Cook, R., Bird, G., Catmur, C., Press, C., & Heyes, C. (2014). Mirror neurons: from origin to function. Behavioral and Brain Sciences, 37(02), 177-192.
This article argues that mirror neurons originate in sensorimotor associative learning and therefore a new approach is needed to investigate their functions. Mirror neurons were discovered about 20 years ago in the monkey brain, and there is now evidence that they are also present in the human brain. The intriguing feature of many mirror neurons is that they fire not only when the animal is performing an action, such as grasping an object using a power grip, but also when the animal passively observes a similar action performed by another agent. It is widely believed that mirror neurons are a genetic adaptation for action understanding; that they were designed by evolution to fulfill a specific socio-cognitive function. In contrast, we argue that mirror neurons are forged by domain-general processes of associative learning in the course of individual development, and, although they may have psychological functions, they do not necessarily have a specific evolutionary purpose or adaptive function. The evidence supporting this view shows that (1) mirror neurons do not consistently encode action “goals”; (2) the contingency- and context-sensitive nature of associative learning explains the full range of mirror neuron properties; (3) human infants receive enough sensorimotor experience to support associative learning of mirror neurons (“wealth of the stimulus”); and (4) mirror neurons can be changed in radical ways by sensorimotor training. The associative account implies that reliable information about the function of mirror neurons can be obtained only by research based on developmental history, system-level theory, and careful experimentation.
Cook et. al mark several areas that need exploring - one of which is the distinction between "action understanding" and "action perception". I can't help but observe that while many favour the idea that mirror neurons are central to understanding because they enable language or imitation, these ideas only follow the rhetoric of Pylyshin's computationalism and the Turing Test, respectively, and are vulnerable to the same critiques. This limitation leads me to believe that while we may be perhaps a step closer to solving the easy problem, figuring out the workings of these neurons is at best a baby step, and at worst a red herring.
ReplyDeleteDoes knowing that there are mirror neurons help explain how the brain produces (or any T3 robot) imitation capacity?
Delete(The hope behind the notion of "action understanding" and "action perception" was that language and meaning were somehow like movement and imitation, in that you can both produce them and perceive them. Is there anything in that? -- We'll come back to it when we discuss the origin and evolution of language, in weeks 8 and 9.)
This reading showed that mirror neurons, an inherent and widely distributed type of neuron in the human brain, is not responsible for or influenced by action goals, but rather by sensorimotor associative learning. This is a proof that understanding the goal or meaning of an action or event produced by another human is not genetically encoded in mirror neurons’ function. These findings could support computationalism’s second requirement: hardware-independence. If understanding another’s actions is not a product of the brain’s biology, could this be the result of a higher level of cognition not solely dependent of neuronal activity? And would it therefore be possible to produce this outside of a human brain?
ReplyDeleteWhether mirroring capacities (perception or production) are learned by "associative learning" (what is that?) or inborn, how does this help reverse-engineer how the brain -- or any system -- produces them?
DeleteIn any case, if imitation turns out to be an analog activity (like Shepard's mental rotation), then it is not computation (though "associative learning" can certainly be computational).
“The latter implies that the firing of MNs during action observation is, in itself, a form of “action understanding”; it does not need to have further consequences in order to qualify as “action understanding.” This claim is not subject to empirical evaluation; it is true, or otherwise, by virtue of the meanings of words”
ReplyDeleteWhen I started reading, I was wondering how it tied back to everything else we’ve been talking about for the past month. To me I was just reading about a cool neurological phenomenon. Not until reaching page 14 had I begun to realize that the authors were discussing understanding in the same way that Searle was discussing understanding (or lack thereof). For some reason, the idea that mirror neurons contribute to understanding was not convincing to me, nor for the authors it seems. I felt like there was more to the idea of understanding than mirror neurons. It makes more sense to consider it a piece of the puzzle (one of those 1000-piece puzzles) but not much more than that. In that sense, I find myself agreeing more with the associative hypothesis of MNs being a result of environmental experience as opposed to an evolutionarily selected for phenomenon for understanding.
“therefore that MNs are to a very large extent built through social interaction”
But my first thought when I was reading about MNs being experience dependent was touched on by the authors near the end of the paper “we cannot assume that the mirror mechanisms found in the members of one human culture are representative of the whole human species”. If this is the case, the purpose of mirror neurons can’t be for understanding (which further argues against the genetic reasoning). I imagined having someone socially isolated in a room for years reading books and interacting with objects, surely, they’d still understand things the way we do. I think if this person had a baseball in their room, they’d understand the concept of throwing around a baseball the same way I’d understand it.
@Lyla I also reflected upon Searle’s Chinese Room Argument as soon as the paper started discussing “action understanding” and what this really means.
Delete“Attempts to clarify have emphasized that, in comparison with purely visual processing of action, MN activity relates to the “meaning” of an action and yields a “richer understanding,” “real understanding,” or “understanding from within”” (page 180)
I pictured Searle trying to simulate a mirror neuron with computation and just exclaiming that he certainly had no sense of understanding by simply associating some input symbols with certain output symbols. In this case, I wonder if the understanding of mirror neurons is not accessible to us via Searle’s periscope, in that Searle cannot become the entire system of a mirror neuron which includes the analog input observation signal and output motor sequence signal. In this sense I am not so interested in trying to equate the feeling of understanding that I as an entire human experience with whatever understanding is/isn’t present in a single mirror neuron. I am more interested in if the behavioural competence that the text mentions (see 9.1.2.) could arise from mirror neuron activity. From my understanding of this section, if the mirror neurons do play a distinct role in the whole system of psychological function (namely that of action understanding), then their effect should be testable given it will influence the output.
