Theorizing Thinking - What is Cognition?
Tracing the mental map
Hello!
This is the second part in a series of posts exploring common definitions and framings of concepts in philosophy of mind. These posts will be generalist, mostly, though when particularly appropriate I’ll be specifying when we’re discussing the application of these concepts to only humans, to specific non-humans, etc.
In my first entry to this series, we discussed the mainstream lineup of theories on consciousness, working through phenomenalism, cognitive access, self-property, functionalism, higher-order representation, biological primacy, panpsychism, and illusionism. These theories resist compatibility while ultimately devolving into either circular logical reasoning and question-begging, or (in the case of panpsychism and higher-order representationalism) the potential lack of productive philosophical work. I proposed a processual perspective that gladly admits its own unfalsifiability and instead focuses on a pragmatism that aims for virtuous, instead of vicious, circularity.
Now, the sun-drenched, parched survivor, emerging from the wreckage of consciousness discourse, at least sees that we must now talk about cognition – “Surely, there must be joy in thinking!” they think to themselves. “Nothing could be as opaque now. Time to head to that distant oasis and drink all the delicious water.”
Sadly, we will be wandering for a while yet. A veritable dust storm of cognitive theories now presents itself to us:
- Classical Computationalism, in which thinking is an elaborate process of symbol manipulation. In this sense, cognition is the “language” of thought, providing the framework by which symbols interface and combine. A famous analogy that represents this is that of the brain as a sophisticated computer.
- Connectionism, in which cognition is a marvelous web of distributed pattern recognition within neural nets.
- Embodied Cognition, in which sensorimotor engagement with the environment forms a mutual flow of interconnected data, input-output, and temporalized context that thinking sprouts from.
- Enactivism, closely related and partially overlapped to embodied cognition, wherein cognition is active sense-making and meaning-making through autonomous engagement.
- Extended Cognition, in which cognition isn’t only something you, a human, does – it extends into the environment and through tools, meaning that when you use a calculator, the calculator isn’t separate from the thinking itself.
- Predictive Processing, in which cognition is a sophisticated and hierarchical process of minimizing error between our predictions and observations of reality.
- 4E Cognition – Embodied, Embedded, Extended, Enactive – that attempts to tie several positions together.
We’ll start our little walk-through with computationalism, solely because it’s the dominant perspective in a gut-feeling sort of sense with most everyday people.
Computationalism: Thinking As Rule-Bound
Much of academic cognitive science is, implicitly or explicitly, quite computationalist. It’s quite attractive and practical for such a demographic because it frames cognition as a framework of formal rules, quite akin to syntax and semantics – a “mentalese” of internal symbols that’s overlaid with algorithms to produce various connections, contexts and meanings. Our senses and perception provide the symbols which our cognition processes until stimulating or producing an action. This theory also neatly ties together the interactions of both humans and computers and identifies thought in a very concrete way as a “thing,” or at least an easily traceable process.
As a result, computationalism has the advantage of being pretty intuitive once we get past the initial confusion of “Does this mean the symbols are real, or not? And what’s doing the manipulating?” Once you’ve gotten past that confusion is also when you’ve become decidedly overconfident – because they’re vital problems in two distinctly different ways.
The first one is the problem of symbolic grounding – how do symbols become symbols, how do they develop meaning? Computationalism doesn’t insist on any ultimate source that symbols spring from. When expressed in a formal way, symbols are necessarily relational in character – this is an intuitive enough idea that it’s almost tautological. However, if there’s no site, no “ground” that provides this meaning, we can dig our way through symbols without ever really being certain what a symbol is properly about.
For instance: imagine a dictionary. You thumb your way to “human” on the dictionary. The dictionary tells you a human is an ape-like mammal. So you look up ape, and then hominid, and then mammal, and then mammalian, mammaliaform, warm-blooded, synapsid... you construct this incredible diagram of symbols that describe a human from different angles, but the conclusion ultimately creeps up on you that you’re just shuffling terms around in place of doing any actual semantic work. This suggests to the skeptic that some ground must be found.
