Epistemic Status:My Best Guess. This is an opinion as to what emotion is and why it comes about. Confidence level: 60%.
TL;DR: Emotion is one representation which describes and directs, built in largely integrating regions like the ventromedial prefrontal cortex, anterior cingulate cortex and anterior insular cortex, which is why the famous patient, Elliot, lost function and feeling together. Emotion is best thought of as being used for conscious agentic steering, while many reflexive behaviors do not need it (e.g. simply eating VS foraging for food at a Wendy’s drive through). Current accounts most similar to mine (like Lisa Feldman Barrett's) stop short of the identity claim I make, preserving this same neo-cartesian error of Damasio's.
Reading Guide: The essential sections for my argument are probably 1, 2, and 5. Sections 3 and 4 mostly build the philosophical framing behind it, of which many of you (especially here) may already implicitly agree with.
What does Morality have to do with a pair of socks?
1 Meet Elliot
He was not flagrant, belligerent, incoherent, nor capricious. He scored normally on IQ tests, passed basic moral exams, and completed just about every laboratory test as anyone with a fully intact prefrontal cortex may have.[1] And yet, he could not pick out a pair of socks.[2] In his seminal work, Descartes' Error, Antonio Damasio pieces together the amalgamation of years of research regarding what he calls the somatic marker hypothesis into a coherent narrative explaining exactly how it is we make decisions. At the heart of it was his patient, 'E.V.R.' More famously, he is known as Elliot.
From Descartes' Error, 1994:
Consider the beginning of his day: He needed prompting to get started in the morning and prepare to go to work. Once at work he was unable to manage his time properly; he could not be trusted with a schedule. When the job called for interrupting an activity and turning to another, he might persist nonetheless, seemingly losing sight of his main goal. Or he might interrupt the activity he had engaged, to turn to something he found more captivating at that particular moment. Imagine a task involving reading and classifying documents of a given client. Elliot would read and fully understand the significance of the material, and he certainly knew how to sort out the documents according to the similarity or disparity of their content. The problem was that he was likely, all of a sudden, to turn from the sorting task he had initiated to reading one of those papers, carefully and intelligently, and to spend an entire day doing so. Or he might spend a whole afternoon deliberating on which principle of categorization should be applied: Should it be date, size of document, pertinence to the case, or another?
This inability to act on knowledge was sharply contrasted with his ability to have it:
His ethical judgments followed principles we all shared. He was aware of how social conventions applied to the problems. His financial decisions sounded reasonable. There was nothing especially sophisticated about the problems we set, but it was remarkable to discover, nonetheless, that Elliot did not perform abnormally.
What separated the decisions he could not easily make from those he could? On the one hand, his knowledge and reasoning were perfectly sound, and yet, for a whole class of decisions, he simply failed to come to any conclusion. Because of it, his life was falling apart.
For Damasio, there was one telling clue in particular.
At first glance, there was nothing out of the ordinary about Elliot's emotions. He was, as I said earlier, an emotionally contained sort, but many illustrious and socially exemplary people have been emotionally contained. He certainly was not overemotional; he did not laugh or cry inappropriately, and he seemed neither sad nor joyful. He was not facetious, just quietly humorous (his wit was far more engaging and socially acceptable than that of some people I know). On a more probing analysis, however, something was missing, and I had overlooked much of the prime evidence for this: Elliot was able to recount the tragedy of his life with a detachment that was out of step with the magnitude of the events. He was always controlled, always describing scenes as a dispassionate, uninvolved spectator. Nowhere was there a sense of his own suffering, even though he was the protagonist.
While Damasio and his perplexing patient sat, conducting test after test, certain gears turned in the mind of the scientist. It was something of a peculiarity that function and feeling seemed to come apart in tandem. Damasio, in my view, gets it right that there is a function to feeling. Only, where one places emotion is not inconsequential.
Damasio's answer to this question was to assign to the damaged region, the ventromedial prefrontal cortex ('vmPFC' from now on), the role of a sort of learned mapping between bodily signals (termed 'somatic markers') and what areas are to be dispositionally activated.[3] Feeling is placed in the brainstem and body as fundamental. In his framework, the vmPFC can secondarily induce feeling by top-down re-representing a primary inducer of it, such as being jolted by an electric shock. This re-representation constitutes emotion.
In this essay I argue that in a twist of irony, Damasio, in Descartes' Error, repeats his predecessor's mistake. Effectively, he holds on to the split between mind and body, only in new clothes. His account relocates affect to the body, leaving the vmPFC as a descriptive organ of cognition.
What is more, the pioneers of the AI revolution are drawing much inspiration from neuroscience. In an interview with Dwarkesh Patel, Ilya Sutskever speaks of the possible value of emotion in making us a 'viable agent,' and its potential analogy for ML.
Ilya: ‘... it may or may not be possible to get that from pre-training.’
Dwarkesh: ‘But, what is that? Like, how do you think about emotions, what is the ML analogy for emotions?’
Ilya: ‘It should be some kind of value function thing, but I don’t think there is a great ML analogy because right now value functions do not play a very prominent role in the things people do.’
There is much to draw from neuroscience in regards to AI. How many technological revolutions have a biological antecedent which dwarfs them in complexity and efficacy? Only, getting it right is everything.
In my estimation, the reason feeling and function are sometimes lost together may be because there is a function to the feeling. Namely, I hold that certain key areas integrate descriptive perception and directive, subjective drives in order to create a perception with pull, or a pull with description: One representation which describes and directs. This, I take to be the role of our emotion.
By descriptive, I will mean a representation meant to match the world (as seen in the detective's list in the example below).
By directive, I will mean a representation the world is meant to match (as seen in the wife's list in the example below).
2 Damasio’s Error
Direction of Fit
I get the feeling Damasio saw indecision, and so he went looking for a way to narrow down the space of decisions. In doing so, he found a lookup table, presupposing affect. The vmPFC prunes the search space using subjectivity.
My question is, what does morality have to do with a pair of socks?
I say there is zero logically approachable structure to the answering of such questions, and so emotion is required. Personal, societal, and moral questions are a different type — they are just that, personal and individual — and the difference must be accounted for. What moves Elliot is almost only ever fatigue, alarm, a poke, a prod. He had emotion and feeling, only, not in the realm of self-referential, social, and human problems.
Professor Elizabeth Anscombe, regarding her study of intention, once thought of a man with a shopping list.[4] This list, as it came from the man's wife, was to be used as an order. If the list said 'apple,' and the man brought back a pear, the man was at fault. However, there was another list — a detective's list. His list was not an order but a description, as it served to catalog the man's purchases as the detective followed him. If this list said 'apple,' but the man clearly bought a pear, it was the detective who was at fault. This example is an illumination of the concept of direction-of-fit. The validity of the man and a list, and which can be said to have erred, was not isolated to the contents of the list nor the actions of the man, but the roles in relation to one another. This concept has since been extrapolated for use in the philosophy of mind. John Searle coined the terms 'world-to-mind' and 'mind-to-world'.[5] The former would be not unlike the wife's grocery list, where the world is to conform to it: the man is to go get what is written on it. The latter is more akin to the detective's list, in that it describes: it should have on it what the man actually bought from the store. Either the world is meant to conform to the mind, or the mind to the world. The drives endeavor to have the world conform to their pull, whereas perception conforms to the world. When perception is wrong, it can be said to be representing things incorrectly.
When, then, are the drives wrong? Colin Klein, in his book What the Body Commands, makes the case that pains are imperatives. On his account, pain, just as hunger or thirst, tells you what to do. Klein defends what he calls 'pure imperativism,' on which pains only command.[6] The drives are not themselves right or wrong in a descriptive sense; they simply command. They may be said to command incorrectly, as in phantom limb syndrome, but such a claim invites the question, 'says who?'
The reason this distinction is relevant here is simple. 2 + 2 = 4 is provable, and exact. Whether that grass is green, sure, is in a way subjective, but it is directly checkable against the world and the matching accounts of others, and at base comes from concrete physical properties. Who was the 42nd president of the United States? None of these require any sort of difficult personal or interpersonal consideration or contemplation. They exist as — whether nominal, abstract, or concrete — beliefs perceived about the world or logic via the mind. Beliefs are mind-to-world. Desire, however, places a subjective stake on reality. There is no ground truth to these sorts of representations. What job to hold down and what socks to wear are alike in that their validity or invalidity is not a thing out there in the world to be described or proven, but comes from a thing inside the biological agent to be satisfied, accounted for, or not.
