Consciousness Live! with Steve Fleming (S2 E5)

0:00 well we're starting all right we are lying joining me today is very special guest that's deep Fleming Steve for those who don't know you can you say a little bit about who you are and what you do sure so I'm a cognitive neuroscientists and psychologists based at the Welcome Center for human neural imaging which is a University College London so we have a work where a group of researchers looking at different aspects of the mind and brain function

0:33 and we all have a common interest in using brain imaging technologies MRI and M eg and my research group is particularly focused on studying ethical mission and conscious awareness very cool you left out your most your most impressive qualification which is former keyboardist for the William James tree that's correct yeah yeah so I know Richard well through together in fact I was looking at our set list

1:06 they recently because it came up in my Dropbox somewhere and I discovered the sheet music for how insensitive and I became really nostalgic yeah I know what I know I was hoping that we could have a reunion at the last assc but it didn't work out it didn't work out that time that one was going for it but we'll leave that as a standing threat to anyone yeah okay so yeah you do a lot of interesting work on metacognition you also do a lot of like kind of popular

1:39 writing on on these issues right because you're right I've seen that stuff by you and like a on and you know various kind of I guess more generally accessible places so do you think of yourself as kind of a science communicator as well or is that kind of not how you think of yourself or no I think it's important to have that outlet I mean not everyone feels comfortable in doing that and I can understand that like different people have different levels of interest

2:11 in kind of science communication for me it's something that I do feel is important to kind of sometimes step back and and consider the bigger picture like why would people be interested in this stuff more generally so I find actually doing that kind of writing really helpful in terms of like interacting with people who maybe don't think about this day-to-day you know when you write a scientific paper then you know that there's going to be a handful of people in the world who are going to review it and then maybe a slightly bigger and

2:44 fuller we're gonna read it but you're kind of all in this bubble and you're just talking about this one issue it might become so important like you know what are the properties of method d-prime and everyone debates about this for like months and years and years but you know no one picking up a copy of Scientific American is gonna be necessarily that bothered about that particular issue but they might be broadly interested in metacognition so you're kind of doing that kind of writing I think is quite helpful in figuring out the bigger picture yeah and it's good I mean I think you're good at it too and it's nice to have

3:17 people there that are doing the science still and also doing that because I mean sometimes I guess I have the perception that a lot of people doing that or not doing as much of the science anymore they kind of step back from running the lab or actually designing experiments and I'm focusing on that sort of stuff it's a balance right because I actually think it's getting better in terms of just the culture surrounding people doing a bit of science communication as part of the day job really because I think a few years ago it was considered

3:55 maybe not a great career move to do a lot of science communication on the side because people would maybe perceive you as being less serious and I don't think that's you know a legitimate critique and I thought really in terms of like people often as you say people who are doing the science often can also you know write about it in their in their area of study and so I think there was just a bit of a reluctance to do it and that's been

4:27 that's been dispelled different recently by lots of you see me a well-respected people here and put out popular science book to written articles for places like iam so I mean it's getting much more normalized and which is great exactly ya know it really is good to see that sort of stuff well yeah so I guess I'll ask about some of that other stuff later but let me just sort of switched it to ask directly about metacognition so one of the things that comes up a lot in thinking about

5:02 metacognition is the difference different ways of thinking about it so obviously one way of thinking about it is thinking about thinking another way of thinking about it is as monitoring and specifically with reference to notions of control like which doesn't necessarily have to do with self-reflection so I'm wondering exactly when you say you're studying metacognition do you distinguish between these sorts of things and you think of yourself as studying one versus the other or do you think they're really at bottom the same yeah it's a good question I I think I mean I think in

5:38 often in studies of metacognition in humans there's almost like this kind of taken for granted assumption that we are seeing something relatively explicit what asking people to reflect on how they're performing on different tasks so I think often it's taken as read that the the monastery on the kind of self reflective aspects go hand in hand and but you when you move to the animal medical mission literature these debates are really fierce like people don't want

6:12 to maybe ascribe the ability to self-reflect to to non-human animals but they're happy to say well they can monitor that before that's right so I think that distinction comes up really strongly in that literature but in the human literature I think often people just think of them going hand-in-hand for the most part but I mean I made my views on this it kind of changed over the years and I now do you think that like there's good evidence that you can do quite sophisticated monitoring and control without any conscious awareness of that monetary in control so I think they're kind of

6:45 with clever experiments like things like looking at dissociations between things like error monitoring and your explicit awareness of whether you've made an error when you can kind of start to dissociate those things out then it seems like these are different qualities of the system experiment in mind you're saying that kind of catches us out yes a there's a really nice experiment done by