There are, of course, rejoinders to this. The main defense incorporates informational semantics in the style of Jerry Fodor and Fred Dretske. In this revision to computationalism, symbols are reliably and causally connected to existent objects. “Human” ends up meaning human because it’s consistently and reliably tokened out, or symbolically mediated, in response to real piles of flesh and blood and muscles and bones. In this way, it’s really about connecting external referents to internal states with an attempt at a direct, causal mechanism, which means it sidesteps having to think about embodied acts or separate humans from computers. And then there are rejoinders to the rejoinders: How can false beliefs ever happen this way? Which specific symbol wins out in fuzzier contexts, like when a “dog” impression emerges for a wolf or a picture of a coyote? What does it mean to say things are “causally connected” in these ways beyond mere assertion? As a results, most computationalists are no longer classical or formal about their computationalism; most admit a need for sensorimotor grounding when it comes to perception, and place anything more abstract or higher in order within the computational bucket.
The second question that so vexed us is called the homunculus problem. Basically: if cognition is symbol manipulation, who or what is doing the manipulating? It seems as if there must necessarily be an actor, agent, or puppet master pulling the cognitive strings. But then someone must be directing that... and that... until we’re trapped in a whorl of infinite puppets and strings.
There are a couple genuinely quite clever responses to this problem, both of which are maintained by Daniel Dennett. Dennett looks at the homunculi being generated and sees hierarchical decomposition, in which what one really has is more like a tall building, with committees on each floor. Each floor has instruments and readouts that seem to correlate to the floor below to various degrees, but the “committee members” – various cognitive processes within that metacognitive tier – can never be sure of a direct association. Even worse, each committee gets dumber and dumber each floor down, until the lowest committees so far known are just physics and beams, mechanical causal processes and neurons. Dennett labels this competence without comprehension. As a tidy illusionist, Dennett thus can deny that understanding “comes from somewhere” in the sense of some ultimate ground.
Dennett’s own favorite example for all this, intriguingly enough, is Large Language Models (LLMs). Regardless of one’s position on their pragmatic benefit when used as a tool for humans, we can admit that they as of now express serious competence and facility at language and language and language tasks. They can construct complex arrays of stylistically and semantically coherent words, sentences, paragraphs and pages that extend into code use and resultant digital manipulation. All of this is so coherent, in fact, that it leads to a distinct appearance of reasoning undergirding the whole endeavor. Crucially to Dennett, however, there’s nothing to identify as genuine comprehension in them and their processes – and the same goes for humans as well. To Dennett, humans are just vastly more complex and sophisticated in their architecture, a skyscraper to an LLM’s small office building. The lack of genuine comprehension holds across both meat and silicon. (Dennett, before his death and posthumously, ended up highly critical of AI in ways that seem at a real tension to all of this illusionism.)
This also relates to the second response to the problem, which is to lean toward a parallel processing model of how thinking works. Dennett distributes cognition, for instance, throughout the entire brain, and frames the development as one of competing influences and hierarchies cascading throughout the system. The infinite regress of agents is rejected for an interdependency of systems.
Connectionism: Thinking As Patterning
In the 1980s, criticisms to classical computationalism, particularly as the field of artificial intelligence grew in sophistication and theory, had mounted to the point that directly competing frameworks broke into the scene. The most successful was connectionism, thanks to Rumelhart and McClelland’s work in Parallel Distributed Processing (PDP).
In connectionism, symbols are no longer the basis of thought. Thinking is instead a monumentally vast network of nodes, linked together by weighted chains or connections. (This is a direct description of a neural network.) At various thresholds of activation, nodes light up and spread through the network based on the weights. When we learn, weights are adjusted and rearranged in cascading algorithmic fashion.
