feature values are corrupted. This overlapping support for detection is what we mean here by “redundancy.”

Now imagine otherwise identical versions of the neural network that apply the same triggering rule to progressively longer feature vectors that provide more redundant support. For each version, we specify a correspondingly longer distal vector that encodes the relevant properties common to all rabbits. The corruption rate and the threshold proportion of matching feature values remain fixed.

As the feature vectors grow longer, the distribution of match proportions in each condition becomes more tightly concentrated around its expected value. A missed detection occurs when the proportion of matches is below the threshold in a rabbit-present input; a false alarm occurs when the proportion of matches meets or exceeds the threshold in a rabbit-absent input. Although the probabilities of these errors need not decrease with every added feature, both probabilities approach zero as the number of features tends to infinity.

We take this unattainable zero-error limit to be the ideal case. RABBIT in this limiting system serves as our ideal concept.

As the finite feature vectors grow longer, tracking approaches the perfect reliability of the idealized case without ever reaching it. The ideal concept serves as a comparison object for studying a real concept’s tracking reliability. The real and the ideal are a duo.

We can then ask whether the resulting informational structure uniquely determines content. Idealized cases may help reveal exactly which questions that structure settles—and which it leaves open.

At this idealized limit, the probabilities of missed detections and false alarms are both zero. A rabbit and its undetached parts, however, are still present together. Would a zero probability of detection error in this model, by itself, settle whether RABBIT means rabbit or undetached-rabbit-part?

Does the way the network combines these feature values to achieve reliable tracking also help determine the representation’s content?

Reliable tracking and determinate content may be distinct achievements—but they may still be related. Idealization may help reveal which aspects of informational structure are semantically relevant and which are not.

Other forms of idealization may prove useful as the research develops.

Why “comparison objects”?

In Philosophical Investigations, Ludwig Wittgenstein (1953, §§ 130–31) describes simple language-games as “objects of comparison” (Vergleichsobjekte). They illuminate our language through differences as well as similarities. He compares such an object to a measuring rod: something we use to examine reality, not a preconception to which reality must conform.

We borrow this methodological idea for our ideal concepts. They are tools for understanding real cognitive systems, not templates those systems must fit. Differences between an ideal concept and a real concept can be as informative as similarities.

We combine this comparative approach with scientific idealization. Ideal concepts may be constructed by abstracting away from selected complications, as in an ideal gas model. The Wittgensteinian point we borrow concerns how we use these idealizations as comparison objects.

Could capsule networks extend this comparative approach?

A suitably configured capsule network could implement finite versions of the model above. A capsule-based extension could also provide a deliberately richer comparison model by introducing explicit part-whole relations into the way feature information is combined (Hinton et al. 2018).

We could compare this richer model with one that uses the simple triggering rule and achieves similar tracking reliability. This comparison could help us investigate whether part-whole organization contributes something semantically relevant beyond reliable detection. The aim would be to examine whether—and how—such organization helps constrain RABBIT’s content, without assuming that part-whole organization uniquely determines that content.

Could one mathematical description have two interpretations?

Capsule networks suggest another way to use these comparison models. A mathematical description of rabbit detection could admit two ontologically distinct interpretations: a cognitive interpretation, in terms of mental states and processes, and a neuronal interpretation, in terms of physical states and processes.

A matrix transformation, for example, could be interpreted cognitively as a process that uses a representation of a part to estimate the configuration of a whole. Under a neuronal interpretation, the same transformation could describe how a circuit transforms a pattern of neuronal population activity. The mathematical operation would be the same under both interpretations, while its terms would refer to different kinds of things.

We can make the example concrete with V=MWV = MW, where MM is a pose matrix describing a part’s configuration, WW is a part-whole transformation matrix, and VV is a predicted pose matrix, or vote, for a whole (Hinton et al. 2018). In our comparison model, an estimated ear position and a specified ear-to-rabbit relationship yield an estimated position of the rabbit’s center. Cognitively, this is a transformation from the parameters of a representation of a part to those of a prediction about a whole. Neuronally, it is a transformation from specified input population coordinates to output population coordinates. The vote is an intermediate prediction, not yet a detection.

In this comparison model, we keep the two ontologies neatly separated. A cognitive interpretation is not simply a neuronal interpretation with different labels. Nor does sharing a mathematical description identify a mental state with a physical state. Treating an item from one ontology as though it belonged to the other is a category mistake within this comparison model.