Mirror neuron activity is correlated with the capacity to detect and imitate movements. What does this correlation explain, if we want to reverse-engineer the capacity? And how is it related to language, meaning or understanding? Language is an action. We can imitate the sounds of speech, just as we can imitate bodily movements and gestures. But how do we get from actions like that to meaning? The question is the same for imitation capacity whether it is based on "associative learning" or inborn.
DeleteI think that already we have established the T2 computer on its own is not going to be sufficient to reverse-engineer cognition, but I would say this mirror neuron correlation between detecting and imitating movements is further evidence for this. Even “the genetic hypothesis does not necessarily assume that experience plays a minimal role” (180) and certainly for associative learning sensorimotor experience is crucial. Somehow our reverse-engineered cognition must have access to sensorimotor experience. This is really important because the ability of my nervous system to, for instance, imitate a dance can be purely based on just watching my dance instructor. That observation alone is linked to the representation (via mirror neurons) in my own brain when I imitate the dance. No translation into language was needed, my dance teacher never explained in words the move that we were meant to do. The movement itself was inherently meaningful to me because of my previous sensorimotor experiences; I directly understood what I needed to do just by watching. I suppose besides the capacity to understand language, our reverse-engineered cognition also needs to be able to extract meaning from just observing movements, and one possibility to extract this meaning seems to be mirror neurons. The YouTube video discusses a transition from “show to tell” so I am interested to learn more how language is connected to this imitation.
DeleteStephanie, "understanding" and "meaning" are ambiguous. "Know-how" is more neutral:
DeleteYou have sensorimotor know-how: You know how to do things, including how to imitate. So do, for example, squirrels, young (and old) monkeys, many species of birds (vocal imitation) and preverbal infants. And in sentient species (which all of these are) it feels like something to have that sensorimotor know-how.
And you also have semantic know-how: the capacity to put (some of) your know-how into words. Not in the sense of being able to solve cogsci's "easy" problem of explaining how and why you can do what we can do. But you can verbalize the verbalizable part of your know-how: "Where is the cat?" "On the mat." For anyone who knows what "where" "cat" "on" and "mat" mean, they have understood what you have said, and what you meant to say. (And their uncertainty about the whereabouts of the cat is reduced: they have been informed by your words.)
The question is: what is the link between sensorimotor know-how -- as conveyed by showing and imitation -- and semantic know-how, as conveyed by words? How is word meaning and understanding related to sensorimotor know-how?
I believe the associative approach to understanding mirror neurons helps us in our reverse engineering of T3 capacities because it relies on experience dependent learning through sensorimotor connections. Thus, mirror neurons provide an example of how our T3 robot can use its sensors to not only describe a photo or a physical sensation, but to learn how to behave like a human.
DeleteThis reminds me of the argument made that we should approach learning how the human brain works by starting with a child’s brain (I believe this was a suggestion by Turing). Accordingly, if we programmed a T3 robot with the ability to develop MNs through associating sensory stimuli with motor output then we are starting with a child computer brain.
The next question is what is the link between sensorimotor know-how and semantic know-how, as conveyed by words?
How is word meaning and understanding related to sensorimotor know-how?
All I can think to say is that the sensorimotor know-how is identified as being contingency- and context-sensitive. So we attribute an action to a contextual situation and perhaps that is where some of the meaning of the action comes from. Another kernel of thought that I have is perhaps logically-related MNs hint at understanding. The paper suggests "MNs acquire sensorimotor properties whenever individuals experience a contingency between 'seeing' and 'doing'"(184). This means that there is no requirement that action performance and observation are about the same action, but there is still most-likely a reason for the relation made. This explains the existence of logically-related MNs. How does this relate to semantic know-how as conveyed by words. Well, in the case with words, one can be said to have semantic know-how of a word if they can properly use it in many different contexts. So if logically-related MNs allow for out putting an action depending on a variety of different inputs, maybe they can be said to have know-how? Thus, complex sensorimotor know-how could be the framework for semantic know-how, meaning using language to describe sensorimotor correlations facilitated by MNs produces semantic know-how.
"if we programmed a T3 robot with the ability to develop MNs through associating sensory stimuli with motor output..."
DeleteExactly. But that's not what the article was doing. It was mostly bickering about whether MN capacity was innate or learned. No model of how it would work, either way. But such models are beginning to be designed and tested. Not clear, though, whether any of the T4 information about MNs has been of any help in reverse-engineering this T3 capacity.
Not sure I understand your linking hunch from sensorimotor imitation capacity to the meaning of words, but we'll be discussing this more later in the course.
Following up on this fascinating thread and professor Harnad’s question: “what is the link between sensorimotor know-how -- as conveyed by showing and imitation -- and semantic know-how, as conveyed by words? How is word meaning and understanding related to sensorimotor know-how?”