This allows for representations to be fully distributed, rather than suffering from the infinite causal regresses of a single chain. As a subsymbolic form of processing, it’s the differences between various patterns of activations across nodes on a map that allow us to identify a human from any other ape. It matches up very well with the human brain, has active application in digital contexts, and it’s generally quite resilient in concept; systems still function when nodes are damaged or inputs are incomplete, while parallel distributed processing helps manage weight.
Connectionism has held strong appeal ever since its introduction, but criticisms are numerous. Some dislike the continued detachment of cognition to the environment or sensorimotor processes, reasoning that surely some interaction must be involved and have meaning for thinking to work the way we imagine it. We also continue to struggle to find causal relationships that explain why conclusions within neural networks are reached, let alone why the patterns themselves are so meaningful. We revolve around back to the symbol grounding problem, except this time it’s a pattern grounding problem.
The most potent critique, however, has typically been seen as that of Fodor and Zenon Pylyshyn in Connectionism and Cognitive Architecture: A Critical Analysis (1988). Fodor and Pylyshyn highlight that connectionism struggles deeply with providing logical systematicity and compositionality to the process of thought. For example, if I understand that Jane is resentful toward John, I necessarily also understand the concept of John being resentful toward Jane, and can likely then understand the inverse or opposite, that of friendliness. How can patterns represent these direct relationships? Each different expression would be a different pattern within the space.
The routes around this critique are quite sophisticated and are expressed to varying degrees of formal precision:
- In perhaps the most direct defense, one reasons that all the feedback loops from the activation cascades interact with temporal dynamics and “ripple” into following patterns. This means “Jane hates John” and “John hates Jane” are linked through shared sub-patterns and weighted contexts.
- In the more abstract sense, we can argue that strict logical necessity doesn’t ground “proper” cognition, and neural networks simply generalize to the degree where everything is good enough for government work. After all, both humans and machines, under the connectionist perspective, cognize perfectly fine while also making mistakes.
- In the most technical sense, distributed representations use tensor product operations in mathematics (algebraic operations representing relations between objects in vector spaces) to encode “role-filler bindings,” in which “Jane dislikes John” has packed inside it whispers of “Jane and John are rivals,” “an oppositional relationship,” etc. The mathematical structure allows everything to recombine systematically.
At this point, if we take a step back and observe the patterns that connectionism would like us to observe, we start to notice that sophisticated symbolic representations verge toward connectionist arguments when resolving its issues, and sophisticated connectionist interpretations start looking like they’re sneaking in symbolic logic – with direct connections between logical structure and outcome implying causality, and traceable activation cascades within temporality.
Embodied Cognition and Enactivism: Environmental Advocacy
So far, computationalism and connectionism has attempted to create systems in which the “mind” can be talked of without any real need or interest in referencing a body, or even environment more broadly (beyond perhaps the basic situations of external referents). This is vaguely Cartesian, for eagle-eyed readers, and modern neuroscience no longer treats the mind and body as easily separable things. This is due to key objections from a developing crowd of embodiment-emphasizers within theories of cognition in the 1990s and 2000s.
Embodied cognition can be read as the tamer, older sibling of enactivism. It rejects the idea of a “brain in a vat” where thinking is done all from the shoulders-up and extends cognition outside of the brain to include body-brain interaction. Advocates note that single chains or pattern-processes that model, compute, and execute in a series don’t seem to map up with what we experience in reality. Taking a page from earlier phenomenologists like Merleau-Ponty and Heidegger, they situate the world in direct perception and possibilities for action. A table provides affordances that make possible putting things on them in ways that map to human interest and meaning. When we place things on a table, we’re moving as well as perceiving as well as doing any computing – everything is a process of dynamic coupling and coherence between all the various physical and mental relations. Sensorimotor schemas give meaning to symbols and claim to solve the grounding problem.
This makes truly abstract thought, though, seem quite opaque. What’s sensorimotor in adding 2+2, or thinking about the state of the American economy next year? If we’re not sure, how can we be sure the symbols involved have grounding? More concerningly, does this mean cognition changes qualitatively when people are disabled, embodied differently, or otherwise show peaks and valleys in cognitive complexity at different tasks?
