This essay is also available as a podcast on anchor.fm, Spotify, and other platforms
Hail and welcome. I mentioned in a prior episode that my theory of ecumenical phenomenology, which I’ve been exploring over the last few episodes, is rooted in two separate theories—platonism and phenomenology—along with a third which, given a small tweak, translates between them, allowing the phenomenological ideas to answer the questions posed by the platonist ideas and vice versa. Today we’ll be exploring that third theory, the connectionist theory of mind.
First, though, I’m going to tell a joke. I’ll be delivering the punchline later on in the episode, not as any sort of teaser to get people to listen longer (I wrote the joke myself and definitely don’t think it would be worth it), but rather just to make a point about something.
A master thief gathers a troupe of other thieves to pull off an art heist. They meet and the master thief explains the plan: “Things could go one of two ways, so on the day of the heist, you’re going to see me wearing one of these two hats, this red one, or this yellow-green one. Which plan we execute depends on which color hat I’m wearing. If our inside agent is working today, she’ll disable the alarm system at just the right time for us to get the painting out. But there’s a chance that her work schedule will change, in which case, we go with plan b: I shit my pants, and you all can get the painting out in the ensuing chaos. Any questions?”
One of the thieves raises a hand and says: “You didn’t tell us which plan goes with which hat! Which one goes with the yellow-green hat?”
A theory of mind is a theory about what the mind is and how the mind works. Because the focus is on the mind—a conscious system which might, depending on the theory, include the brain—and not just the brain by itself, it falls within the domains of philosophy and psychology, but there’s significant crossover with neuroscience just because we have good reason to believe that the activity of the mind is rooted in the brain and its physical structure. Now, the standpoint of seeing the “real world” as something external to us which were are experiencing in some way is one that I’ll be troubling over the course of this and future episodes, but for now, let’s just proceed from the natural attitude that this is the case: the world is an objective state of affairs that exists apart from our experience of it, and mind—conscious human thought and experience—is our standpoint of experience of that external world. So then we have to ask the question: what is mind exactly? What is its relationship to the brain? Where does mind come from and how does it work? Our answers to those questions are theories of mind. Over the last century, theories of mind have been leaning increasingly towards computational models, which describe the mind as being a computational system, like a computer. Such a model would mean that the mind is less a window through which we passively recieve our experiences of the world and more something active, moving from state to state according to certain rules or processes.
This approach largely started with the work of British mathematician Alan Turing’s 1936 paper “On Computable Numbers, With an Application to the Entscheidungsproblem.” Turing’s paper, as well as another one he published in 1950, “Computing Machinery and Intelligence,” contained a description and exploration of an idealized computer, a Turing machine, which, he theorized, could, in principle, replicate all of the activity of the human brain. According to the Stanford Encyclopedia of Philosophy, there is general agreement that Turing was correct about this (Rescorla, 2020). That doesn’t tell us that the mind is a Turing machine, only that a Turing machine could potentially be a mind.
Computers are extensions of our ability to think, and as we’ve developed computers, we’ve contemplated whether the machines themselves might be capable of thought. But this essay is not about whether computers can be or have minds, but rather about whether minds are a kind of computer, as well as what kind of computers they might be and what this might mean for us as humans and for my theory of ecumenical phenomenology. We certainly don’t know whether even a computer that can fully mimic human thought can actually be said to be thinking, but we already know that humans think and are conscious and we wonder whether the mind being a computational system might account for those phenomena. There are three possibilities: one, the mind is a not a computational system; conscious, thinking mind arises from the brain in some other way. Let’s call that noncomputationalism. Two, the mind is a computational system, but this is insufficient on its own to generate consciousness and thought; something else is present which allows humans to truly think, something which machines lack, precluding their being able to do anything more than mimic thought. Let’s call that computational insufficiency. Three, the mind is a computational system, and this sufficiently explains the phenomenon of conscious thought in humans. Let’s call that computational sufficiency.