The trolley problem is the famous question of whether or not you would pull a switch to divert an incoming trolley from killing five people to instead running over just one person. Most people answer 'yes,' and elect to save the five in favor of the one unfortunate bystander, a result illuminating a general utilitarian bias in people. Now, think of the trolley problem's more visceral counterpart, the Footbridge problem. In this scenario, participants are told to imagine they are on a footbridge above the tracks. This time, there is no lever, but on the footbridge alongside you is a large stranger. The only way to save the five people about to be run over on the tracks is to push this stranger off the footbridge and down onto the tracks. About 12% of people say yes.[7] The numbers remained the same, but the feeling changed.
Elliot was generally indecisive for a particular class of problems, but that's not where the story ends. In the article eerily titled "Harming Kin to Save Strangers,"[8] the authors continued the long lineage of research detailing the exacerbated utilitarian choices of vmPFC lesion patients. This time, they used a somewhat more sinister 'set of moral dilemmas,' one of which being a question of whether or not to harm one's own family member over a greater number of strangers. Not surprisingly, in all of the dilemmas, vmPFC patients showed a greater propensity to choose numbers over nostalgia.
I am not aware of any time-constrained or measured studies of this sort, but I think it is relatively fair to say that, if anything, vmPFC patients may make this sort of decision a bit too easily. Is his affliction really about indecision, about large option spaces?
This deficit presents in different ways, and we can think of it, in part, as such: The ability to make a decision may depend upon whether a ‘computable proxy’ is around. For the client's documents, Elliot could sort them competently by similarity, disparity, and content, but for the life of him, he could not settle on which principle of organization he liked best. The latter aspect requires choosing what matters for the client, or how he would like to organize them himself. Perhaps he could compare different methods of organization along some dimensional axes, but this sounds like more of a solution than it probably truly is. Suppose you and I are two quite hyperrational vmPFC lesion patients. I come to you with a philosophical treatise titled 'The optimal method for organizing client files in X line of work.' It is 1800 pages, full of citations, math, and proofs. But still, at the end of it all, you say, 'well, why did you value those things, such as time and money, over, perhaps, the color of the binder notes?' What else, you could come up with your own 1800-page treatise, only ever so slightly different, and say, 'no, this is the optimal method, for you see I have named it as such.'
At the end of one session, after he had produced an abundant quantity of options for action, all of which were valid and implementable, Elliot smiled, apparently satisfied with his rich imagination, but added: "And after all this, I still wouldn't know what to do!"
In contrast to this, the trolley problem, the footbridge problem, and the even darker twists, all give vmPFC patients a number, and so they compute. Socks would hand Elliot nothing tangible, because the necessary proxy is not some number or fact, and it's not in the basic context of the body either. What is necessary for such questions is a subjective valuation of the socks' worldly characteristics. Perhaps Elliot values conformity and would like a traditional white pair to fit in, or maybe he prefers the color neon green because of some positive association. Any sufficient answer invokes 'he,' the self, the subjective account.
There is much in the way of evolutionary support for this broader framing of orbitofrontal valuation circuitry having to do with subjective questions. Rodent and primate orbitofrontal cortices,[9] in part, do something like subjective outcome valuation over time: a question of 'what does this mean to me now?' How much do I value this opportunity at stealing a bit of food from this lady's kitchen? More than the risk of being caught? What if I were really hungry? What if I wanted desperately to fulfill my lifelong dream of being a world-renowned chef despite being a rodent? Of course, such a question is a far cry from morality, but the common substrate is there. Perhaps this same, more primitive area of self-valuation operations was elaborated upon in evolution to not only include self-valuation, but inter-self valuation.
3 The Computer Metaphor
Paul Cisek, in "Beyond the Computer Metaphor: Behaviour as Interaction,"[10] writes, 'One can say that the general working assumption in brain science today is that the function of the brain is to convert stimuli into reactions.' Cisek goes on to illuminate the core issues underpinning such a working assumption, and how the input-output functional viewpoint of the brain leads to unnecessary difficulty in the approach of certain problems, namely those of 'consciousness, emotion, motivation, and meaning.' For Cisek, things are backwards: meaning precedes symbols.
More than anything, the drives begin to look different when the working metaphor of the brain switches from a computer to a control system. Cisek, again, writes, 'For computers there is no notion of desirable input within the computing system, and hence there is the riddle of meaning, a.k.a. the symbol grounding problem.' The symbol-grounding problem is defined as follows: "How can the meanings of the meaningless symbol tokens, manipulated solely on the basis of their (arbitrary) shapes, be grounded in anything but other meaningless symbols?"[11]
In his view, the evolutionary reality that meaning precedes symbols effectively neutralizes the symbol grounding problem. While I agree with the gist, I think an explanation is still needed.
Often, perhaps due to the computer metaphor, people tend to treat the drives as objects. The common currency view, for instance, sees certain OFC sites like the vmPFC as containing a neural common scale of value.[12]
In contrast, Damasio merely presupposes feeling. His account of the vmPFC resembles a lookup table, mapping intelligent and affective bodily context in order to dispositionally re-represent it. That is, to the very complicated and shrouded mystery of emotion, he places it fundamentally in the body and brainstem. Solms takes a similar view.[13] There are two current camps available for how to view the drives: One being through a computational lens, often treating them as objects like labeled scalars, and the other, taking them as given, somehow coming from the body and brainstem.
In the control-oriented picture, all of this begins to make little sense. Drives are not computational objects, but links in a control system. Their interpretation, problematically, is made either empirically, by a third-person observer, or phenomenologically, by a first-person experiencer.
The imperative force of what we may call a directive signal is in its position in an architecture, not in the signal itself. Think of a perspiring organism: One sensory neuron which senses heat and connects to a neuron which activates a sweating process. The signal we observers read, roughly, as 'sweat, now,' in the perspiring organism is so defined by where it goes and what it induces. Now think of a scientific observer looking at the brain of a hungry mammal. Magically decompose the ensemble(s) which appears to function as 'eat' or 'be hungry' and you recover a list: ghrelin at X, glucose at Y, stomach distension at Z. The imperative is nowhere in the list, and it is nowhere in the ensemble either, but in what the ensemble is wired to (observably), and in the history which made that wiring successful (teleologically).[14]
The implication of force being a property of arrangement for something like hunger would be that, in a way, there are many 'levels' to hunger. At one observable level of some ensemble which seems to mean hunger, there may be a reflexive wiring to some automatic process like salivation, or ghrelin release from the stomach. At another, perhaps 'higher' level, hunger may be empirically observed as (in part) some activity in the anterior cingulate and anterior insular cortices.[15] Rather than hunger as a whole being a thing many parts of the brain have computed, it is a general direction of control activity based on the bodily senses, cognition, and so on. Sometimes, this may route to very simple behavior, or it may instead need to focus agentic attention, intention, and activity upon the sometimes complicated and holistic process of scavenging for food. Similarly, the same activity in the brainstem could be 'consumed' early and reflexively by some innate purging reflex, or perhaps could be merely the innate origin (observer-relative and interpreted) of some conscious feeling of disgust which is dispositionally activated (and/or partially constituted) by the ventromedial prefrontal cortex after hearing some esoteric words, such as, 'she vomited.'
In sum, the control-oriented picture would perhaps be better suited to a verb like 'flow' over 'compute.' In this way, what the observer takes to be drives are identified by their consumers. The signals originate in the body, brainstem, but the drive is not properly located there for the observation of all functions. For example, conscious emotion often heavily implicates the cortex: Why is it that both the cortex and limbic areas are active during emotions such as disgust or feelings such as hunger?
The cortex's involvement in drives is seemingly paradoxical. It is a wonder that hunger can be represented to us experientially as a descriptive feeling with both intensity and a certain quality, and that this same representation, seemingly, is what moves us. For reflexes, this is no problem. Merely reflexive creatures can jump between impulse and action in a cool, continuous stride. But for us — how is it possible we can decide at all?