DeleteI would like to play with this question by defining characteristics of our T3 machine. Suppose that 1) beyond its computational abilities, we gave our machine an identical motor-sensor apparatus as a human being such that, it could effectively through its senses learn the weight, shape, texture, colour, smell, taste, etc… of an apple. Given that it senses enough apples it will soon through statistical inference be able to tell what sensory properties qualify and apple and distinguish it from a cherry or a pear. 2) Now suppose that it also has MN like components that allowed it to have an “action understanding” capability. It would effectively by imitation learn to cut its apples, cook its apples and maybe even learn to distinguish and make good cider by imitation of “taste-norms”. Now this robot could start associating registered experiences of good cider tasting with the proposition “apple” and soon enough he has his own Proust Madeleines and poetry makes his heart sensors skip a beat?
More seriously, would a computer coupled to its environment with the same motor-sensor apparatus as a human being and the capacity to imitate with MN-like components be able to understand and be “grounded” the way a human being understands? If it were able to evaluate the sweetness of an apple with human-like sensors and simulate shivers on its skin when reading a good book, would we be able to say that a machine is grounded and “understands”? Is overcoming the threshold between the passive understanding of symbol manipulation and an active sensorially grounded understanding of symbol referants enough to qualify a machine as cognizing?
"would a computer coupled to its environment with the same motor-sensor apparatus as a human being and the capacity to imitate with MN-like components be able to understand and be “grounded” the way a human being understands?"
DeleteIf it could pass T3. Not if it was just an imitating, fruit-sorting toy.
(I still don't understand what "action understanding" is: do you?)
I'm writing this having only finished a portion of the paper so I may return to edit my thoughts, but there is something almost funny about this paper to me. Which is that it presents itself as debating an 'either or' approach to the genetic versus associative approaches, however in practice it is only arguing a matter of degree. In its presentation of the alternative 'genetic' explanation it states a genetic basis as the main component, but also acknowledges the impact and importance of associative learning during development. Then in turn the explanation of the paper's side of the debate, the 'associative' approach, it too argues itself as the base and then acknowledges the importance of the genetic basis as well.
ReplyDeleteIt all just seems a tad silly in its degree when both sides of the debate are admitting to the importance of the other's impact.
You're right; a lot of nature vs. nurture debates reduce to that kind of silliness.
DeleteBut focus, in reading the rest, on the existence of mirror neurons, and what, if anything, that adds to help in reverse-engineering our T3 capacities (of which imitation is certainly one).
Clarification for the Mid-Term - Will the readings and skywritings for this week (week 4 in the syllabus, but week 6 in the semester) be included in our Mid-Term?
ReplyDeleteActually, this week (4) is actually week 5 of the course (not week 6). But the midterm will only be on on the first 4 weeks of the course.
DeleteWhether (as per the genetic account) mirror neurons evolved as an adaptation for action understanding or (as per the associative account) they "acquire their capacity to match observed with executive action through domain-general processes of sensorimotor associative learning", or a mix of the two as Deirdre mentioned, the importance of these neurons in imitation cannot be ignored. Although understanding MNs specifically may not yield the answers we need to solve the easy problem, understanding the way in which they fit into cognition as a whole may help us figure out whether cognition really is hardware-independent.
ReplyDeleteAdditionally, if we were able to reverse-engineer the functions of MNs (which would require understanding them thoroughly), we could reverse-engineer imitation (or at least part of it). This could be the key to creating a robot that can pass T3 through learning by imitation - this is also reminiscent of the idea of a robot perhaps needing to pass T4 in order to pass T3. If we got to a point where we could reverse-engineer a process as complex as imitation, we would be closer to reverse-engineering cognition as a whole.
Edit: Imitation could be an analogue process. In this case, understanding how mirror neurons fit into the process of associative learning (or whether they are necessary for it) may be the key to the hardware-independence question/a step forward in the reverse-engineering quest.
DeleteReflect a little on what the existence of MNs (whether their capacities are inborn or learned) has (or has not) taught us along the road of reverse-engineering... MN capacity (imitation)! What do we know about how to generate MN capacity (now that we know that there are MNs) that we did not already know, knowing that we had a brain, and that we could imitate?
DeleteThat makes me think of Fodor's vehemence against the need for brain studies. Knowing of an imitation mechanism that might explain our capacity for... imitation sounds underwhelming, I agree, but, as Ishika said, it might be a step forward nonetheless. Unless it is determined that our hardware, including MNs, is entirely irrelevant to cognition. If imitation is part of the recipe for learning, and learning is part of the recipe for cognition, rejecting MNs outright as something that could help us figure out the easy problem of consciousness seems to me hasty.
DeleteI am not sure what you mean by "rejecting MNs outright”.
DeleteNo one is denying the correlation data. But is it not a good idea to ask what those data tell us about how the brain generates imitation capacity? Not just when and where.
Remember that it is not a surprise that the brain is capable of producing imitation (since we) are producing imitation. What would be a surprise would be if nothing in the structure or activity of the brain were correlated with imitation activity.
So you can take it for granted that for every capacity, and for the exercise of every capacity, something in brain structure or activity is correlated with it. But reverse-engineering has to cash in correlation as causation.
Even weather forecasting is not just about correlation. Simplifying [because I’m not a meteorologist!]: it is not just that atmospheric pressure is correlated with precipitation (rain), but that physics explains how pressure causes rain.
It’s that “how” that cogsci is trying to reverse-engineer.