The alignment between the interpretations is mathematical, not causal. Both interpretations must respect the relations and transformations specified by the model. Causal relations may figure within an interpretation, but the correspondence between the interpretations is not itself a causal connection.

The compatibility requirement is straightforward. Applying the interpreted process and then describing its output mathematically must give the same result as describing its input mathematically and applying the model’s transformation. This requirement applies under each interpretation. The accompanying technical note develops a small example in which it can be checked explicitly.

This offers a tool for exploring the mind-body problem. We can vary a relation or operation in the model and ask what that variation amounts to under each interpretation. Which variations still admit both interpretations? Where does either interpretation become strained? Whether the model adequately describes actual cognitive or neuronal processes remains a further question.

The aim is not to solve the mind-body problem but to offer a comparison object: something modest enough to inspect and vary. We can compare this arrangement of one mathematical description and two ontologically distinct interpretations with positions and approaches such as dualism, materialism, psychophysical parallelism, epiphenomenalism, double-aspect theory, functionalism, and anomalous monism, looking for similarities and differences, analogies and disanalogies.

As a further variation, the physical interpretation might describe silicon-based circuits instead of neuronal circuits, while the cognitive interpretation remains the same. How would this possibility relate to traditional arguments from multiple realizability?

Read the worked technical note

The mind’s relational architecture

Here the river widens into the delta.

The Lab’s broader ambition is to study the relational architecture of cognition: how meaning, information, language, mathematics, representation, computation, intelligence, consciousness, technology, and human experience relate to one another.

Coco and Pippa standing together, with Pippa holding a white rabbit.

The meaning-information relation is our point of entry.

Rather than assuming that any one of these phenomena reduces to another, we ask first what relations hold among them and what those relations can explain about cognition.

Some of these relations may eventually support reductive explanations. Some may also prove illuminating when studied in their own right.

The aim is not to presuppose the architecture but to make enough of it visible to isolate, model, and compare specific relations.

The meaning-information relation can therefore serve as a first case study for a broader strategy: Characterize relations among these phenomena and then ask what underlying organization explains the forms those relations take.

Could consciousness help make semantic content determinate?

If information-carrying relations leave open several interpretations of a representation’s content, would considering conscious thought or experience narrow the alternatives? John Searle (1992) argues that intrinsic intentionality—the aboutness a mental state has in its own right, rather than merely through an interpreter’s attribution—depends on actual or potential consciousness. In resisting Quinean indeterminacy, he assigns an important role to the way things are represented from the subject’s point of view.

Saul Kripke’s skeptical reading of Wittgenstein’s discussion of rule-following raises a different difficulty. On this reading, a conscious image, a formula held in mind, or a distinctive feeling of understanding does not by itself determine which rule someone intends to follow or what counts as following it correctly (Kripke 1982). What would make the conscious state itself have one content rather than another?

G. P. Baker and P. M. S. Hacker (1984) dispute this skeptical reading of Wittgenstein. On their reading, the failure to find a private mental occurrence that independently fixes meaning does not show that meaning is indeterminate. They examine the techniques and standards involved in meaningful use rather than seek an inner occurrence that fixes meaning independently of those practices.

The Lab does not settle these questions in advance. We ask how conscious experience, information-carrying relations, and criteria of correct use are connected. When information-carrying relations support multiple interpretations of a representation’s content, does this reveal semantic indeterminacy in the cognitive system, a limitation of our understanding, or a problem with what we require of an explanation of meaning?

The Lab is at an early stage. We are making ground—the polder way.

A sunny Dutch polder landscape with green fields, water channels, a windmill, and a farmhouse.

A polder—land enclosed by dikes and kept dry through water management—is neither discovered nor simply declared into existence. Building and sustaining one require boundaries and channels, ongoing maintenance, cooperation, and an accurate sense of what the water will do next.

We welcome your thoughts and questions.

Email Roan Wisse at roanwisse@roanwisse.com.

Thematic reference list and further reading

This reference list brings together philosophical and scientific work relevant to the Lab’s inquiry into meaning and information. It includes theories of mental representation and critiques of those theories, discussions of reference and consciousness, and work on mechanistic explanation, scientific idealization, informational redundancy, and neural error correction. These works offer theoretical background, methods, and points of comparison for the questions explored on this site.