4 The Ant and the Aunt
The ant, if you cut off 50% of its leg length, walks about half the way there.[16] If you put your Aunt Sue on stilts, she still makes it to the destination, providing she can balance well enough. This is not to argue ants fit, in all aspects, this sort of criterion, only at least in this one ability of individually mediated walking. But Aunt Sue herself must account for more than where in geometric space a half-eaten bag of Doritos was left on the ground and how to get there. She needs to take into account not merely all of her drives and base desires, but esoteric and abstract concepts such as 'it matters what people think of me' and 'it is unruly in the eyes of society to get on all fours and eat off the ground', very real concepts of cognition which must at every moment battle with some drives whose sentiments are something akin to 'those Doritos look mighty felicitous right now' and yet work in support of others such as 'I care about my place in the group and would be fearful of any meaningful embarrassment.' There is no reason the ant's ambulation be modulated by such a consideration, while it is crucial for the aunt's. The ant's problem is solved by simple odometry, a biologically cheap and easy implementation. Aunt Sue's requires the drive to become a term, and yet to continue to remain a driver of her choice, and for these drives to be contrasted against one another in light of abstract, esoteric, descriptive considerations.
This ability may be called Situational Arbitration: The exchange rates of reward are set by the situation as understood from both ends of the system, interoceptively and exteroceptively, and their communication.
Less clearly than the Aunt, but entirely as necessary (and evolutionarily antecedently), the problem arises when a perception in the world does not have a preordained answer for it. An ant's ambulation can respond to a multitude of physical perceptions in a complex manner, but a cat must also be able to know what to do with the opaque grey box labeled 'Fancy Feast,' taking into account in deliberation all of her subjective desire and all of her perceptive belief. There are a host of physical beliefs about the world unavailable to be accounted for in light of the ends of agents by a preordained manner of input and output, however complex and adaptable.
The comprehension of the words 'Wendy's 2-for-1' can influence the decision of whether or not to stop and eat. Blood sugar levels may also influence the choice. Physical worldly perceived factors with different directions of fit meet in the middle to convergently decide what to do. One's force is in the perception of the stimulus itself as learned in terms of relevance and subjective value, while the other is seen in the arrangement of a long and complicated control system. If you were pretty hungry but unsure whether or not to stop, and only understood Spanish, the Wendy's ad may not provide an adequate nudge. How is this accomplished, and can we overlook the direction of fit distinction, merely treating this function as a prediction with priors, or a value function, or what-have-you?
A control system must see the mind-to-world representations of the objective in light of its world-to-mind posited drives and desires. In this way, there is a genetically posited (and developmentally influenced) subjective interpretation of reality, not only in the way it appears but 'what it means to me.' To the cat, I am some vague symbol, at once made up of the feelings of warmth, satiety, comfort, maybe love — as well as the perceptions of treats, pets, and brunette hair. A perception may need to say how things are, and yet it must move you. Cats do quite well to thrive in quite modern and artificial environments consisting of inert plastic prey who have on the side of their skin the symbols 'Fancy Feast' or 'Delectables!'
It is easy to overlook the subjectivity of choice, given it arises from physical, almost mechanical systems. Seemingly, the crowded meritocracy of subjectivity is modulated by the affect. In reality, it also creates it.
5 Emotion
It is feasible to split description and direction, and it seems to work. There can be a common-currency-like scalar taken in by the function which actually directs. Or, there may be a magical fountain of affect situated in the brain and body, leaving areas like the vmPFC to subsist as descriptive lookup functions. This is not a category of position which is unpopular, wholly irrational, or one I can argue against in a manner which can disprove it. Rather, I simply think there is an alternative which, if abductively reasoned into, yields greater explanatory power. Most accounts continue to split mind and body in what I take to be merely a more scientifically acceptable version, and the computationalist lens does well to allow for this to continue.
Cisek's work on looking through a more evolutionary and control-oriented lens is very useful, as well as his affordance competition hypothesis,[17] which I view as perfectly commensurable with my own. Overall, I think he does well to reframe our picture of cognition. Still, I view it largely as a diagnosis (and a valuable start to an answer) in need of more answers, and I hope to contribute something here.
A representation with both directions of fit is widely believed to be incoherent, or at least unnecessary.[18] Others, well before me, argue for it.[14:1] I think Millikan’s pushmi-pullyu representations (‘PPRs’; representations which are at once directive and descriptive) are seen primitively because of their principality in biological agentism, not because they are merely primitive. Millikan holds that they split off later in evolution, into descriptive and directive faculties.
What if we simply say that Elliot lost function and feeling at once because their integration occurred at the vmPFC itself? My claim is that at certain points of the brain, one function feeds both directive and descriptive consumers. In my view, the vmPFC is,
A connector hub-like region, integrating many other regions[19]
Of those regions it connects to, many are concerned more or less with belief, and many more or less with desire (to generalize simply)
This takes place for a particular class of decisions pertaining to self or inter-self questions
What if the representational format it constitutes from the integration of the somewhat differential (and perhaps incommensurable) representations of the limbic and cortical regions, the former of which has no obvious form or meaning on its own, is a valenced perception? A perception with pull, or a pull with perception. An apple can make me more hungry, and I can speak of my hunger or weigh my hunger against other subjective ends. This function of integration occurs, perhaps, so that the idea of the trolley and footbridge problems do not present as inert questions but offer a visceral reaction at once? A visceral reaction which takes all that is relevant into account of all else. So that the visual representation of socks is not inertly presented but subjectively valenced as seen? What if this is something of a necessary functional invention, a stamping of the formless, posited and subjective onto the worldly in order to decide? Nothing about socks themselves constitutes truth, the choice necessitates a subjective position. As I see it, there is no computed object to be tossed around, but a multitude of acutely meaningless signals of a biological control system which must work in tandem to navigate reality with some posited sense of direction.
This may explain Elliot: He could perceive, and he could feel, but no longer for this class of decisions — for this category of consideration, the integration of the two was lost. Certain perceptions lost their pull.
Extrapolating from this, the brain regions most implicated in the actual feeling of emotion should be those which most integrate limbic and cortical areas. The anterior temporal lobe, anterior cingulate cortex, anterior insular cortex, and ventromedial prefrontal cortex are some of the areas I have identified thus far. Their roles in functions like choice, attention, valuation, and decision, are of note.
For Lisa Feldman Barrett, the brain constructs emotion by applying learned 'emotion concepts' to internal and external sensory input and the resulting bodily states. She relies on the predictive processing lens, viewing this process as a sort of active inference.[20] Arguably, she is among the nearest accounts to my own, pinpointing many of the same (but not identical) regions of interest and treating emotion as more or less a functional construction. She writes, in "Interoceptive Predictions in the Brain":[21]
In this way, the anterior insula and the cingulate cortices help to create a multisensory representation of the body in the world so that what you see, hear and feel is influenced by your interoceptive predictions. The result is a multisensory representation of the world from the perspective of someone with a body. The portions of the vmPFC that make up part of the visceromotor cortex belong to the DMN (which is also known as the 'mentalizing network') and could therefore transmit anticipated homeostatic consequences to the rest of that network as predictions during phenomena that robustly engage this network, such as semantic and conceptual processing, imagery, mind wandering, memory, empathy, person perception, mental-state inference, self-related processing, valuation and language processing, among others.
I would caution against some of the stronger wording, like 'The result is a multisensory representation of the world from the perspective of someone with a body.' Nevertheless, and almost hand-in-hand, I see these areas as pertinent and doing much the same things functionally. Only, with the predictive processing lens, she continues to stop short of applying the identity claim I do. Klein argues that Predictive Processing can't get desire without smuggling it in, a sentiment I agree with.[22]
Empirically, I think there are some potential clues pointing my way. Elliot lost function and feeling in tandem (as do other vmPFC patients). One akinetic mutism patient, after having recovered, said about her experience: "I really had nothing to say."[23] Insular-damaged rats cannot properly recall the value of the food, affecting their decisions.[24] All of these cases can be absorbed by other frameworks perfectly well. I am only offering another potential explanation, born out of a different lens, and a dissatisfaction with the current accounts, which, in the sentiment of Cisek, seem to struggle most with certain problems in the realms of 'consciousness, emotion, motivation, and meaning.'