As I’ll be getting into, inference to the best explanation leads us to conclude that either computational insufficiency or computational sufficiency are more likely to be correct than noncomputationalism, and that leads us to the question: what kind of computational system is the mind? Is it some manifestation of the Turing machine concept? That claim is known as the classical computational theory of mind (CCTM), and it has several variations that we won’t be getting into. Either computational insufficiency or computational sufficiency are sufficient for at least a weak version of ecumenical phenomenology, and it might even be possible to reconstruct the theory should noncomputationalism prove to be true. However, the strongest version of ecumenical phenomenology is predicated on a specific computational theory of mind: connectionism.
I’ll quote here the general description of connectionism from the Stanford Encyclopedia of Philosophy’s page on the computational theory of mind:
Connectionists draw inspiration from neurophysiology rather than logic and computer science. They employ computational models, neural networks, that differ significantly from Turing-style models. A neural network is a collection of interconnected nodes. Nodes fall into three categories: input nodes, output nodes, and hidden nodes (which mediate between input and output nodes). Nodes have activation values, given by real numbers. One node can bear a weighted connection to another node, also given by a real number. Activations of input nodes are determined exogenously: these are the inputs to computation. Total input activation of a hidden or output node is a weighted sum of the activations of nodes feeding into it. Activation of a hidden or output node is a function of its total input activation; the particular function varies with the network. During neural network computation, waves of activation propagate from input nodes to output nodes, as determined by weighted connections between nodes.
Rescorla, 2020
The YouTube mathematics channel 3Blue1Brown has a great series on neural networks, as implemented in computer programs. I recommend watching those for a visual explanation for how exactly this works, but I’ll summarize the explanation provided in the first video. To be clear, what I’m about to describe is a neural network algorithm as applied in computer science; this is not itself the structural description of mind provided by connectionism but rather an implementation of that description as a computer program. And furthermore, while this describes a sort of minimal architecture of a neural network, actual implementations in practice can vary quite considerably from this model.
Imagine we want to write a computer program to recognize a handwritten digit, 0 through 9. We give the program as input a greyscale image, a grid in which every pixel has a value which describes how dark or light it is. Each pixel activates one of the input nodes of the neural network to a degree corresponding to the lightness or darkness of that pixel. Each of the nodes so activated has weighted connections to the nodes in the first hidden layer. In other words, each node in the input layer has, for each node in the first hidden layer, a value (which we can normalize to real numbers between 0 and 1, inclusive) representing the strength of its connection to that hidden node. The values of those weights have been set by training the program, basically showing it digits and telling it what those digits are so that the weights reflect those patterns. The activated input nodes “pass along” the value of their individual pixels to the hidden layer at strengths corresponding to the weights of the connections. The nodes of the first hidden layer are activated based on the strengths of the signals they receive from the first layer and pass along that information to the next hidden layer. This continues throughout however many hidden layers until an output layer is reached, which has 10 nodes corresponding to the 10 digits we’re trying to identify, and if everything is working correctly and the program has been properly trained, the node corresponding to the correct digit will be activated with a very high value and the other nine will be activated with much lower values.
Each layer is, in essence, recognizing structural patterns in the layer feeding into it, moving from more finely-grained structure to broader and coarser structural descriptions of the entire image. Importantly, neural networks respond to training, dynamically adjusting the connection weights based on their own activity and their success or failure at a given task. This is the “learning” part of machine learning.
Let’s step back for a second to look at the human nervous system so that we can ground this theory in biology to at least such a degree is possible given our present level of understanding.
Nervous systems in general are complex networks of electrical circuits made from a particular kind of cell—the neuron—as well as other cells which provide supporting roles. Neurons are specialized: there are sensory neurons which receive and transmit information received from the outside world (“outside world” here referring to not only the world outside the body but also the body outside the nervous system); motor neurons cause muscles to contract; and interneurons form complex circuits within the brain.