The works are grouped by their role in the inquiry, with short overviews introducing each theme. Each work appears once, although several are relevant to more than one theme. Within each section, entries are alphabetized by author. The reference list offers different points of entry into the literature; it is not an exhaustive survey, and the works express a variety of positions.

1. Informational semantics and mental representation

What makes an internal state a representation of something in the world? These works cover causal and covariational approaches. They provide starting points for examining how a representation’s relations to its environment might determine what it means and how to distinguish the relations that determine content from those involved in misrepresentation or incidental association.

  • Artiga, Marc, and Miguel Ángel Sebastián. 2020. “Informational Theories of Content and Mental Representation.” Review of Philosophy and Psychology 11 (3): 613–27. https://doi.org/10.1007/s13164-018-0408-1.
  • Boghossian, Paul A. 1991. “Naturalizing Content.” In Meaning in Mind: Fodor and His Critics, edited by Barry Loewer and Georges Rey, 65–86. Blackwell.
  • Buckner, Cameron. 2022. “A Forward-Looking Theory of Content.” Ergo: An Open Access Journal of Philosophy 8 (37): 367–401. https://doi.org/10.3998/ergo.2238.
  • Buckner, Cameron. 2023. “Cameron Buckner—‘A Forward-Looking Theory of Content.’” Ergo Blog, April 30. https://ergoblog.org/cameron-buckner-a-forward-looking-theory-of-content/.
  • Cohen, Jonathan, and Aaron Meskin. 2006. “An Objective Counterfactual Theory of Information.” Australasian Journal of Philosophy 84 (3): 333–52.
  • Dretske, Fred. 1981. Knowledge and the Flow of Information. MIT Press.
  • Dretske, Fred. 1986. “Misrepresentation.” In Belief: Form, Content, and Function, edited by Radu J. Bogdan, 17–36. Clarendon Press.
  • Dretske, Fred. 1988. Explaining Behavior: Reasons in a World of Causes. MIT Press.
  • Eliasmith, Chris. 2000. “How Neurons Mean: A Neurocomputational Theory of Representational Content.” PhD diss., Washington University in St. Louis.
  • Eliasmith, Chris. 2005a. “Neurosemantics and Categories.” In Handbook of Categorization in Cognitive Science, edited by Henri Cohen and Claire Lefebvre, 1035–54. Elsevier.
  • Eliasmith, Chris. 2005b. “A New Perspective on Representational Problems.” Journal of Cognitive Science 6 (2): 97–123.
  • Fodor, Jerry A. 1984. “Semantics, Wisconsin Style.” Synthese 59 (3): 231–50. https://doi.org/10.1007/BF00869335.
  • Fodor, Jerry A. 1987. Psychosemantics: The Problem of Meaning in the Philosophy of Mind. MIT Press.
  • Fodor, Jerry A. 1990. “A Theory of Content” and Other Essays. MIT Press.
  • Fodor, Jerry A. 1994. The Elm and the Expert: Mentalese and Its Semantics. MIT Press.
  • Fodor, Jerry A. 1998. Concepts: Where Cognitive Science Went Wrong. Oxford University Press.
  • Godfrey-Smith, Peter. 1989. “Misinformation.” Canadian Journal of Philosophy 19 (4): 533–50. https://doi.org/10.1080/00455091.1989.10716781.
  • Godfrey-Smith, Peter. 1992. “Indication and Adaptation.” Synthese 92 (2): 283–312. https://doi.org/10.1007/BF00414302.
  • Landrum, Kyle. 2025. “Good Explanations and the Causal Metasemantics of Inductive Concepts.” Mind 134 (535): 735–57. https://doi.org/10.1093/mind/fzaf024.
  • Loewer, Barry. 1987. “From Information to Intentionality.” Synthese 70 (2): 287–317. https://doi.org/10.1007/BF00413940.
  • Luzon, Bar. 2024. “Explaining Content.” PhD diss., New York University.
  • Neander, Karen. 2017. A Mark of the Mental: In Defense of Informational Teleosemantics. MIT Press.
  • Roche, William, and Elliott Sober. 2021. “Disjunction and Distality: The Hard Problem for Purely Probabilistic Causal Theories of Mental Content.” Synthese 198 (8): 7197–230.
  • Rupert, Robert D. 1996. “The Best Test Theory of Extension.” PhD diss., University of Illinois at Chicago.
  • Rupert, Robert D. 1999. “The Best Test Theory of Extension: First Principle(s).” Mind & Language 14 (3): 321–55.
  • Rupert, Robert D. 2001. “Coining Terms in the Language of Thought: Innateness, Emergence, and the LOT of Cummins’s Argument Against the Causal Theory of Mental Content.” The Journal of Philosophy 98 (10): 499–530.
  • Scarantino, Andrea. 2015. “Information as a Probabilistic Difference Maker.” Australasian Journal of Philosophy 93 (3): 419–43.
  • Scarantino, Andrea, and Gualtiero Piccinini. 2010. “Information Without Truth.” Metaphilosophy 41 (3): 313–30.
  • Skyrms, Brian. 2010a. “The Flow of Information in Signaling Games.” Philosophical Studies 147 (1): 155–65.
  • Skyrms, Brian. 2010b. Signals: Evolution, Learning, and Information. Oxford University Press.
  • Stampe, Dennis W. 1977. “Toward a Causal Theory of Linguistic Representation.” Midwest Studies in Philosophy 2 (1): 42–63. https://doi.org/10.1111/j.1475-4975.1977.tb00027.x.
  • Usher, Marius. 2001. “A Statistical Referential Theory of Content: Using Information Theory to Account for Misrepresentation.” Mind & Language 16 (3): 311–34.
  • Weissglass, D. E. 2020. “Greatest Surprise Reduction Semantics: An Information Theoretic Solution to Misrepresentation and Disjunction.” Philosophical Studies 177 (8): 2185–205. https://doi.org/10.1007/s11098-019-01305-0.