As I see it, the site damaged in Elliot was not a lookup table using brainstem or bodily feelings. It was the site where largely world-referential processing and subjective steering were both taken into account, and the format it constitutes is one of both direction and description, at once world-to-mind and mind-to-world. To us, it is that red apple which moves us, that feeling of hunger that at once directs attention and can be descriptively used in considerations — even talked about. Such is what I take emotion to be, its role, and generally where it would be most directly constituted in the brain.
6 Empirical Commitments of this View
For these, I will focus on the vmPFC, anterior cingulate cortex (ACC), and the anterior insular cortex (AIC). They are well-studied (and hotly contended on the particulars of, I think largely owing to the ideas circled in this post) and my view should hold each of them to certain predictions which may support or falsify this account. This is my best effort currently, and I would love feedback.
Also, it may be important to note that I am not arguing these areas contain emotion, are its only constitutors, or the only places in the brain we can find affect… The anterior temporal lobe, for instance, is also of note based on this theory, and I am working out some intuitions which pertain to that now. What I am saying, is that the nature of their being highly integrative of both world-referring and subjectively posited content, for use in valuation, attention, decision, choice, and so on, makes them highly implicated in agentic decision as well as emotion, and so they are important areas of note for this view of emotion, not the only ones. The very ‘force as a property of arrangement’ and control vs computation points from earlier should further shed light on the extent to which this stuff is not easily narrowed down, and why I am quite anti-localizationist.
Prediction #1: Reweighting from Belief Alone
This harkens back to 'Situational arbitration.’ Belief update, even on its own, should be able to re-weight different desires and impact choice. Pretty much every account should predict this, but I’ve included it anyway.
Do I think belief and desire can come apart, empirically? No. What I mean is that merely seeing that Wendy's sign should constitute emotion, unless particular sites of this integration were damaged (even though you could still see the sign).
Prediction #2: Feeling and Function Do Not Dissociate, for the Same Class of Decisions
The same domain of decision/valuation/behavior should be affected both in terms of the typical function and the feeling. Elliot understands but it doesn't move him. Smokers with insular damage have less of a conscious urge to smoke as well as a lessened habit of smoking.[25]
The way I would explain that is the AIC does not direct their attention to the habit as much. The feeling of the urge and the habit should lower in tandem. Another: Anterior midcingulate cortex stimulation leads to increased heart rate AND a will to persevere.[26] I would explain this by saying this area both initiates involuntary reflexes and prepares the agent consciously. Certain similar regions like the posterior cingulate should, if stimulated, lead to some effect but no/less conscious affective change — function and feeling are more separate.
To be clear, the cingulotomy and pain asymbolia literature[27] may seem like a discriminating case against OR for me depending on one's interpretation. This may stem from potential confusion with the directive and descriptive dichotomy: It may appear that the ACC is broadly directive and the AIC descriptive. This is not the sense in which I mean the two terms. I view both of them as being both directive and descriptive, only in different ways. Look at it like this:
What do its descriptive consumers do with it?
What do its directive consumers do?
For instance, the Cingulotomy patient that feels pain but does not care for it has the AIC intact, so it seems 'oh description but not direction.' This is a rather phenomenological interpretation. My claim is more specific. Each site constitutes emotion in a different way, and one presents more imperatively and the other descriptively, phenomenologically speaking. However, both implicate belief and desire at once. The AIC being intact still allows for (or partially aids in) both:
The direction of attention based on pain (e.g. it is the ankle which hurts, maybe this is presented to your current thought so you can choose to not walk on it)
The qualitative feeling of pain
A drive is described and it still ‘directs’ its function. Now, I am aware of the studies that generally show things like (but not exactly this) such a patient (no imperative aspect of pain) may still walk on their broken ankle, or may chew on their tongue. Of course, with no phenomenally imperative aspect of pain, patients would care less. I would say it would be far worse with zero phenomenally evaluative aspect of pain too.
Prediction #3: Reflexive vs. Deliberate Behavior of the Same Underlying 'Drive' Displays dissociative potential for observed activation in AIC and ACC vs brain stem and hypothalamus (other regions may be carefully added in this contrast too)
That is, from outside, hunger can look like ghrelin increasing OR like your roommate being rude to you because his blood sugar is low, to the observer. The former sort should be able to (not that it always will) occur without activation (or with less of) of AIC and ACC as well as without conscious knowledge of any hunger. The latter should not.
Prediction #4: Choices with an Answer Should Not Depend on These Sites
Elliot can probably answer a multiple-choice moral exam with numbers as proxies. He probably cannot pick out a pair of socks very well.
7 Objections & Replies
Does Damasio not predict all of this too?
I would say, empirically, our accounts end up being quite similar. The difference is more in the why/how than the 'what happens.' I think both matter a lot. Accounts which do not place affect in the body à la The Hidden Spring but still maintain a descriptive and directive split at these cortical regions may offer more in the way of discriminating comparative tests, and that is a good thing because I view them as the stronger accounts compared to his own.
Are you comfortable with the fact your first post on LessWrong is about consciousness?
Not particularly, but this is how the cookie crumbled.
In all seriousness, this is a theory of emotion. That said, delving into this has changed my mind about consciousness and there is a route to taking this theory and simply extrapolating it to feeling in general: The reason it feels like anything at all is to perceive and think about the concrete world in light of arbitrary goals, and phenomenality is that representational format. A representation with both directions of fit, the invention of the subject. I think it is more important to not put the cart before the horse, and begin with this idea. Nevertheless, I also think we do need more functionalist/materialist theories of consciousness which can be applied to AI as well, and my account may lend well to this.
Would a Value Function Count?
This deserves its own post, but in short, not exactly, and I think Ilya (in the aforementioned interview) is thinking about it slightly off. My intuition which I am working on developing is that a value function is reflexive, eo ipso.
Is this falsifiable?
I intuit the best route to falsifiability may be prediction #2 above. Still, I don’t think any of the current accounts on emotion are particularly falsifiable and this sort of thing will be simply settled upon over time. See Mark Solms and Lisa Feldman Barrett talk with one another, here[28] and here[29], and you may get a sense of what I mean. What is the explanation that strains least in order to answer questions, or fit evidence?
Related Accounts
My ideas depend on, overlap with, etc., to some degree or another many accounts which have already been cited, like Barrett, Solms, Millikan, Cisek, etc.
Dickinson and Balleine are overall likely the closest account if you factor in the tentative consciousness point.[30] However, I would treat the feeling as one representation with both directions of fit while they have an interface. The realm of argumentation and philosophical treatment is also quite disparate, and I offer some (different) neuroanatomical claims.
In terms of overall proximity to another account which has not yet been mentioned, Bud Craig is quite close, and I recommend his book How Do You Feel?[31]
Joseph LeDoux’s work is somewhat similar, and I view his terminology separating ‘survival circuits’ from emotion to be quite useful and basically how I think about it.[32]
The framework concerning many of the same topics, well-known around here, is of course Steven Byrnes'.[33] I am a fan of his work, but I locate the primary difference between our views (insofar as I understand his) in that the verdict on certain choices, for me, occurs in his Learning Subsystem. In other words, for me, valence can be supplied by the cortex, not just predicted by it.
Antonio Damasio, Descartes’ Error (1994). Primary source of Elliot’s case and all of Damasio’s quotes about it. ↩︎
The socks example is likely a retelling, here from Ilya Sutskever on the Dwarkesh Podcast (emotion discussion starts at ~9:39). I keep it instead of client files or any other factual decision Elliot had, because it lends itself to my rhetorical ends. ↩︎
Tataranni et al., “Neuroanatomical Correlates of Hunger and Satiation” (1999). ↩︎
Wittlinger, Wehner & Wolf, “The Ant Odometer: Stepping on Stilts and Stumps” (2006). To be clear, my sentence is not to be taken as a hard stance on ant ability, nor do I think the comparison is entirely fair to them. ↩︎
Paul Cisek, “Cortical Mechanisms of Action Selection: The Affordance Competition Hypothesis” (2007). ↩︎
Michael Smith, “The Humean Theory of Motivation” (1987). ↩︎
van den Heuvel & Sporns, “Network Hubs in the Human Brain” (2013). For the term “connector hub” only. ↩︎
Lisa Feldman Barrett, “The Theory of Constructed Emotion” (2017). ↩︎
Barrett & Simmons, “Interoceptive Predictions in the Brain” (2015). ↩︎
Colin Klein, “What Do Predictive Coders Want?” (2018). ↩︎
Epistemic Status: My Best Guess. This is an opinion as to what emotion is and why it comes about. Confidence level: 60%.