“…[T]he associative account suggests that, in the long term, it may be possible to overcome these problems by establishing a rodent model and using sensorimotor training to induce between-group variation in the number and type of MNs present in rodent brains. If the associative account is correct, rodents are likely to have the potential to develop MNs because they are capable of associative learning.”
ReplyDeleteI think the possibility of testing this account in a rodent model varies, depending on what type of mirror neuron behaviour you’re expecting. If it’s simply a matter of stimulus-motor association, then I think that previous behaviourist paradigms demonstrate that it is possible for the rat. But if mirror neurons include ‘action understanding (as operationalised as understanding the intent of an association), I question whether this is possible. The subjects that we have studied so far include humans and monkeys, and all of the former and some types of the latter belong to a subset of species who can pass the mirror test. That is to say that they demonstrate some behavioural to recognise their reflections as themselves. The mirror test could be construed to imply both an association between an action and the perception of the reflection as well as the understanding of that association. It would be interesting to see in other species (although I’m not sure how you would do this humanely), whether this particular deficit means that intent of an association overall can’t be understood broadly or if this is just one instance that they don’t grasp. If the former is the case, does this mean they lack mirror neurons in areas that humans and certain monkeys do?
No, there is no way to do any of this humanely with other species.
DeleteBesides, even with humans and brain imaging it does not tell you how imitation capacity (or mirror self-recognition) are produced.
And the link between imitation (or MN or mirror self-rec) capacity and language or even just "action understanding" is still very vague and speculative, don't you think?
I completely agree with you that "action understanding" is almost like a weasel word, because I still don't know what Cook means by that. Sounds like he's trying to solve the hard problem of consciousness? I feel like no matter how much we know about mirror neurons (their existence, how specific their working domains are etc.) it is not at all pertinent to our quest to reverse-engineering. If a T4 robot had mirror neurons, isn't it unnecessary? I feel like if it passes T4, that is good enough. We can look into biology to find all these cool facts about our brain. Maybe I'm being a skeptic like Fodor here, but I think the more you get into the nitty-gritty of neuroscience, the more fruitless the findings are.
DeleteI think the discoverers of MNs (not Cook et al. but Rizzolatti et al.) were not trying to solve the hard problem, but they were trying to relate MNs and imitation to language. ("That I can imitate you means I can 'understand' your action: maybe understanding language is somehow like that.") I don't get it.
DeleteBut Cook et al. were mostly focused on whether MN capacities were innate or learned by "associative learning." I don't know about "associative," but Turing already noted that the capacity to learn was a huge part of the TT.
One thing that I appreciate from the associative account hypothesis is that they separate the explanations for origin and function of MNs. When it comes to figuring out how to reverse-engineer how the brain works, sometimes it feels hard to see past the individual differences to really think about how the underlying mechanisms are working. However, the way the author insists in the beginning to separate the two concepts helped clarify that for me today.
ReplyDeleteIn a previous post, professor Harnad also asks what is the link between sensorimotor know-how and semantic know-how? The first thing that came to mind was the Sapir Whorf hypothesis (unrelated: is the Sapir Whorf hypothesis the “strong” version of linguistic relativity?). But I’m not really sure how to elaborate on that more, or if that’s even relevant because I’m not sure how imitation is part of the hypothesis. I also don’t know what comes first: the semantic know-how (“this is blue”) or the sensorimotor know-how (“this is the physical word that we use to describe this thing that is considered “blue””).
Maybe in terms of reverse-engineering a T3 robot, MNs just demonstrate the need for dynamic processes, and that it cannot all be computation? Or is that already obvious…
Yes, the relation between sensorimotor and semantic know-how may be related to the (Weak) Whorf Hypothesis. The Strong Whorf Hypothesis is that language determines how we perceive the world. The Weak Whorf Hypothesis is that language influences how we perceive the world. (But what is language?)
DeleteThe sensorimotor part is being able to learn what to do with what (including what to call it). The semantic part is being able to use the words that name things in sentences that describe things. More in Chapters 8 and 9.
There are by now lots of reasons cognition can't be just computation; the imitation of bodily movement is one of them -- because bodily movement cannot be just computation (movement is dynamics, not computation).
I'd argue that sensorimotor know-how comes first. After all, babies seem to have an understanding of the uses of objects before we have indication of semantic know-how from them.
DeleteNot only does sensorimotor know-how come before semantic know-how (indeed, semantic know-how is grounded in sensorimotor [= robotic] know-how): in children the capacity to understand language comes before the capacity to produce language.
DeleteThere might be a clue in the three kinds of learning:
(1) Unsupervised Learning is learning passively from just perceiving patterns in sensory input. There are correlations in that sensory input and the brain is capable of abstracting those correlations as predictive features. Unsupervised neural net models can do that too. (GPT-3 text-processing shows how many patterns there are in human texts, and how they can be abstracted without understanding any of the words.)
(2) Supervised learning (also called” reinforcement learning” or “feedback-corrected learning”) is not passive: the learner does something in response to the input, and then gets feedback (from the consequences of its actions) as to whether it has done the right or wrong thing. This too leads to abstracting patterns and features, but this abstraction is guided not just by the correlations within the input, but also the correlations between the input and the output, and its consequences. Supervised neural net models can do that too
(Gibsonian Sensorimotor "affordances" are already either a hybrid of (1) and (2), or perhaps fully (2): Sensory “inputs” are not just static snapshots we view and process passively. We move and act in the world, so there is a correlation between our inputs and our outputs and their joint consequences.)