7:20 loganing Crump published in science a few years ago where they looked at skilled typists and they had this clever stuff that where the typist were typing they were audio typing – they're hearing things over the headphones and they were typing at speed and what they could then do is occasionally the typist would just make an error just naturally but and they would then ask the typist to kind of monitor whether they were making errors or not but the clever

7:54 manipulation was that sometimes if they made an error they would actually correct it for them so what appeared on the screen was the correct letter because I really common is that they would sometimes in certain area without the typist having made an error right so there's these kind of there's this like two-by-two dissociation did you actually make it error and what what did you see on the screen right was it an error or not and what they found was that like implicit markers of error monster eating things like slowing down on the next keystroke

8:27 we're affected by the objective status whether you made an error or not but your met your explicit reflective judgments or did I make it our or not we're more influenced by what you saw on the screen not your your actual status of making an error so that that's a really cleaned association between the monitoring process which is going on and the explicit kind of declarative I saw I I made an error or not level so they kind of use that data to say there's like two loops of our monitoring

8:59 there's they're more implicit one and then there's this more explicit reflective one so the the implicit one is the one that's governing the behavior of the slowing down yeah and that's the idea not around and the more explicit one then is seeming to track I guess what their subjective experience would be like whether they where they felt like they made an error that's right that's right and do you think that this generalizes I do you think this is specific to this kind of task or do you think that this kind of because I know there's a kind of general

9:33 what we're discovering it's a general kind of disassociation between the ability to do tasks and our ability to be aware of how good we're at doing the task but this seems to be going a step further and saying that in fact we had to make a further distinction between different ways of being aware of how good we are doing the task yeah I mean I think I think the I think well that kind of experiment shows that it's possible to have some level of monitoring and then control without any conscious judgment of how we're doing on the task

10:08 and so I think I think it's just you this this starts to pull apart kind of different levels of representation so I think there's a the dominant model in psychology in terms of the study of metacognition has been there's like lower level stuff object level stuff and then there's matter level stuff that's monitoring the object level and I think that what these kind of experiments are showing is that it's just not a simple fact that like you could have kind of intermediate levels of monitoring that don't shade into full-blown

10:44 reflexive awareness of your performance but I think that's a somewhat different question that's almost like saying what signals are being used to control behavior which is a somewhat different question to saying what's the can objective performance and metacognitive judgments dissociated so in the sense it's like going beyond the standard paradigm because like here yeah often when we're studying medical meter in the lab were not really looking at situations where people can

11:19 adaptively control their behavior on trial a trial basis by using these kind of monitoring signals right we're looking at cases where we're just interested in this dissociation between explicit awareness and task performance but there's probably you know shades of gray in between those two right so do you think this is like an area for more research obviously but do you think this is a confound and some some of the more general results that we're finding for making too hard and faster the dichotomy here yeah it's interesting I wouldn't

11:55 associate is confound I guess it just suggests that things are more complicated than initially presumed and I think also it puts a bit of pressure on the terminology we use so I think there's you know a pull towards simple dichotomies right you know it's clear that the brain is much more complicated than a dichotomy in terms of these levels of processing and presumably you know the way the way behavioral control

12:29 is organized is also much more complicated than that now right there's an interesting I guess this is where we showed more into consciousness because there's an interesting thing there where people are still quite tempted by the idea that you know there still is this dichotomy between implicit monitoring control and more subjective judgments about performance yeah but I think I think you just made it just makes it clear that kind of the original two-level architecture is just not gonna be sufficient yeah so well I'm almost getting distracted by what you

13:03 just said but I guess what I was trying to what the thought was was that if we're looking for especially at animals behavior and we're finally well ok so here's some kind of control of behavior which implies or suggests that kind of monitoring and then we say okay well they're doing metacognition and a metacognition in humans is associated with self-reflection and so here's control and social reflection but it seems like what we're what what you're suggesting is that that's too obviously too quick because it may be in between cases and so that's the kind of question

13:37 that's the kind of thing I meant by that confound was maybe that's I see I see some of the things that we're saying are count as evidence of metacognition might in fact be something in between that well is a kind of metacognition but something we'd want to distinguish from I see yeah no I think I think I'd agree with that and I think that you know often when I talk to animal medical mission researchers they're like you got you guys study humans you have such an easy ride because like no one in your journal articles are you reviewing them and say is this really conscious you know but I think it's a you know we

14:13 stick a human in there like a physics lab and we get this through six with the trials of the boring task and we think every single confidence judgment they're making is somehow this like you know Cartesian reflective really clearly not right I think that you know having done loads of these experience myself I think it is the case that you know in general you are kind of adopting a more reflective stance when you're doing the confidence reading where is things like just opting out of a decision might be