TL;DR: Emotion is one representation which describes and directs, built in largely integrating regions like the ventromedial prefrontal cortex, anterior cingulate cortex and anterior insular cortex, which is why the famous patient, Elliot, lost function and feeling together. Emotion is best thought of as being used for conscious agentic steering, while many reflexive behaviors do not need it (e.g. simply eating VS foraging for food at a Wendy’s drive through). Current accounts most similar to mine (like Lisa Feldman Barrett's) stop short of the identity claim I make, preserving this same neo-cartesian error of Damasio's.
Reading Guide: The essential sections for my argument are probably 1, 2, and 5. Sections 3 and 4 mostly build the philosophical framing behind it, of which many of you (especially here) may already implicitly agree with.
What does Morality have to do with a pair of socks?
1 Meet Elliot
He was not flagrant, belligerent, incoherent, nor capricious. He scored normally on IQ tests, passed basic moral exams, and completed just about every laboratory test as anyone with a fully intact prefrontal cortex may have. [1] And yet, he could not pick out a pair of socks. [2] In his seminal work, Descartes' Error, Antonio Damasio pieces together the amalgamation of years of research regarding what he calls the somatic marker hypothesis into a coherent narrative explaining exactly how it is we make decisions. At the heart of it was his patient, 'E.V.R.' More famously, he is known as Elliot.
From Descartes' Error, 1994:
This inability to act on knowledge was sharply contrasted with his ability to have it:
What separated the decisions he could not easily make from those he could? On the one hand, his knowledge and reasoning were perfectly sound, and yet, for a whole class of decisions, he simply failed to come to any conclusion. Because of it, his life was falling apart.
For Damasio, there was one telling clue in particular.
While Damasio and his perplexing patient sat, conducting test after test, certain gears turned in the mind of the scientist. It was something of a peculiarity that function and feeling seemed to come apart in tandem. Damasio, in my view, gets it right that there is a function to feeling. Only, where one places emotion is not inconsequential.
Damasio's answer to this question was to assign to the damaged region, the ventromedial prefrontal cortex ('vmPFC' from now on), the role of a sort of learned mapping between bodily signals (termed 'somatic markers') and what areas are to be dispositionally activated. [3] Feeling is placed in the brainstem and body as fundamental. In his framework, the vmPFC can secondarily induce feeling by top-down re-representing a primary inducer of it, such as being jolted by an electric shock. This re-representation constitutes emotion.
In this essay I argue that in a twist of irony, Damasio, in Descartes' Error, repeats his predecessor's mistake. Effectively, he holds on to the split between mind and body, only in new clothes. His account relocates affect to the body, leaving the vmPFC as a descriptive organ of cognition.
What is more, the pioneers of the AI revolution are drawing much inspiration from neuroscience. In an interview with Dwarkesh Patel, Ilya Sutskever speaks of the possible value of emotion in making us a 'viable agent,' and its potential analogy for ML.
There is much to draw from neuroscience in regards to AI. How many technological revolutions have a biological antecedent which dwarfs them in complexity and efficacy? Only, getting it right is everything.
In my estimation, the reason feeling and function are sometimes lost together may be because there is a function to the feeling. Namely, I hold that certain key areas integrate descriptive perception and directive, subjective drives in order to create a perception with pull, or a pull with description: One representation which describes and directs. This, I take to be the role of our emotion.
By descriptive, I will mean a representation meant to match the world (as seen in the detective's list in the example below).
By directive, I will mean a representation the world is meant to match (as seen in the wife's list in the example below).
2 Damasio’s Error
Direction of Fit
I get the feeling Damasio saw indecision, and so he went looking for a way to narrow down the space of decisions. In doing so, he found a lookup table, presupposing affect. The vmPFC prunes the search space using subjectivity.
My question is, what does morality have to do with a pair of socks?
I say there is zero logically approachable structure to the answering of such questions, and so emotion is required. Personal, societal, and moral questions are a different type — they are just that, personal and individual — and the difference must be accounted for. What moves Elliot is almost only ever fatigue, alarm, a poke, a prod. He had emotion and feeling, only, not in the realm of self-referential, social, and human problems.
Professor Elizabeth Anscombe, regarding her study of intention, once thought of a man with a shopping list. [4] This list, as it came from the man's wife, was to be used as an order. If the list said 'apple,' and the man brought back a pear, the man was at fault. However, there was another list — a detective's list. His list was not an order but a description, as it served to catalog the man's purchases as the detective followed him. If this list said 'apple,' but the man clearly bought a pear, it was the detective who was at fault. This example is an illumination of the concept of direction-of-fit. The validity of the man and a list, and which can be said to have erred, was not isolated to the contents of the list nor the actions of the man, but the roles in relation to one another. This concept has since been extrapolated for use in the philosophy of mind. John Searle coined the terms 'world-to-mind' and 'mind-to-world'. [5] The former would be not unlike the wife's grocery list, where the world is to conform to it: the man is to go get what is written on it. The latter is more akin to the detective's list, in that it describes: it should have on it what the man actually bought from the store. Either the world is meant to conform to the mind, or the mind to the world. The drives endeavor to have the world conform to their pull, whereas perception conforms to the world. When perception is wrong, it can be said to be representing things incorrectly.
When, then, are the drives wrong? Colin Klein, in his book What the Body Commands, makes the case that pains are imperatives. On his account, pain, just as hunger or thirst, tells you what to do. Klein defends what he calls 'pure imperativism,' on which pains only command. [6] The drives are not themselves right or wrong in a descriptive sense; they simply command. They may be said to command incorrectly, as in phantom limb syndrome, but such a claim invites the question, 'says who?'
The reason this distinction is relevant here is simple. 2 + 2 = 4 is provable, and exact. Whether that grass is green, sure, is in a way subjective, but it is directly checkable against the world and the matching accounts of others, and at base comes from concrete physical properties. Who was the 42nd president of the United States? None of these require any sort of difficult personal or interpersonal consideration or contemplation. They exist as — whether nominal, abstract, or concrete — beliefs perceived about the world or logic via the mind. Beliefs are mind-to-world. Desire, however, places a subjective stake on reality. There is no ground truth to these sorts of representations. What job to hold down and what socks to wear are alike in that their validity or invalidity is not a thing out there in the world to be described or proven, but comes from a thing inside the biological agent to be satisfied, accounted for, or not.
The trolley problem is the famous question of whether or not you would pull a switch to divert an incoming trolley from killing five people to instead running over just one person. Most people answer 'yes,' and elect to save the five in favor of the one unfortunate bystander, a result illuminating a general utilitarian bias in people. Now, think of the trolley problem's more visceral counterpart, the Footbridge problem. In this scenario, participants are told to imagine they are on a footbridge above the tracks. This time, there is no lever, but on the footbridge alongside you is a large stranger. The only way to save the five people about to be run over on the tracks is to push this stranger off the footbridge and down onto the tracks. About 12% of people say yes. [7] The numbers remained the same, but the feeling changed.
Elliot was generally indecisive for a particular class of problems, but that's not where the story ends. In the article eerily titled "Harming Kin to Save Strangers," [8] the authors continued the long lineage of research detailing the exacerbated utilitarian choices of vmPFC lesion patients. This time, they used a somewhat more sinister 'set of moral dilemmas,' one of which being a question of whether or not to harm one's own family member over a greater number of strangers. Not surprisingly, in all of the dilemmas, vmPFC patients showed a greater propensity to choose numbers over nostalgia.