(3) And last there is verbal learning: language. We learn language through unsupervised and supervised learning, but once we have used those two means of learning to ground enough symbols in our capacity to recognize and manipulate their referents in the world (apples, colors, people, actions), words themselves give us a new nuclear weapon for learning still more things. (Question: How many directly grounded symbols is “enough”?)
(1) and (2) are both learning by induction (examples, correlations, feedback, consequences). (3) is learning by instruction.
So far no computational model can learn by instruction — only by (1) and (2) — because its symbols are ungrounded: unconnected to their referents (except through the minds of the human users of the models). But grounded autonomous robotic models (which cannot be purely computational because sensory and motor function are not computational) may become capable of (3) as they scale up to T3.
Here’s a zoom seminar series going on every Thursday, half of whose sessions are on grounding.
In this article by Richard Cook et al., we learned about the theories behind the origin and function of mirror neurons. According to the genetic account, MNs genetically evolved for a particular socio-cognitive purpose and to facilitate understanding. According to the associative account, MNs play an important role in sensorimotor associative learning. Sensorimotor associative learning occurs when a neuron fires both when an animal performs an action and when the animal observes the same action performed by another agent. It is also a reward-based adaptive linkage between sensory inputs and pre-existing motor patterns.
ReplyDeleteThe associative account strengthens Professor Harnad's view that T3 is the objective level of the five TTs. The authors point out that MNs are implicated in communication (including embodied simulation, emotion recognition, intention-reading, language acquisition, language evolution, manual communication, sign language processing, speech perception, speech production...), required for passing T2 (Searle's Chinese Room). T3 robots, like us, develop an understanding of the world by coordinating experiences from physical actions and interactions with others.
I read another scientific paper written by Zarr et al. They discovered that MNs even activate when reading sentences describing an action! I think this is a prime example of symbol-grounding (and maybe action understanding), which is a requirement for consciousness. You can find the article here: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3865370/
Easy for Cook et al. to say "Sensorimotor associative learning occurs when a neuron fires both when an animal performs an action and when the animal observes the same action performed by another agent... a reward-based adaptive linkage between sensory inputs and pre-existing motor patterns.". Much harder to turn that into the T3 capacity to imitate movement.
DeleteDitto for "understanding of the world by coordinating experiences from physical actions and interactions with others". Easier said than done. The demands of reverse-engineering (Turing) keep your feet on the ground.
MN activity correlates with movement imitation; saying words correlates with hearing words; saying (and meaning) words that refer to bodily actions influences other words about bodily actions. (The senior author of that paper, Art Glenberg, was here in Montreal 2 weeks ago (virtually) when one of his co-authors was giving a seminar in a series of (zoom) seminars (mostly) about symbol grounding. There will be other relevant seminars* in that series that you are welcome to join on Thursdays at 10:30-12:00. Email me if you want the zoom link.
*17-Sep Thériault Christian L’apprentissage des catégories visuelles par les humains et les réseaux neuronaux « deep learning
24-Sep Guenther Fritz Grounding word meanings in perceptual experience: A computer-vision approach
08-Oct Kennington Casey Towards Understanding Understanding: Dialogue, Robots, and Meaning
29-Oct Alami Rachid Le robot cognitif et interactif: vers les robots assistants ou équipiers
12-Nov Bredeche Nicolas Lifelong social learning in swarm robotics
19-Nov Clark Stephen Grounded Language Learning in Virtual Environments
26-Nov Lopes Marcos Grounding dictionary meaning
03-Dec Pulvermueller Friedemann Constraining networks neurally to explain grounding
Mirror neurons are fascinating but I think (as has been mentioned above) that this whole debate opposing the associative hypothesis and the genetic one isn’t very useful for the purposes of cognitive scientists. I think the more important question would be about whether or not MNs are necessary for cognition and if so, what is it that they do that a T3 passing robot would need to integrate? For instance, the associative hypothesis claims that MN development is strongly dependent on experience. I think that one of the implications of the article is that MNs are a consequence of learning mechanisms (i.e correlated excitation). Mirror neurons’ function can essentially be boiled down to mapping meaningful sensory information to motor actions so as to encode action understanding. Can a mechanism doing what MN is doing be implemented computationally or is this more of a dynamic process?
ReplyDeleteYes, that's the question. And of course coming up with the mechanism. (And seeing whether it can scale up from toy robot imitation to T3...)
DeleteHi Solim! The same thoughts were running through my mind as I read the article. . As others have also mentioned, experience and learning are important factors when it comes to mirror neurons — it goes far beyond genetics and brain structure. This further complicates the task of reverse engineering a brain. Even if we develop a thorough understanding of the brain, it is not an isolated system. The constant interactions with the world outside it are integral to our understanding of cognition.
DeleteI enjoyed reading Matt's reference to Turing's child brain proposition. Perhaps a first step to reverse engineering a brain is understanding these physically based concerns of brain structure in order to create a child brain. After this, one could focus on creating the interactions necessary for the development of cognition. These interactions might be the link between sensorimotor know-how and semantic know-how?