2. Reference, rule-following, and semantic determinacy

These works frame questions about what fixes reference and what makes an application of a concept or rule correct. Quine’s discussion of reference and the disagreements surrounding Kripke’s reading of Wittgenstein provide distinct points of entry rather than a single argument for indeterminacy. Wittgenstein also supplies the comparison-object idea used in the Lab’s idealization section, especially in Philosophical Investigations §§ 130–31.

  • Baker, G. P., and P. M. S. Hacker. 1984. “On Misunderstanding Wittgenstein: Kripke’s Private Language Argument.” Synthese 58 (3): 407–50. https://doi.org/10.1007/BF00485249.
  • Hacker, P. M. S. 2001. Wittgenstein: Connections and Controversies. Clarendon Press. https://doi.org/10.1093/019924569X.001.0001.
  • Kripke, Saul A. 1982. Wittgenstein on Rules and Private Language: An Elementary Exposition. Harvard University Press.
  • Quine, Willard Van Orman. 1960. Word and Object. MIT Press.
  • Wittgenstein, Ludwig. 1953. Philosophical Investigations. Translated by G. E. M. Anscombe. Basil Blackwell.

3. Consciousness and intentionality

This section places consciousness alongside questions about representation and intentionality—the aboutness of mental states. Dretske and Searle provide different approaches to these connections. For the Lab, the guiding question is whether considering conscious experience helps explain what determines a representation’s content and how that contribution relates to information and cognitive organization.

  • Dretske, Fred. 1995. Naturalizing the Mind. MIT Press.
  • Searle, John R. 1992. The Rediscovery of the Mind. MIT Press.

4. Mechanistic explanation and scientific idealization

Craver supplies background on multilevel mechanistic explanation; Weisberg supplies background on the aims and uses of scientific idealization. These are methodological resources for the Lab’s relational approach and construction of ideal concepts, not sources for those proposals themselves. The complementary idea of using ideal concepts as comparison objects draws on Wittgenstein’s Philosophical Investigations, listed in section 2.

5. Redundant, unique, and synergistic information

Williams and Beer provide a formal framework for distinguishing three kinds of information about a target variable: information shared by sources, information supplied uniquely by a source, and information available only when sources are considered jointly. This framework provides background for clarifying what is meant by “overlapping support” in the Lab’s feature-vector example. Its role here is to inform the analysis of informational structure rather than to serve as a theory of semantic content.

6. Neural networks, denoising, and error correction

These papers provide technical examples of pattern recovery, denoising, redundant representation, and fault-tolerant computation in neural networks. They connect the Lab’s thought experiment with mechanisms that compensate for incomplete or unreliable signals and, in some cases, failures within the computing system itself. The capsule-network papers also provide resources for developing richer comparison models that incorporate part-whole organization. Together, these works inform the construction and comparison of artificial systems; what the organization of those systems contributes to semantic content remains the Lab’s question.