I am not aware of any time-constrained or measured studies of this sort, but I think it is relatively fair to say that, if anything, vmPFC patients may make this sort of decision a bit too easily. Is his affliction really about indecision, about large option spaces?
This deficit presents in different ways, and we can think of it, in part, as such: The ability to make a decision may depend upon whether a ‘computable proxy’ is around. For the client's documents, Elliot could sort them competently by similarity, disparity, and content, but for the life of him, he could not settle on which principle of organization he liked best. The latter aspect requires choosing what matters for the client, or how he would like to organize them himself. Perhaps he could compare different methods of organization along some dimensional axes, but this sounds like more of a solution than it probably truly is. Suppose you and I are two quite hyperrational vmPFC lesion patients. I come to you with a philosophical treatise titled 'The optimal method for organizing client files in X line of work.' It is 1800 pages, full of citations, math, and proofs. But still, at the end of it all, you say, 'well, why did you value those things, such as time and money, over, perhaps, the color of the binder notes?' What else, you could come up with your own 1800-page treatise, only ever so slightly different, and say, 'no, this is the optimal method, for you see I have named it as such.'
In contrast to this, the trolley problem, the footbridge problem, and the even darker twists, all give vmPFC patients a number, and so they compute. Socks would hand Elliot nothing tangible, because the necessary proxy is not some number or fact, and it's not in the basic context of the body either. What is necessary for such questions is a subjective valuation of the socks' worldly characteristics. Perhaps Elliot values conformity and would like a traditional white pair to fit in, or maybe he prefers the color neon green because of some positive association. Any sufficient answer invokes 'he,' the self, the subjective account.
There is much in the way of evolutionary support for this broader framing of orbitofrontal valuation circuitry having to do with subjective questions. Rodent and primate orbitofrontal cortices, [9] in part, do something like subjective outcome valuation over time: a question of 'what does this mean to me now?' How much do I value this opportunity at stealing a bit of food from this lady's kitchen? More than the risk of being caught? What if I were really hungry? What if I wanted desperately to fulfill my lifelong dream of being a world-renowned chef despite being a rodent? Of course, such a question is a far cry from morality, but the common substrate is there. Perhaps this same, more primitive area of self-valuation operations was elaborated upon in evolution to not only include self-valuation, but inter-self valuation.
3 The Computer Metaphor
Paul Cisek, in "Beyond the Computer Metaphor: Behaviour as Interaction," [10] writes, 'One can say that the general working assumption in brain science today is that the function of the brain is to convert stimuli into reactions.' Cisek goes on to illuminate the core issues underpinning such a working assumption, and how the input-output functional viewpoint of the brain leads to unnecessary difficulty in the approach of certain problems, namely those of 'consciousness, emotion, motivation, and meaning.' For Cisek, things are backwards: meaning precedes symbols.
More than anything, the drives begin to look different when the working metaphor of the brain switches from a computer to a control system. Cisek, again, writes, 'For computers there is no notion of desirable input within the computing system, and hence there is the riddle of meaning, a.k.a. the symbol grounding problem.' The symbol-grounding problem is defined as follows: "How can the meanings of the meaningless symbol tokens, manipulated solely on the basis of their (arbitrary) shapes, be grounded in anything but other meaningless symbols?" [11]
In his view, the evolutionary reality that meaning precedes symbols effectively neutralizes the symbol grounding problem. While I agree with the gist, I think an explanation is still needed.
Often, perhaps due to the computer metaphor, people tend to treat the drives as objects. The common currency view, for instance, sees certain OFC sites like the vmPFC as containing a neural common scale of value. [12]
In contrast, Damasio merely presupposes feeling. His account of the vmPFC resembles a lookup table, mapping intelligent and affective bodily context in order to dispositionally re-represent it. That is, to the very complicated and shrouded mystery of emotion, he places it fundamentally in the body and brainstem. Solms takes a similar view. [13] There are two current camps available for how to view the drives: One being through a computational lens, often treating them as objects like labeled scalars, and the other, taking them as given, somehow coming from the body and brainstem.
In the control-oriented picture, all of this begins to make little sense. Drives are not computational objects, but links in a control system. Their interpretation, problematically, is made either empirically, by a third-person observer, or phenomenologically, by a first-person experiencer.
The imperative force of what we may call a directive signal is in its position in an architecture, not in the signal itself. Think of a perspiring organism: One sensory neuron which senses heat and connects to a neuron which activates a sweating process. The signal we observers read, roughly, as 'sweat, now,' in the perspiring organism is so defined by where it goes and what it induces. Now think of a scientific observer looking at the brain of a hungry mammal. Magically decompose the ensemble(s) which appears to function as 'eat' or 'be hungry' and you recover a list: ghrelin at X, glucose at Y, stomach distension at Z. The imperative is nowhere in the list, and it is nowhere in the ensemble either, but in what the ensemble is wired to (observably), and in the history which made that wiring successful (teleologically). [14]
The implication of force being a property of arrangement for something like hunger would be that, in a way, there are many 'levels' to hunger. At one observable level of some ensemble which seems to mean hunger, there may be a reflexive wiring to some automatic process like salivation, or ghrelin release from the stomach. At another, perhaps 'higher' level, hunger may be empirically observed as (in part) some activity in the anterior cingulate and anterior insular cortices. [15] Rather than hunger as a whole being a thing many parts of the brain have computed, it is a general direction of control activity based on the bodily senses, cognition, and so on. Sometimes, this may route to very simple behavior, or it may instead need to focus agentic attention, intention, and activity upon the sometimes complicated and holistic process of scavenging for food. Similarly, the same activity in the brainstem could be 'consumed' early and reflexively by some innate purging reflex, or perhaps could be merely the innate origin (observer-relative and interpreted) of some conscious feeling of disgust which is dispositionally activated (and/or partially constituted) by the ventromedial prefrontal cortex after hearing some esoteric words, such as, 'she vomited.'
In sum, the control-oriented picture would perhaps be better suited to a verb like 'flow' over 'compute.' In this way, what the observer takes to be drives are identified by their consumers. The signals originate in the body, brainstem, but the drive is not properly located there for the observation of all functions. For example, conscious emotion often heavily implicates the cortex: Why is it that both the cortex and limbic areas are active during emotions such as disgust or feelings such as hunger?
The cortex's involvement in drives is seemingly paradoxical. It is a wonder that hunger can be represented to us experientially as a descriptive feeling with both intensity and a certain quality, and that this same representation, seemingly, is what moves us. For reflexes, this is no problem. Merely reflexive creatures can jump between impulse and action in a cool, continuous stride. But for us — how is it possible we can decide at all?
4 The Ant and the Aunt
The ant, if you cut off 50% of its leg length, walks about half the way there. [16] If you put your Aunt Sue on stilts, she still makes it to the destination, providing she can balance well enough. This is not to argue ants fit, in all aspects, this sort of criterion, only at least in this one ability of individually mediated walking. But Aunt Sue herself must account for more than where in geometric space a half-eaten bag of Doritos was left on the ground and how to get there. She needs to take into account not merely all of her drives and base desires, but esoteric and abstract concepts such as 'it matters what people think of me' and 'it is unruly in the eyes of society to get on all fours and eat off the ground', very real concepts of cognition which must at every moment battle with some drives whose sentiments are something akin to 'those Doritos look mighty felicitous right now' and yet work in support of others such as 'I care about my place in the group and would be fearful of any meaningful embarrassment.' There is no reason the ant's ambulation be modulated by such a consideration, while it is crucial for the aunt's. The ant's problem is solved by simple odometry, a biologically cheap and easy implementation. Aunt Sue's requires the drive to become a term, and yet to continue to remain a driver of her choice, and for these drives to be contrasted against one another in light of abstract, esoteric, descriptive considerations.
This ability may be called Situational Arbitration: The exchange rates of reward are set by the situation as understood from both ends of the system, interoceptively and exteroceptively, and their communication.