Enactivism goes further than embodied cognition in its core claims, with Evan Thompson and Francisco Varela being the most famous within the field. For enactivists, cognition isn’t just inseparable from the body but also inseparable from the environment and the relationships in between all of them. Systems with autonomy create meaning necessarily through the work they do in interaction, and that meaning-making is itself what cognition is. In this way, enactivism strongly prejudices toward the biological as the general field where thinking resides, though advances in robotics have led to active discussion within enactivist circles. It’s also a theory that prizes autonomous and normative action within organisms as directly harmonizing with various thought experiments that were posed to computationalist perspectives, such as the infamous Chinese Room experiment formulated by John Searle (which will be covered in a future Extra Reading).
One might notice that enactivism inherits the problems of embodied cognition perhaps even more severely. On the pragmatic level, by devaluing sensorimotor schemas, it increases the difficulty of formally modelling it for research. More pressingly, however, on paper it still completely fails to account for abstract and hypothetical reasoning. If there’s no longer any interest in representations and symbols as important, how do we end up with language or memory? Answers to this reincorporate sensorimotor schemas more fully, attempt to extend cognition into tools as forms of material and cultural scaffolding (verging into 4E theory), or in the case of “radicals” Daniel Hutto and Eric Myin admit that human-level cognition is a special linguistic and cultural case on top of a virtual dominance of sensorimotor coordination.
Finally, as previously alluded to, the original insistence by Thompson and Varela that autonomy requires biological life or metabolism can and has been contested as question-begging in the age of artificial intelligence or thought experiments regarding non-biological architectures – so where do we draw our boundaries without being completely arbitrary?
Extended Cognition, 4E, and Predictive Processing: The Mental Frontiers
Once we start incorporating tools into the process, we lead to the conclusions of Andy Clark and David Chalmers in their extremely influential 1998 paper “The Extended Mind.” Under this framework, cognition does work and has presence outside of the bounds of the brain and body, most importantly extending within the tools themselves. A smartphone isn’t just something you do cognition to, it’s something that is necessarily involved and participatory in your cognition when you’re using it. The famous example is that of a woman who remembers a museum address and a man with Alzheimer’s who uses a notebook where he wrote down the address prior. To Clark and Chalmers, the notebook is just as cognitive as the biological memory. There is a parity principle: if you can even just hypothetically calculate in the head but use a calculator to do the work, the calculator is part of the cognitive process when you use it. In my opinion, this is cool stuff – it solves the abstraction problem by extending systems to materials and symbols, and grounds symbols functionally in terms of how productive they work within both body, environment, and tool usage. That said, we again reach boundary problems. Is the internet part of our cognition? Religion, or lack thereof? Is it in the physics we manipulate? Plus, it seems hard to say falsifiably that tools cause or are a necessary part of cognition; if we can extend these boundaries, where do we stop? Clark rejects this framing as itself unnecessary; he cares about results more than resolving the metaphysics.
Many of the above preoccupations – extendedness, embodiment, embeddedness, enactivity – have been slotted into an umbrella typically named as 4E cognition, despite the fact that the four E’s are often free to contradict each other depending on the interests of the person wielding them.
Clark later tied extended cognition further into an attempt at a more parsimonious framework, dovetailing with Karl Friston and the Free Energy Principle, Anil Seth, and others. With his 2015 book Surfing Uncertainty, Clark describes the brain as a sophisticated prediction engine that loops downwards and upwards. Every instant, predictions are being generated about sensory input, and deviations between what’s predicted and what’s received as input influences further behavior in cascades. This is seen in tiers of relative abstraction; when walking on the sidewalk, there’s the sidewalk scene, the items along the sidewalk, then the specific observation of a ridge that propagates upward to inform the body not to trip on that ridge today. These prediction heuristics adjust based on the clarity of input, and importantly, action integrates with error minimization – you act so as not to trip, improving your prediction of how not to trip in the future. It can be expressed formally, scales across cognitive domains, and synthesizes the best parts of a lot of prior theories. But fundamental questions within both theory of mind, epistemology and metaphysics remain: is this back to computationalism, obfuscating problems under sheer sophistication? What attaches discrete meaning to the error-minimization, and why does prediction error minimization produce subjective experience (the hard problem)? If everything can be boiled down to error prediction, how can the theory be falsified?