14:48 tapping into a more implicit level but I think I think yeah absolutely these questions are not raised often enough in in human and none more research I think on mass cognition because there's just a there's a paucity of the labels we have available and to kind of start studying these it start kind of discussing these these intermediate cases but also there's just a lack of paradigms I guess to like start associating them out so often you know I think it would be I always think it would be cool to try and study these things like together in one

15:23 paradigm like have a marker of performance of implicit master cognition and more explicit reflective judgments and there have been a few paradigms that have done this exactly like this typist study that I mentioned but there's you know there's not enough the mother yeah and it that there should be more of that yeah yeah so when just when you were just a second ago this is kind of tangental but you mentioned talking about consciousness and saying that there there seems to be

15:56 a temptation for our dichotomy were you suggesting that maybe these kind of results might suggest we shouldn't be tempted towards the dichotomy with respect to [ __ ] or oh I wasn't exactly sure what your what that little tangential point you were trying to make their way yeah maybe it was just tangential someone who thinks that their shades of consciousness or levels of consciousness I know some people talk this way but I'm not sure if you're you were one of them I think that it is helpful to have different to have

16:32 terminology that allows us to discuss different types of consciousness so I you know I and you know more about this than I do but like I it feels to me as though when the block was writing his famous papers on distinguishing different phenomenal access conscious monitoring consciousness there was a sense in which okay we need to start you might not agree with that distinction but we at least need to start kind of me adopting finer grained terminology yeah and I just get the sense I don't know my perception of like recent SEC meetings maybe we're sliding back towards a more

17:06 monolithic and terminology and we just now talked about consciousness now you know we distinguish things like levels and states levels and contents right but we when we talk about conscious contents there doesn't seem to be so much this discussion between dissociating different meanings of that word right so I mean yeah because I think that's an important point I mean it seems when you talk about levels of consciousness we

17:40 can talk about you know how vividly you experience the contents whether you know and and obviously we do this kind of stuff in the lab to write say rate the visibility of this stimulus or or something like that and so visibility ratings or something like that might correspond the generally you know how strongly experience the stimulus was but that's a difference in the kind of content you might think the level of the content not whether there's a vague kind of consciousness around so I think it's hard to make sense of what it what it

18:14 might mean to have vague consciousness as opposed to big content now of course if we're talking about being a learning aware so kind of you know arousal levels then there's a different kind of case to be made there so you know hypnagogic reverie or some in-between kind of state where you're kind of sort of semi asleep but send me awake that that that's still a case whether you either conscious or not conscious so at least that's what it seems to me so I'm not sure if you were suggesting that maybe we need some gray areas 'm or a few words agreeing with the points about the content being vague

18:48 but not the level of consciousness or i guess i guess i was thinking of it more along the lines of you know it might be that it's useful to adopt a distinction between two kind of reexamine the terminology that we used to use in terms of things like distinguishing between and aspects of self reflective consciousness versus aspects of maybe primary phenomenology that doesn't come along with self-reflection i think sometimes these these these uses the

19:24 word consciousness get get conflated you know you can have two people who are just talking past each other because you know one person means kind of metacognitive awareness and the other person means and well no I want to strip all that away and I think this you know we we saw some of those debates at the recent SSC conference you know people like rocky my like saying no you want to strip away all the kind of high-roller stuff whereas you know people that had cram now were saying you know that that's what i think is the most interesting part to study when we're

19:57 thinking about consciousness and i guess like you know maybe you can reconcile those two and we'll may be one of the you maybe you're both right partly but we we can't just have one or the other we need to kind of think about how both index different aspects of conscious experience yeah so I mean it sounds to me like what you're saying is is partially that the way the science of consciousness has kind of unfolded is there's been a kind of fleeing from thinking about theoretical

20:30 concepts like distinguishing different notions of consciousness and being a being a somewhat more worked in tune with some of the more philosophical questions like what kind of thing are we talking about when we're talking about this kind of thing consciousness as opposed to that kind of thing monitoring or something like that and then if you don't have that kind of debate and obviously there's always kind of lip service to it but but a lot of times it seems to me that scientists sort of see that as a waste of their time to be cutting these distinctions

21:04 and making all these fine and fine distinctions they want to get to the data but of course if you don't have the kind of theoretical stuff first then the data is not really going to be all that meaningful – yeah yeah no I think that's right I think we want to least agree on a kind of set framework for running the experiments in right otherwise you just you would agree that there isn't one right now I mean that's basically what you're saying yeah I think it's evolving and we just shouldn't you know I think it was really nice your discussion with Matthias Michele about the history of