Beyond these specializations, one neuron is largely similar to the next. They feature branched extensions called dendrites which receive information from other neurons, and a long branching tendril called an axon which transmits information to other neurons. The main differentiation between neurons is in the structure of these axons and dendrites, which wire together to form neural circuits in complex patterns. A single interneuron might be wired to receive inputs from as many as 100,000 other neurons. These connections are highly dynamic, responding and changing based on the brain’s own activity (Reese & Campbell, 2011).
Given that neurons are largely identical beyond their connections, it seems very likely that the information content of the brain is not stored in the neurons themselves but rather in the way in which the neurons are wired together. Scientists have referred to this “wiring diagram” as the connectome, relating it etymologically to the genome of DNA sequences. What we have in the connectome is a network of nodes which stores information in the connections between those nodes and which responds dynamically to its own activity. While we still have a great deal to learn about how exactly the connectome translates into cognitive experience, it gives us at the very least a viable biological substrate for connectionism and ecumenical phenomenology.
I find connectionism appealing because, one, it is derived not from some abstract model of computation but from the actual structure of the human brain; and two, because the most successful AI programs are structured around neural networks. And when I say “the most successful AI programs,” what I mean is those computer programs which are most successful at creating the appearance or effect of intelligence. Again, whether or not that can actually be called intelligence proper is not a matter I’ll be getting into; it’s simply not relevant. However, consider that the simplest explanation for why neural-network-based computer programs are so good at imitating mind would be that they’re structured like mind. Looking at it from the other direction, we can say that if the connectionist model explains mind, then we can (and do) test that theory by implementing it in computer programs and seeing if they behave like mind. Whether or not said programs can be said to think and whether or not they’re conscious, we can indeed say that they behave as though they are, at least to some limited extent. Although we’re far from knowing everything about the physiology and chemistry of the brain, connectionism at least accords with everything that we do know. We have a great many unanswered questions about how exactly mind would function as a connectionist network and how that would relate to the physiology of the human brain, but we have as yet no direct contradictions to the answers we already have. And, under ecumenical phenomenology, connectionism would explain not only the phenomenon of mind but also those of societies, civilizations, other social systems, and the intersubjective, abstract reality of human experience. So, when tested as an explanation for mind by the method of abductive reasoning, also known as inference to the best explanation, connectionism passes with flying colors. You might say that what I’ve presented here so far seems a bit cursory for me to be able to make such a claim, and that’s completely fair, but I can’t present everything at once and keep this manageable as a podcast episode, so please bear with me and I hope you’ll continue to evaluate this claim as we progress.
But before we get too enthusiastic about connectionism, it’s important to note that there are significant differences between neural network computer programs and what we’ve observed of the human brain. For example, human neurons emit discrete spikes of activity to signal other neurons, whereas the nodes in neural networks emit continuous rather than discrete data. Differences such as these may be superficial; or it may be the case that neural network computer programs implement connectionism differently than the human mind but that the mind can regardless still be said to be essentially connectionist in nature. Or, we might find that these differences ultimately show us that the human mind isn’t connectionist. Furthermore, arguments have been put forth by classical computationalists and noncomputationlists that mind exhibits certain features which cannot be explained by connectionism. It’s also important to note that connectionism doesn’t yet offer any explanation for consciousness itself: based on what we know so far, it’s possible to imagine a connectionist mind for which there is no associated conscious experience. So this is far from a slam dunk. Regardless, I maintain that the connectionist model, while far from certain, remains the best explanation for mind (and other related phenomena) that we have at present. It’s also possible that the human mind is not connectionist in nature but that it can be modeled as a connectionist network regardless. This would be to say that we can use connectionism to understand mind and predict its behavior even if it doesn’t really reflect what’s going on “under the hood.” Such modeling is common in science: for example, science textbooks commonly show diagrams of atoms which are scientifically inaccurate but which nevertheless succeed at conveying, in a general way, the structural relationship between protons, neutrons, and electrons.
I’ll proceed from here under the assumption that either connectionism is in some way correct or that it is sufficient for modeling purposes. Our next step will be to look at the theory in terms of its specific description of mind.