Another consequence of this framework is that, as an explicitly functionalist sort of theory, it’s relatively uninterested in biological primacy. Viewpoints differ, but Clark himself strives for consistency and thinks it’s an open question as to whether Large Language Models engage in “genuine” cognition or not. However, his framing is usually that of LLMs as their own powerful extenders of human cognition. The implication here might be that there’s a prioritization of a certain conception of autonomous action that disqualifies AI from having their own “standing,” as it were. But where do we define autonomous action? If it’s in the predictive error minimization itself, could we flip it around and say that LLMs also use humans as tools for their own potential cognition?
A Provisional Meta-Theory: Pragmatic Representationalism
So where does all this leave us? Well, it leads us to several conclusions:
- Formal computationalism is hopelessly mired in the symbol grounding problem;
- But other theories do better jobs never resolve the problem at the end of the day;
- As theories get more sophisticated, we hit intractable issues about where we draw our lines and boundaries for what cognition consists of;
- None of them establish a clear normative answer for what should qualify for autonomy or cognitive standing (Octopi? Corvids? Aliens? AI?);
- None of them can answer the hard problem of consciousness, naturally;
- And the more sophisticated the theory, the more unfalsifiable it becomes.
In part one of Dissolved Distinctions, we triangulated between consciousness theories to come up with this definition:
“Consciousness is a relational display of meaning-making, existing within a gradient, wherein high and high displays of cognitive and metacognitive reasoning merit greater recognition and regard.”
(This relational display is in the enactivist vein, toward a mind-body-world field in which “body” is extended outside of strict biological embodiment to instead signify “ability to act on metacognitive capacity.” Meaning-making here means: coherent conceptualizations, expressed by the subject, that serve pragmatic functions within their own environmental context.)
This provisional theory of consciousness admits its own unfalsifiability, attempts to make circularity productive, and consists as a sort of meta-theory incorporating the insights and failures of the theories below. We can do similarly with cognition by emphasizing a pragmatic relationalism that reacts to critiques in a very particular way.
Cognition is a relational display of adaptive pattern-coordination, existing within a gradient, wherein systems demonstrate integration across information streams, temporal responsiveness, and capacity for elaboration through learning.
This relational display emerges in interaction between system and environment across timescales, even without a perceptually flowing temporal state, so long as the functional equivalent of memory and learning is established. Pattern-coordination involves recognizing regularities, generating context-sensitive responses, and adjusting behavior based on feedback. The gradient is profound, not bounded, ranging from the most minimal systems conceptualizable still capable of work to the complex metacognitive architecture of the human brain, and not ending at those two conceptual poles.
Such a position consistently rejects symbolic grounding, the hard problem of consciousness, and the boundary problem as malformed questions and/or category errors. Symbols are created through distributed, parallel interactions that are pattern-matched to make meaning without requiring grounds. The hard problem presupposes a subjective/objective dichotomy, while pragmatic relationism considers subjectivity to be what the processual interactions of pattern-coordinations look like from the inside; in other words, there is no real difference or ground that separates subjectivity from objectivity per se. The boundary problem is revealed as anxiety over a loss of stability and applicability that was never fundamental or fixed in the first place – boundaries are set by purely pragmatic and relational grounds, adjusting based on the observations presented and never standing as concrete barriers. Predictive processing frameworks fit naturally here – error minimization is pattern-coordination extended to include anticipatory dynamics, applying equally to biological and non-biological systems.
In my next post, we’ll be exploring Searle’s Chinese Room, and suggest some renovations for expansion.
The Classical Computationalism part is brilliant; very fonudational for AI.