21:36 consciousness science you know we need to keep in mind our history right – really not forget you know useful distinctions that were perhaps made a few years ago that now get brushed under the carpet right yeah I mean part of a part of that it seems like something that the society like the asse could could take a role in I mean one of the things we kind of would help would be like a canonical text line or something like that which there really isn't one for this kind of know there you go maybe maybe you just you know you're offering

22:10 to sit to take on this role yeah they have my writer yeah you're offering to supervise no it would hell it would help though I agree no I mean I wonder if this is something more general but I wonder some people have kind of speculated that you know cognitive science or cognitive neuroscience is kind of dissipating in a sense and you don't find like a departments being named cognitive neuroscience anymore as much as in it's a couple of famous ones they've even

22:43 changed their names recently and it's been kind of absorbed into just the neuroscience and the psychology departments in a lot of these in a lot of places but do you I mean you still see cognitive science as a distinct thing from neuroscience and in psychology that should have its own methodologies and so forth or how do you view that relationship I mean I I think coming to science is draws on lots of different fields so you know it takes

23:19 elements of neuroscience takes elements of a eye and computer science and elements of psychology I think I think yeah I think it's true that you know certainly be sharp distinction between say a neuroscience department psychology department have certainly become blurred with the advent of brain imaging and cognitive neuroscience but I still think that you know there's a there's a useful you know cognitive science slash cognitive neuroscience I think you know

23:54 basically inherits a particular set of ways of thinking about information processing kind of adopting more computational views of what the brain is doing and how that might support various mental functions so I I think that it's almost like you're adopting a particular set of shared assumptions when you're doing research with in cognitive neuroscience and usually that's a more it kind of information processing model of how brains working I see okay yeah interesting and actually

24:30 brings me this kind of your recent proposal about some of the a computational model about what might be underlying metacognition which I wanted to ask you about this is your your higher-order state space model which are yes or no underlying metacognition is I think that's quite a judgment of awareness sorry sorry yeah so I think going back yeah this kind of feeds into the discussion we're having before so yeah I think it's more this is more this kind of one of my first forays into try and trying to do more on conscious awareness per se rather than just

25:05 thinking of it as something that kind of comes along with or often comes along with not to come you soon yeah so but yeah so sorry about that I misspoke but this is the higher-order state space model yeah yeah and you call it Hoss I guess I guess this is one of the first times that I'm called yeah Cooley hops out loud so yeah cool okay so can you tell us I mean a little bit about what the model is aiming to do

25:38 and and then I want to talk about some of the predictions that makes which I think yeah sure so I guess I guess the first thing that I think is an important point to make is that my my view on like the way in to starting building a model of awareness is very much by starting with the data we have and that really are reports you know we have as humans disability to report about our

26:13 experience and I think again this is maybe a broader discussion about the history of consciousness science but I think reports have also got a bit of a bad rap recently you know they kind of they they just seem so impoverished you know it's not as if my conscious experience is just a flick yes/no binary thing right so and this is also I think interestingly underpins these you know Speirs debates about overflow and so on I think in a sense we've been distracted by models such as signal detection

26:52 theory into thinking that reports are just one big artifact that we need to get rid of when we when we're doing consciousness science and I I think that you know that's one signal detection theory I'm a big fan of and I use it a lot but I think this is one thing where section theories gotten it quite wrong that actually reports do have this richness to them that our simple models have not yet captured but I think they could capture in the future okay so I think the starting point for the Haas model was the idea that we should start

27:27 with reports and try and figure out the kind of computations that might support conscious reports and power plants and states so and in particular this is just like reports of I saw it or yeah it's it's their versus that's right not there that's right simple reports about visibility of stimulating things that's right yeah so one one starting point is the you know this is what we can take from the empirical literature on awareness reports in simple tasks you know one

28:00 thing as you as you just say no is that like these reports are actually surprisingly simple they're very abstract but they're pretty pretty low you know they're pretty simple in terms of they range often from you know in perceptual experiments you're dealing with reports about you know am i aware of something or not and that might be operationalizes is the stimulus visible or not and you know that's distinct from other kind of judgments we can make about the stimulus

28:33 which are more judgments about things in the world rather than judgments about you know my experience did I did I was I aware of that stimulus on that trial rifle is doing that left or right or things of that nature exactly so then the key the the kind of the starting point or the computational insight is that if we want a computational account of these kind of abstract awareness reports we need to start thinking about what is known as factorized state space and so we can have two or we can have two