Less clearly than the Aunt, but entirely as necessary (and evolutionarily antecedently), the problem arises when a perception in the world does not have a preordained answer for it. An ant's ambulation can respond to a multitude of physical perceptions in a complex manner, but a cat must also be able to know what to do with the opaque grey box labeled 'Fancy Feast,' taking into account in deliberation all of her subjective desire and all of her perceptive belief. There are a host of physical beliefs about the world unavailable to be accounted for in light of the ends of agents by a preordained manner of input and output, however complex and adaptable.
The comprehension of the words 'Wendy's 2-for-1' can influence the decision of whether or not to stop and eat. Blood sugar levels may also influence the choice. Physical worldly perceived factors with different directions of fit meet in the middle to convergently decide what to do. One's force is in the perception of the stimulus itself as learned in terms of relevance and subjective value, while the other is seen in the arrangement of a long and complicated control system. If you were pretty hungry but unsure whether or not to stop, and only understood Spanish, the Wendy's ad may not provide an adequate nudge. How is this accomplished, and can we overlook the direction of fit distinction, merely treating this function as a prediction with priors, or a value function, or what-have-you?
A control system must see the mind-to-world representations of the objective in light of its world-to-mind posited drives and desires. In this way, there is a genetically posited (and developmentally influenced) subjective interpretation of reality, not only in the way it appears but 'what it means to me.' To the cat, I am some vague symbol, at once made up of the feelings of warmth, satiety, comfort, maybe love — as well as the perceptions of treats, pets, and brunette hair. A perception may need to say how things are, and yet it must move you. Cats do quite well to thrive in quite modern and artificial environments consisting of inert plastic prey who have on the side of their skin the symbols 'Fancy Feast' or 'Delectables!'
It is easy to overlook the subjectivity of choice, given it arises from physical, almost mechanical systems. Seemingly, the crowded meritocracy of subjectivity is modulated by the affect. In reality, it also creates it.
5 Emotion
It is feasible to split description and direction, and it seems to work. There can be a common-currency-like scalar taken in by the function which actually directs. Or, there may be a magical fountain of affect situated in the brain and body, leaving areas like the vmPFC to subsist as descriptive lookup functions. This is not a category of position which is unpopular, wholly irrational, or one I can argue against in a manner which can disprove it. Rather, I simply think there is an alternative which, if abductively reasoned into, yields greater explanatory power. Most accounts continue to split mind and body in what I take to be merely a more scientifically acceptable version, and the computationalist lens does well to allow for this to continue.
Cisek's work on looking through a more evolutionary and control-oriented lens is very useful, as well as his affordance competition hypothesis, [17] which I view as perfectly commensurable with my own. Overall, I think he does well to reframe our picture of cognition. Still, I view it largely as a diagnosis (and a valuable start to an answer) in need of more answers, and I hope to contribute something here.
A representation with both directions of fit is widely believed to be incoherent, or at least unnecessary. [18] Others, well before me, argue for it. [14:1] I think Millikan’s pushmi-pullyu representations (‘PPRs’; representations which are at once directive and descriptive) are seen primitively because of their principality in biological agentism, not because they are merely primitive. Millikan holds that they split off later in evolution, into descriptive and directive faculties.
What if we simply say that Elliot lost function and feeling at once because their integration occurred at the vmPFC itself? My claim is that at certain points of the brain, one function feeds both directive and descriptive consumers. In my view, the vmPFC is,
What if the representational format it constitutes from the integration of the somewhat differential (and perhaps incommensurable) representations of the limbic and cortical regions, the former of which has no obvious form or meaning on its own, is a valenced perception? A perception with pull, or a pull with perception. An apple can make me more hungry, and I can speak of my hunger or weigh my hunger against other subjective ends. This function of integration occurs, perhaps, so that the idea of the trolley and footbridge problems do not present as inert questions but offer a visceral reaction at once? A visceral reaction which takes all that is relevant into account of all else. So that the visual representation of socks is not inertly presented but subjectively valenced as seen? What if this is something of a necessary functional invention, a stamping of the formless, posited and subjective onto the worldly in order to decide? Nothing about socks themselves constitutes truth, the choice necessitates a subjective position. As I see it, there is no computed object to be tossed around, but a multitude of acutely meaningless signals of a biological control system which must work in tandem to navigate reality with some posited sense of direction.
This may explain Elliot: He could perceive, and he could feel, but no longer for this class of decisions — for this category of consideration, the integration of the two was lost. Certain perceptions lost their pull.
Extrapolating from this, the brain regions most implicated in the actual feeling of emotion should be those which most integrate limbic and cortical areas. The anterior temporal lobe, anterior cingulate cortex, anterior insular cortex, and ventromedial prefrontal cortex are some of the areas I have identified thus far. Their roles in functions like choice, attention, valuation, and decision, are of note.
For Lisa Feldman Barrett, the brain constructs emotion by applying learned 'emotion concepts' to internal and external sensory input and the resulting bodily states. She relies on the predictive processing lens, viewing this process as a sort of active inference. [20] Arguably, she is among the nearest accounts to my own, pinpointing many of the same (but not identical) regions of interest and treating emotion as more or less a functional construction. She writes, in "Interoceptive Predictions in the Brain": [21]
I would caution against some of the stronger wording, like 'The result is a multisensory representation of the world from the perspective of someone with a body.' Nevertheless, and almost hand-in-hand, I see these areas as pertinent and doing much the same things functionally. Only, with the predictive processing lens, she continues to stop short of applying the identity claim I do. Klein argues that Predictive Processing can't get desire without smuggling it in, a sentiment I agree with. [22]
Empirically, I think there are some potential clues pointing my way. Elliot lost function and feeling in tandem (as do other vmPFC patients). One akinetic mutism patient, after having recovered, said about her experience: "I really had nothing to say." [23] Insular-damaged rats cannot properly recall the value of the food, affecting their decisions. [24] All of these cases can be absorbed by other frameworks perfectly well. I am only offering another potential explanation, born out of a different lens, and a dissatisfaction with the current accounts, which, in the sentiment of Cisek, seem to struggle most with certain problems in the realms of 'consciousness, emotion, motivation, and meaning.'
As I see it, the site damaged in Elliot was not a lookup table using brainstem or bodily feelings. It was the site where largely world-referential processing and subjective steering were both taken into account, and the format it constitutes is one of both direction and description, at once world-to-mind and mind-to-world. To us, it is that red apple which moves us, that feeling of hunger that at once directs attention and can be descriptively used in considerations — even talked about. Such is what I take emotion to be, its role, and generally where it would be most directly constituted in the brain.
6 Empirical Commitments of this View
For these, I will focus on the vmPFC, anterior cingulate cortex (ACC), and the anterior insular cortex (AIC). They are well-studied (and hotly contended on the particulars of, I think largely owing to the ideas circled in this post) and my view should hold each of them to certain predictions which may support or falsify this account. This is my best effort currently, and I would love feedback.
Also, it may be important to note that I am not arguing these areas contain emotion, are its only constitutors, or the only places in the brain we can find affect… The anterior temporal lobe, for instance, is also of note based on this theory, and I am working out some intuitions which pertain to that now. What I am saying, is that the nature of their being highly integrative of both world-referring and subjectively posited content, for use in valuation, attention, decision, choice, and so on, makes them highly implicated in agentic decision as well as emotion, and so they are important areas of note for this view of emotion, not the only ones. The very ‘force as a property of arrangement’ and control vs computation points from earlier should further shed light on the extent to which this stuff is not easily narrowed down, and why I am quite anti-localizationist.
Prediction #1: Reweighting from Belief Alone
This harkens back to 'Situational arbitration.’ Belief update, even on its own, should be able to re-weight different desires and impact choice. Pretty much every account should predict this, but I’ve included it anyway.
Do I think belief and desire can come apart, empirically? No. What I mean is that merely seeing that Wendy's sign should constitute emotion, unless particular sites of this integration were damaged (even though you could still see the sign).
Prediction #2: Feeling and Function Do Not Dissociate, for the Same Class of Decisions
The same domain of decision/valuation/behavior should be affected both in terms of the typical function and the feeling. Elliot understands but it doesn't move him. Smokers with insular damage have less of a conscious urge to smoke as well as a lessened habit of smoking. [25]
The way I would explain that is the AIC does not direct their attention to the habit as much. The feeling of the urge and the habit should lower in tandem. Another: Anterior midcingulate cortex stimulation leads to increased heart rate AND a will to persevere. [26] I would explain this by saying this area both initiates involuntary reflexes and prepares the agent consciously. Certain similar regions like the posterior cingulate should, if stimulated, lead to some effect but no/less conscious affective change — function and feeling are more separate.