I’ve been trying to understand why it is that we are stopping on these Mirror Neurons in the context of reverse-engineering our minds. It seems to me that mirror-neurons and “action-understanding” bring additionality to our cognitive capacities in that they allow us to access generations of skills embedded in our culture (e.g. eating, writing and cooking) in a much faster fashion than having to develop these skills by ourselves. But would I go as far as to say that we would not be able to cognise without this ability to “download” culture? Maybe.
ReplyDeleteAs I was re-reading Searle’s experiment I started wondering if Searle would start grasping some Chinese were he to have a sensory-apparatus capable of monitoring the reactions of another mind to the symbols he was passively manipulating. There is no reason for an apple to correspond to the label “apple” if there is no need for another mind to recognise it as such in the first place. And perhaps objects only gain a meaning and label when they can be recognised by two or more minds as having that value? When an MN allows us to project and identify ourselves in one-another’s common gesture (as an MN literally allows us to do without acting in some cases!), to point a finger to a same object while ringing the same vocal cords, we learn to give a common meaning to a referent. I believe there is something in the “reflection” that MNs allow that is indispensable to meaning and understanding.
The point is that knowing that there are MNs and knowing where they are active when we are doing imitation does not explain how we do imitation.
Delete"a sensory-apparatus capable of monitoring the reactions of another mind to the symbols he was passively manipulating"? You mean piping neural imaging data from someone's head into the TT-passing computer? What has become of the TT then? It's no longer words in and words out (i.e., no longer T2). But it's not T3 or T4 either. So what would it be testing, and reverse-engineering? (I think you have gotten a bit carried away with free associations on the notion of mirroring!)
Cook et al.'s separation of the two questions: (1) What are Mirror Neurons doing? (function), and (2) Why do we have Mirror Neurons? (origin), is an interesting parallel to the 2 questions of Cog Sci.
ReplyDeleteThe article argues that Mirror Neurons are not created in our genes, but through sensorimotor experience, and I think one of the important ramifications of this is the importance of learning. When we learn to associate different phenomena, that gives us the feeling of understanding of what we're doing. So by establishing Mirror Neurons role in creating the associations, we can analyze the synapses they create to expedite areas that share associative communications. This could help us understand how the brain gives rise to the mind.
This mirrors the goals of figuring out how we think, so that we can figure out why we think.
"How" is a question about what is the causal mechanism of a capacity. "Why" is a question about what is the adaptive value and evolutionary origin of that capacity.
DeleteSensorimotor learning capacity is a good place to start, but "association" is not a mechanism (except perhaps for rote learning of paired associates). The course will later cover unsupervised and supervised learning models.
The connection ("grounding") between a word and its referent is a lot more than an "association."
The existence of mirror neurons, while interesting, I feel do not provide much information in terms of reverse-engineering a T3 robot. “The term ‘action understanding’ was introduced by Rizzolatti and colleagues to characterize the function of MNs.” The article goes on to question how accurate this characterization is. In fact, mirror neurons firing only seem to correlate with what is thought to be the act of understanding action. For a T3 robot, the ability to mimic other humans appears to be a necessary feature in order to be truly indistinguishable from a human. The mechanism by which a T3 robot is able to produce this mimicry seems unimportant. For humans, this mechanism may be mirror neurons, but if a T3 robot is able to copy using different means, why should it matter if it performs exactly like a human? What bothers me about the study of mirror neurons is that “the picture for action understanding is obscured by the fundamental problem of defining exactly what is meant by “action understanding” and how it differs from action perception.” The article argues the extent to which the existence of mirror neurons is learned or innate, but only raises this fundamental question a few times. I do not believe we will ever be able to tell if MNs are the mechanism for action understanding, simply because we will never be able to peer into another’s cognition. At least for a T3 robot, I do not think this fact matters.
ReplyDeleteThe existence of MNs (cells or regions that are active when someone makes a movement or someone else makes the same movement) does not explain anything. If the cells are doing it, we don't know how; if they are part of a larger system that is doing it, we still don't know how. If it's done via "associative learning" we don't know what model(s) can do it. And, as you note, we don't know what "action understanding" means. As for T3, it certainly would need this capacity, but it will have to be scaled up from toy-scale to full Turing scale.
DeleteSo no reverse-engineering to speak of, with the discovery of MNs. But the generalizations ("action understanding," language," "empathy") are interesting (vegan) food for thought...
As has been mentioned in this thread, Cook et al.'s paper deals with the question of whether mirror neurons are innate and adaptive––and therefore came into existence though evolution––or whether these neurons are gained through associative learning, such as Pavlovian or instrumental learning. This debate about mirror neurons does not give us any insight into the "how" of cognitive science––that is, it does not help us reverse engineer a T3 robot that can do the things that we can do and does not help us provide any kind of causal explanation as to how we do what we do.
ReplyDeleteHowever, I believe it is helpful to separate the "how" from the "why" in cognitive science (as Prof Harnad does in an above response). Despite the fact that the paper cannot give us insight into the "how," perhaps it could still give us insight into "why" we do what we do. If the authors' argument is correct and mirror neurons turn out not to be innate but rather gained through associative learning processes, then (I would argue) this gives us some relevant information about the "why": for example, this might imply that it is not mirror neurons themselves that are adaptive, but rather the cognitive processes (i.e., associative learning) which allow us to gain mirror neurons which are innate and adaptive.
The "why" question for the "easy" problem is not "why do we do what we do?" but "why can we do what we can do?"