What, for example, plays the role of the node in the human mind? Is a node a neuron? A bundle of neurons perhaps?
A 1988 paper by cognitive scientist Paul Smolensky aims to put connectionism on its best possible footing, clearly stating its foundational claims and acknowledging the weaknesses and shortcomings of those claims as well as demonstrating their strengths. Smolensky’s approach is to analyze these questions at three different levels according to what he refers to as the subsymbolic paradigm.
To get us started, let’s posit that humans think in terms of concepts which are represented by symbols, and that these symbols exist within two parallel systems: the semantic (the world of meaning) and the syntactic (the world of grammar). This is, to be sure, a drastic oversimplification of a quite complex and mysterious phenomenon, but this explanation is just setting us on the road to understanding the subsymbolic paradigm and will no longer be needed once we get there.
Given these linguistic phenomena, we might assume that the concept is the fundamental unit of thought. Under a connectionist model, concepts would then correspond to nodes, the fundamental processing units of the network, and the meaningful relationships between them (i.e. the syntax) would be reflected in the way the nodes are wired together. This is the symbolic paradigm, which Smolensky directs us away from adopting in favor of the subsymbolic paradigm.
The name “subsymbolic paradigm” is intended to suggest cognitive descriptions built up of entities that correspond to constituents of the symbols used in the symbolic paradigm; these fine-grained constituents could be called subsymbols, and they are the activities of individual processing units in connectionist networks. Entities that are typically represented in the symbolic paradigm by symbols are typically represented in the subsymbolic paradigm by a large number of subsymbols. Along with this semantic distinction comes a syntactic distinction. Subsymbols are not operated upon by symbol manipulation: They participate in numerical—not symbolic—computation. Operations in the symbolic paradigm that consist of a single discrete operation… are often achieved in the subsymbolic paradigm as the result of a large number of much finer-grained (numerical) operations.
Smolensky, 1988, p. 3
Smolensky discusses various hypotheses about the way that subsymbols amalgamate into symbols and we won’t concern ourselves with those here. He makes it clear, though, that we don’t presently have a hypothesis that maps the connectionist model at any level directly to neural activity in the brain, so we can’t say whether neurons correspond directly to subsymbols. If we could, we would likely have enough information to confirm or refute the connectionist model itself, but in 1988, when Smolensky wrote his paper, we simply didn’t have enough information about neurons and the functional structure of the human brain to make that connection, and we still don’t today. But for our purposes, that’s acceptable (if not ideal). All we need to worry about is whether connectionism provides for representation of concepts at some level. If connectionism is correct, we have a lot of questions to answer as to how exactly neural activity generates or operates on subsymbols and how subsymbols amalgamate into symbols, but at present, we have at least hypotheses for how connectionism can answer those questions. By way of additional support, I found a 1999 paper in the Behavioral and Brain Sciences journal by Gerard O’Brien and Jonathan Opie which describes in detail how connectionism could account for phenomenal experience.
Funny story: philosophers and cognitive scientists formulated connectionism by looking at the phenomenon of mind at the individual level. I landed on connectionism myself, before I had ever heard of or read anything about the theory, by looking at the phenomena of societies and civilizations. The story of how that came about is an interesting one, I think, but after writing it out as an essay I realized (after some feedback from my patrons) that it required my introducing a number of difficult concepts which are in no way necessary to understand either connectionism nor ecumenical phenomenology. But it was fascinating to me how Smolensky’s presentation of connectionism was an ontology in need of a phenomenal account, an account of how the model is experienced and influences or causes behavior, and how I discovered connectionism by looking at phenomena and extrapolating to the ontology. This is where that tweak to connectionism comes in, the one I mentioned at the top of the episode: instead of looking at individual minds as being individual connectionist networks, we can look at the minds—or more properly, mind—of the human species as a single distributed connectionist network, and doing so provides us with a formal (if rudimentary) ontology of abstract reality, explaining a wide range of phenomena and answering a similarly wide range of philosophical questions.