To be clear, the cingulotomy and pain asymbolia literature [27] may seem like a discriminating case against OR for me depending on one's interpretation. This may stem from potential confusion with the directive and descriptive dichotomy: It may appear that the ACC is broadly directive and the AIC descriptive. This is not the sense in which I mean the two terms. I view both of them as being both directive and descriptive, only in different ways. Look at it like this:
For instance, the Cingulotomy patient that feels pain but does not care for it has the AIC intact, so it seems 'oh description but not direction.' This is a rather phenomenological interpretation. My claim is more specific. Each site constitutes emotion in a different way, and one presents more imperatively and the other descriptively, phenomenologically speaking. However, both implicate belief and desire at once. The AIC being intact still allows for (or partially aids in) both:
A drive is described and it still ‘directs’ its function. Now, I am aware of the studies that generally show things like (but not exactly this) such a patient (no imperative aspect of pain) may still walk on their broken ankle, or may chew on their tongue. Of course, with no phenomenally imperative aspect of pain, patients would care less. I would say it would be far worse with zero phenomenally evaluative aspect of pain too.
Prediction #3: Reflexive vs. Deliberate Behavior of the Same Underlying 'Drive' Displays dissociative potential for observed activation in AIC and ACC vs brain stem and hypothalamus (other regions may be carefully added in this contrast too)
That is, from outside, hunger can look like ghrelin increasing OR like your roommate being rude to you because his blood sugar is low, to the observer. The former sort should be able to (not that it always will) occur without activation (or with less of) of AIC and ACC as well as without conscious knowledge of any hunger. The latter should not.
Prediction #4: Choices with an Answer Should Not Depend on These Sites
Elliot can probably answer a multiple-choice moral exam with numbers as proxies. He probably cannot pick out a pair of socks very well.
7 Objections & Replies
Does Damasio not predict all of this too?
I would say, empirically, our accounts end up being quite similar. The difference is more in the why/how than the 'what happens.' I think both matter a lot. Accounts which do not place affect in the body à la The Hidden Spring but still maintain a descriptive and directive split at these cortical regions may offer more in the way of discriminating comparative tests, and that is a good thing because I view them as the stronger accounts compared to his own.
Are you comfortable with the fact your first post on LessWrong is about consciousness?
Not particularly, but this is how the cookie crumbled.
In all seriousness, this is a theory of emotion. That said, delving into this has changed my mind about consciousness and there is a route to taking this theory and simply extrapolating it to feeling in general: The reason it feels like anything at all is to perceive and think about the concrete world in light of arbitrary goals, and phenomenality is that representational format. A representation with both directions of fit, the invention of the subject. I think it is more important to not put the cart before the horse, and begin with this idea. Nevertheless, I also think we do need more functionalist/materialist theories of consciousness which can be applied to AI as well, and my account may lend well to this.
Would a Value Function Count?
This deserves its own post, but in short, not exactly, and I think Ilya (in the aforementioned interview) is thinking about it slightly off. My intuition which I am working on developing is that a value function is reflexive, eo ipso.
Is this falsifiable?
I intuit the best route to falsifiability may be prediction #2 above. Still, I don’t think any of the current accounts on emotion are particularly falsifiable and this sort of thing will be simply settled upon over time. See Mark Solms and Lisa Feldman Barrett talk with one another, here [28] and here [29] , and you may get a sense of what I mean. What is the explanation that strains least in order to answer questions, or fit evidence?
Related Accounts
My ideas depend on, overlap with, etc., to some degree or another many accounts which have already been cited, like Barrett, Solms, Millikan, Cisek, etc.
Dickinson and Balleine are overall likely the closest account if you factor in the tentative consciousness point. [30] However, I would treat the feeling as one representation with both directions of fit while they have an interface. The realm of argumentation and philosophical treatment is also quite disparate, and I offer some (different) neuroanatomical claims.
In terms of overall proximity to another account which has not yet been mentioned, Bud Craig is quite close, and I recommend his book How Do You Feel? [31]
Joseph LeDoux’s work is somewhat similar, and I view his terminology separating ‘survival circuits’ from emotion to be quite useful and basically how I think about it. [32]
The framework concerning many of the same topics, well-known around here, is of course Steven Byrnes'. [33] I am a fan of his work, but I locate the primary difference between our views (insofar as I understand his) in that the verdict on certain choices, for me, occurs in his Learning Subsystem. In other words, for me, valence can be supplied by the cortex, not just predicted by it.
Antonio Damasio, Descartes’ Error (1994). Primary source of Elliot’s case and all of Damasio’s quotes about it. ↩︎
The socks example is likely a retelling, here from Ilya Sutskever on the Dwarkesh Podcast (emotion discussion starts at ~9:39). I keep it instead of client files or any other factual decision Elliot had, because it lends itself to my rhetorical ends. ↩︎
Damasio, Descartes’ Error, chs. 5–8. ↩︎
Elizabeth Anscombe, Intention (1957), §32. ↩︎
John Searle, Intentionality (1983). ↩︎
Colin Klein, What the Body Commands (2015). ↩︎
Hauser et al., “A Dissociation Between Moral Judgments and Justifications” (2007). ↩︎
Thomas, Croft & Tranel, “Harming Kin to Save Strangers” (2010) ↩︎
Schoenbaum et al. (2009) on rodents; Padoa-Schioppa & Assad (2006) on primates. ↩︎
Paul Cisek, “Beyond the Computer Metaphor” (1999). Source of all Cisek quotes used. ↩︎
Stevan Harnad, “The Symbol Grounding Problem” (1990). ↩︎
Levy & Glimcher, “The Root of All Value” (2012). ↩︎
Mark Solms, The Hidden Spring (2021). ↩︎
Ruth Millikan, “Pushmi-Pullyu Representations” (1995). ↩︎ ↩︎
Tataranni et al., “Neuroanatomical Correlates of Hunger and Satiation” (1999). ↩︎
Wittlinger, Wehner & Wolf, “The Ant Odometer: Stepping on Stilts and Stumps” (2006). To be clear, my sentence is not to be taken as a hard stance on ant ability, nor do I think the comparison is entirely fair to them. ↩︎
Paul Cisek, “Cortical Mechanisms of Action Selection: The Affordance Competition Hypothesis” (2007). ↩︎
Michael Smith, “The Humean Theory of Motivation” (1987). ↩︎
van den Heuvel & Sporns, “Network Hubs in the Human Brain” (2013). For the term “connector hub” only. ↩︎
Lisa Feldman Barrett, “The Theory of Constructed Emotion” (2017). ↩︎
Barrett & Simmons, “Interoceptive Predictions in the Brain” (2015). ↩︎
Colin Klein, “What Do Predictive Coders Want?” (2018). ↩︎
Damasio, Descartes’ Error, ch. 4. ↩︎
Balleine & Dickinson, “The Effect of Lesions of the Insular Cortex on Instrumental Conditioning” (2000). ↩︎
Naqvi et al., “Damage to the Insula Disrupts Addiction to Cigarette Smoking” (2007). ↩︎
Parvizi et al., “The Will to Persevere Induced by Electrical Stimulation of the Human Cingulate Gyrus” (2013). ↩︎
Foltz & White (1962) on cingulotomy; Berthier et al. (1988) on pain asymbolia. ↩︎
Solms and Barrett in conversation. ↩︎
Solms and Barrett in conversation pt 2. ↩︎
Dickinson & Balleine, “Hedonics: The Cognitive–Motivational Interface,” in Pleasures of the Brain (2010). ↩︎
Bud Craig, How Do You Feel? (2015). ↩︎
Joseph LeDoux, “Rethinking the Emotional Brain” (2012). ↩︎
Steven Byrnes, Intro to Brain-Like-AGI Safety (2022), posts 3 and 6 are relevant here; see also his Valence series (2023). ↩︎