DeleteWe can imitate others' movements. MNs do not tell us how.
Why can we imitate others' movements? What do you think?
One possibility is that it helps us teach and learn to do some things from others, especially parents and their offspring.
Another is that vocal imitation is important in courtship and social communication in many social species, especially birds (and humans).
Another (to be discussed in Weeks 8 and 9) is that imitation (pantomime) may have been part of the origin of language.
Hi Alex! This is the other Alex.
DeleteI think you make a really important point. It seems like we can accept the arguments presented by Jerry Fodor, that a lot of the discourse and experimental imaging methods for "direct" analysis of the brain do not give us sufficient answers or solutions to "how" we can do what we do, but this does not necessitate a complete dismissal of their potential contributions.
The studies that lead to the identification of MN correlational dynamics could still help guide our continuous questioning and perhaps give us clues for where to look in asking "how," like you say, with informing our ideas about "why" we can do what we do. Looking at significant trends in brain activity (meaning, what happens or does not happen and /where/ we perceive that empirical evidence) may prove to be useful or informative (reducing uncertainty) for researchers attempting to reverse-engineer human cognition.
This is possibly due to emphasis placed on it in the paper, but this paper made me realize that an important part of our learning and brain functioning process is dependent on brain plasticity. Thinking of this from a reverse engineering point of view I feel like this would make T4 especially difficult. I think accounting for those changes in hardware might be possible robotically but I don't see how we could do it in a replication of the way our brains change since I feel like that would require actual physical changes in the wiring? And wouldn't that be an unnecessarily difficult functionality to impose if it isn't necessary?
ReplyDeleteI'm not sure what you mean by plasticity, but if you mean the capacity to learn, yes, that involves internal changes, but internal changes go on all the time (e.g., when we eat and raise our blood sugar). Habituation to a repeated stimulus is a change, but it's stretching it to say it's a change in the "hardware".
DeleteAnd if by "wiring" you mean internal connections: these need not be changes in "hard"-wiring, just "soft" plasticity. Neural net models are changing their interconnectivity all the time, under the influence of their inputs and their learning rule for update activations and connections. Most neural net models today are computational simulations, not real neural nets, but they can also be implemented as real parallel distributed nodes that have activation levels and connectivity that can be modified by experience.
Even in a digital computer that is just doing computations it is more accurate to say that the hardware itself does not change with input; it is just being functionally reconfigured by its algorithm, its input, and its computations.
From Cook, R. et al (2014). Mirror neurons: from origin to function
ReplyDeleteCook, R. et al showed that the action goal or the action meaning is not encoded in the mirror neuron but rather, the neuron is influenced by sensorimotor associative learning. This can interestingly be related to Searle’s CRA. The scientific article led me to the idea that, the “understanding” (action perception, goal or meaning) is not explained in any way by mirror neurons’ activity. It is correlated with the activity in some way, but as Searle would have argued, no explanation of understanding is given. It is just a discussion about squiggles and squoggles; laying down the history of blind alterations of “forms” and “shapes”, but I don’t see an explanation there that brings us closer to understanding “action understanding”. Mirror neurons in themselves don’t explain what action understanding is but I wonder, just like I wonder about Fodor’s “where” and “when”, if collecting data about mirror neurons in a “where” and “when” manner may eventually guide our thinking and eventually lead us to infer the underlying mechanism.
There are no squiggles and squoggles in the MN findings. And Searle would not have argued against MNs since they are data about the brain, not computations. Please read the other comments and Replies about where/when and /how/why.
DeleteI recently read a paper which made me think back to this mirror neuron paper. The author Yael Niv argues that most discoveries about cognition haven’t been discovered by looking directly at the brain, and instead by looking at human behavior. He uses examples like watching out how a mouse learns a maze and experiments on conditioning, which are discoveries made thanks to examining behavior, not poking/prodding the brain. He argues that the “the list of true discoveries about perception and cognition from neural measurements or perturbations that were not already known from studies of behavior is embarrassingly short.” I think this is really interesting. When we found these mirror neurons, we weren’t really discovering anything we did not already know. We already knew we could learn from others. Also, from my understanding of the Cook et al. article, this mirror neuron discovery wasn’t profound enough to give us any new groundbreaking insight into cognition. We still don’t know how these neurons are trained or how exactly they affect us. Are they complex enough to help us learn things like empathy? And, are any of these experiments worth the cost of animal suffering?
ReplyDelete(The paper I mentioned: https://psyarxiv.com/y8mxe/ )
This paper presents a persuasive claim that MNs originate from sensorimotor associative learning, thusly refuting the genetic account of mirror neurons. The argument is that sensorimotor associative learning is salient when imitating and mirroring others, and can potentially extend to ideas surrounding communication. That is, the capacity of language and to be able to speak/talk about all that we know and understand. Therefore, these authors seem to support the notion that sensorimotor capacity is necessary to pass the T2 penpal Turing test; specifically, that T3 is necessary for T2. Although it would seem that we would have to understand how mirror neurons function, to properly construct T3.