The central structural feature of ecumenical phenomenology, the ecumenicon, is the information content of the connectome, the wiring configuration of the human mind, as it exists collectively, distributed across minds, forming a recursive, parallel, and highly complex dynamical system. We relate to the ecumenicon in different ways at different levels of analysis. Each individual mind is in itself the ecumenicon, but the recursive self-similarity of the ecumenicon means that we can take different standpoints according to which we can say either that the ecumenicon is within the mind, or that the mind is within the ecumenicon. Both are simultaneously true.
We relate to the ecumenicon both as a collection of individual minds and as a kind of meta-mind, though I caution against thinking of it as anything like a hive mind, like the Borg of the Star Trek franchise. It is not singular, unitary, or harmonious. Quite the contrary: the fractures and fault lines we see in our societies and civilizations are reflections of ecumenical structure. Abstract objects are patterns across this distributed dynamical system; abstract reality—the eiditic world, the world of ideas, of values, of meaning, and so forth—is our experience of the distributed dynamical system itself. Here’s the key point: this experience is of something real and, in a certain sense, objective. Really intersubjective is the better word, but even more precisely, ecumenical phenomenology obliterates the distinction between subjective and objective: our shared experience of “objective” reality exists within the ecumenicon, and our private experience contributes to its structure.
We can examine the ecumenicon at various levels of resolution in which that wiring configuration and its dynamics become more or less finely grained. The ecumenicon is instantiated at the level of individual minds, and the ecumenicon is itself the interaction between these instances, a continuity of the connectionist network of individual minds. Each instance of the ecumenicon—I’ve used the word ecumeniconidion to describe these instances in the past but that neologism is likely too clumsy to endure—is substantially similar to other instances; the differences from one to the next are precisely what individuate us, and the degree to which this similarity holds depends on how broadly one is making the comparison: the instances of individuals with a shared culture will be more similar than those of individuals from different cultures, and in fact cultures are a manifestation of those very differences.
Compare the ecumenicon to DNA. Each individual organism contains an instance of the DNA code which describes that organism’s species. That code, belonging to a single individual organism, differs very slightly from that of other organisms of the same species, slightly more from that of other organisms of closely related species, and substantially more from that of other organisms of distantly related species. Those differences themselves are constitutive of both the individual and the species: the individual is such in terms of its difference from the species, and the species is such because of the similarities between individuals.
At last we’re in a position to get a sense of what this looks like in practice. A corollary of ecumenical phenomenology is what we might call abstract realism: a position that abstract objects, including concepts like justice, are fully real, in the same sense as anything else we might call “real.” Justice is of the same substance as the rest of reality. This is not to say that it is identical to anything we might point out in concrete reality, like a car or a pen or an animal or person, but justice is regardless made of the same thing as those particular concrete objects or any others.
When I say the word “justice,” you get some meaning from that, however vague and indeterminate. You’re not thinking that I’m about to start talking about how to bake a custard tart. You might suspect that my idea of justice is radically different than yours, and that might indeed be the case, but just the fact that we’re able to say that my idea of justice is different than yours but also able to say that we’re talking about the same thing indicates that there’s some kernel of shared understanding.
Now think about all minds, all having some understanding of something called “justice” or that could be translated that way into English. For any individual, this conceptual understanding is a pattern within their connectome, one which functions in a certain way as a medium of the brain’s electrical activity, and, whatever the exact mechanism, this pattern repeats, with extensive variation, across the entire human population, the same way that a particular gene might exist within a certain range of variation across a species of animal or plant. What’s more, these connectomic instances of justice interact with each other in a dynamical system of enormous complexity distributed across billions of minds.
At this very moment, by means of this very podcast episode and these very words I am now saying, I am altering the substance of your mind, altering your reality in fact. What I’ve said becomes part of your memories, what I’m saying now part of your present conscious experience. I inscribe your reality and then refer to those inscriptions later, for example, by concluding the joke I told earlier: when the master thief has concluded the explanation of the plan, one of the troupe says, “You didn’t tell us which plan goes with which hat! Which one goes with the yellow-green hat?”