ReplyDeleteHowever, although this reading seems to adequately weigh the genetic account vs. the associative account, it does not seem to shed light on our understanding of consciousness or our goals of somehow achieving T3. Related MNs are shown to fire at the same time and in a predictive manner (ie: contiguity and contingency respectively), but is this purely correlational? How does this provide any causal evidence for action understanding? A part of me believes that the discovery of mirror neurons have merely confirmed what we already know about our capacity to accurately imitate the motion of others. Though it may seem like progress to identify a potential correlate of this ability, do mirror neurons relate to the higher-level executive functions that researchers seem to think they do?
As many have pointed out, as well as the professor, MNs do provide us with a potential clue as to how we have imitative capacities, but it is far from conclusive.
ReplyDeleteI looked further into the full article you provided to see if I can find something that has not been discussed yet and I came across the paper written by Ayse P. Saygin & Frederic Dick (P. 219), they discuss about how MNs in humans have a large abstract mechanism. MNs in humans are involved in abstraction because they can process novel stimuli that has not existed in its evolutionary path (E.g. Robots). With this in mind, we can see that MNs can have a form "aid" in terms of helping us abstract things in our environment (e.g. behaviour).
Though it's really hard to know how far the abstractive qualities of MNs can go, especially in humans, it can give us an intimation as to how language may originated. Because as mentioned before, proper imitation signals to the person giving the instructions to imitate that they understood you. This can be seen as a primitive form of communication (though I imagine more primitive forms exist as well), and it's possible that over time, this form of understanding could have evolved into a tool such as language. Because if a person could say a word, and the other person could imitate it by repeating it to them and the person who said it can confirm that the imitator properly identified the referent, then perhaps MNs could be seen as a linguistic encoding mechanism. Though, this is definitely a narrow perspective on language acquisition despite it having support by some authors in the article.
In reading this article and following the comparison between genetic and associative account of mirror neurons, I was reminded of a issue that my pain psychology professor often cited as the source of disagreement in psychological and cognitive fields: the lumping and splitting problem (L/S problem). When do we lump things (neurons, cognitive conditions, functions, tasks etc.) together and say that they belong to the same category - and therefore can have relatively similar causal explanations, mechanisms and purposes - and when do we split them - affirming that they should and must be studied as distinct phenomena.
ReplyDeleteIn reading this paper, I came to think of the L/S problem as not only a problem for description of brains, but a necessary issue that would continue to arise if we were to use the brain as a model for reverse-engineering artificial minds.
If we accept that we do not need or want a T4 level bot to pass the Turing test, we need to decide what apparatuses to replicate, in order to implement given functions. Would a T3 robot need specifically designed mirror neurons (as the genetic account argues we have), or would it be enough to have apparatuses set up to support domain-general associative learning, from which we could trust that some sort of specialization would eventually arise to do the work that mirror neurons do (imitation, action perception, and the infamously vague "action understanding"). What if mirror neurons (or a close artificial analogue) did not arise? From the perspective of an engineer, would it matter?
As many commenters have pointed out earlier in this thread, the existence of mirror neurons and finding that our actions and the actions of others are correlated with their activity does not tell us how they work, and it also does not provide a clues for how to reverse engineer. The authors have made a strong case for the associative learning account as opposed to the genetic account, but I again I echo some of the earlier comments that mention that associative learning is a largely opaque term. What is it exactly? What are the causal aspects of it that would helpful for us to reverse engineer? These questions remain unanswered.
ReplyDeleteHowever, coming from a perspective much later in the course, mirror neurons could have important implications in the area of language evolution. The article states:
"Thus, the localization problem notwithstanding, insofar as current experimental data provide even early signs, they
suggest that MN activity may make some contribution to action perception and imitation"
We know from later articles on language evolution that there is a possibility that language began in pantomime. If this is the case, we have to ask ourselves: what capacity did we have (and that we still have) that allowed us to derive meaning from pantomime (interpret and mimic the pantomimes of others) and 'ground' the pantomimes in their referents (knowing from someone's abstract action what they are referring to in the real world). A possible answer to these questions might have to do with mirror neurons, who have been implicated in imitation and 'action understanding' (although as the authors point out, it's not entirely clear what action understanding means).
Overall, I find Cook et al.’s argument for the development of mirror neurons through associative learning quite persuasive.
ReplyDeleteI think that this perspective could have interesting implications for the study of groups that show abnormal MN activity, such as individuals with autism spectrum disorder. While I know that there is a lot of debate about whether people with ASD do indeed have MN dysfunction, the theory seems to be supported when applied specifically to social contexts. For instance, it has been shown that people with ASD show less MN activity when observing and imitating social actions such as facial expressions, and the degree of activity is correlated with the severity of the disability. This effect does not seem to be present in non-social tasks. (https://www.sciencedirect.com/science/article/pii/B978012391924300017X).
The associative learning theory in a sense reverses the previously assumed direction of causality in the association between MN dysfunction and autism. Rather than believing that people with ASD have a “broken mirror neuron system” which causes them social impairment, it may be that social impairment prevents mirror neurons from developing. This would also align with the fact that children with ASD tend to show a lack of imitation, and lesser attention to social cues such as facial expressions. One study showed that when observing a demonstration in an imitation task, children with ASD watched the action but payed less attention to the demonstrator’s face (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6952170/). This lack of attention to faces and other social information could explain why MN impairment seems to especially be present in this domain, as the associations have simply not been forged.
Anyways, I went down a bit of a rabbit hole looking into this as you probably tell, but I think it is a really interesting rethinking of mirror neurons.