“Oh, of course,” the master thief says. “Shart-ruse.”
At the beginning of this episode, I inscribed the setup for that joke into your mind, into your reality, and was then able to make use of that inscription by finishing the joke. In doing so, I invoked a whole host of other associations, like the idea of an art heist, drawing upon your total experience of reality, something which I understand at least to some degree just by virtue of me being able to make those invocations and to communicate with you.
You might point out that my joke is based on a pun, relying on knowledge of the English language, impossible to translate into other languages, but that’s the case only for this particular joke and not for any joke or story I might tell or anything I might tell you about and then refer back to later. And in fact anyone who didn’t get the joke still understood my referencing it and concluding it. They may have understood the overall story without getting the pun.
Justice is the net result of, at the minimum, quadrillions of communicative interactions over the lifetime of the human species. It is no whole and unified thing but rather something deeply fractured and conflicted, and constantly in flux. And its substance as this distributed mental pattern is the same substance as anything else in our reality: the sound of my voice, or the objects in your environment, or your body. Such things, like justice itself, are fully constituted within your mind: that is your total experience of them, even in terms of the notion that such entities might be described as “objectively real.”
Here’s an objection, though: would it be possible for justice to shift into something that we presently understand as injustice? To answer that, keep in mind that justice is not merely whatever we refer to by that word or its cognates, but also that there is no transcendental referent to which we can refer to determine whether something really is justice or not. We only have justice itself—the distributed, self-similar ecumenical pattern—to refer to. Here’s another way to ask the question that I think will clarify the matter: if justice weren’t justice but rather something else, would it still be justice? No, it would be something else. Regardless, this model tells us something really remarkable about justice and other concepts: they are highly dynamic, stable in certain key ways but fluid around the edges. They are always in process. Thus, conceptual epistemology is necessarily process epistemology: when we talk about what we know about justice, for example, we’re talking about something that is always in process, something stable (to varying degrees, depending on the concept) but not fixed.
Now take not only justice but all concepts, all of your knowledge and beliefs and opinions, your entire reality, and think of it in those terms, as being comprised of patterns across a distributed, recursive, self-similar dynamical system of connectionist minds.
Not only does this model have enormous explanatory power over a wide range of phenomena, answering literally millennia of open philosophical, sociological, and psychological questions, I find it stunningly elegant and beautiful. Pondering or meditating on the ecumenicon, I sometimes feel a bit like Neo in The Matrix when the substance of the simulated reality of the Matrix reveals itself to him in the form of the iconic lines of green-tinted code.
There remain some aspects to this theory that I haven’t explored in this episode. Most central is the relationship between the ecumenicon and the ontic world, the “objective” or “real” world as traditionally understood. This is something I mentioned in the last episode and will be going into more detail on in a future episode. Next up, though, at least in this series, we’ll be putting ecumenical phenomenology into practice by exploring a corollary theory, the cartel model.
I hope you’ve found this piece interesting and informative. If you’ve enjoyed it, I encourage you to look at some of my other essays, and if you find my approach to philosophy and religion at all valuable, I hope that you’ll stop in at my Patreon page, which features bonus content for patrons, and that you’ll stop back by to check on my new content.
Works Cited or Referenced
O’Brien, G., & Opie, J. (1999). A connectionist theory of phenomenal experience. Behavioral and Brain Sciences, 22(1), 127–148. https://doi.org/10.1017/S0140525X9900179X
Reece, J. B., & Campbell, N. A. (Eds.). (2011). Campbell biology (9th ed). Benjamin Cummings / Pearson.
Rescorla, M. (2020). The Computational Theory of Mind. In E. N. Zalta (Ed.), The Stanford Encyclopedia of Philosophy (Fall 2020). Metaphysics Research Lab, Stanford University. https://plato.stanford.edu/archives/fall2020/entries/computational-mind/
Smolensky, P. (1988). On the proper treatment of connectionism. BEHAVIORAL AND BRAIN SCIENCES, 74.