29:08 separate causes of the sensory days and so what what it is but also whether I've seen it so becoming aware of say a red tilted thing on the computer screen might depend on me having this abstract stay of things are generally there or not that's not bound up with the state of things being red or tilted and so on so that's the key that's the kind of key idea in the model right so that they'd be there's an argument in the paper that

29:42 I wanted to ask you about which i think is in the vicinity of what your timeline although I'm not exactly sure but so these some people might think that the kind of awareness judgments that you're trying to explain are not our judgments about things in the environment rather than things about one's perceptual model at a certain time right but you want you say there's two real there's two reasons why we should think of these kinds of judgments as really being based judgments about a perceptual model so

30:16 can you tell us what the what what what reasons do you think there are so you I'm not sure what the question is the some reasons I think that these are judgments about awareness rather judgments about that now in the world yeah exactly yeah yeah so I think I mean I think that it could turn out and this is more of an empirical question that you know the way we do judgments the way that we detect things in the world also sometimes

30:50 involves these higher-order awareness states but I think the the the key argument here is that you know we when we are commenting on our experience so when I am making a subjective judgment about whether I've seen something or not then I'm in a sense making a decision about whether my perceptual model of the world has signal or noise in it right so you know it might there might have been some kind of processing going on but

31:28 this second-order introspective judgment says well you know I'm saying that that wasn't really signal in the world so in you know that's the kind of idea that it's because they're these judgments a second order or they're somehow getting at this level of awareness about perceptual content then we need to bake that into the structure the model right yeah no but I guess that's what I'm asking about is that that first step not that I disagree with that I'm just but

32:01 I'm just curious about what the what the what the reasoning is that that you think leads from one or the other because someone might say look if you're asking and I'm not saying I'm saying this but that someone might say you're asking people about whether they saw something or not then you're asking them about whether basically the thing was there or not so when they're giving you judgments they're giving you judgments about the thing out there not about their experience of the thing yeah – – but I think I would I would say back the

32:34 the fact that we can do that for basically any types of perceptual concept means that this is an abstract state that's factorized away from you know being read or tilted or anything like that right so I can ask you did you see a red thing did you see a blue thing to do you kind of did you see or were you aware of there is abstract and you know completely dissociated away from any particular types of content right I think that you know isn't that just the same as saying was it there so any way

33:11 of phrasing it is like did you see it and that makes it sound like you're thinking about your own experience but another way of saying is is it there and that makes it sound like what they're thinking about like the stuff out there yeah no I think I think I'd be willing to yea concede that point I think I'd be I'd remain agnostic about whether you know as I said before it could be these abstract States because they are you know indexing part of that they're indexing a feature of the world that is very

33:45 general that's not about any particular stimulus type or any particular property and instead it's just whether things are there or not in yeah in your in your perceptual input so so yeah so I think that that sounds that sounds fine to me I think where you might be able to start to tease those apart is by thinking about cases of reality monitoring where people can kind of distinguish between perception and imagination that's

34:17 slightly different you know a different thing but I think this model might also speak to those kind of ideas as well okay yeah that's interesting so okay let's put that aside I guess I'll come back to her in a second but because one of the things that I think is really cool about the model that you present is the since the state space is asymmetric in particular that there's only one way to say that it's not there but many ways to say that it is there that you're you get a couple of nice results from that

34:49 so one is that you you think maybe this is a computational and correlate of explosion of a kind of a widespread ignition not explosion but it widespread ignition in the brain but also it makes them empirical predictions that we should find some some differences and people's reports about not seeing the thing that we wouldn't expect from other models yeah so what so can you walk us through so what these know you you've done it you've done it yeah you know

35:27 you're absolutely right I think this is one of the reasons why I I'm excited about this approach because it yeah I mean it's still a very kind of abstract model in the sense it's not a model of how neurons fire and so on but it it could have the potential to make some course of scale predictions that we could test with brain imaging experiments so this is exactly what we're trying to get set up at the moment in in the lab in London

36:00 so the yeah the one one key idea is that so in a lot of early work looking at contrasting subliminal Prime's against Prime's that were seen and by people like stand the hand and colleagues they showed these when you say you saw something you see this kind of global ignition response across the prefrontal and parts of corset and that was interpreted by them as evidence for their global workspace model okay so the

36:34 information is shared widely throughout the brain and that's correlative conscious processing so the I think was interesting about this approach is that you can potentially explain those kind of global ignition responses as rather than reflecting information sharing across the brain instead now there I consider we think of them as showing a large prediction error on OK in cases where you say you saw something there's

37:09 this case where you didn't and the reason might you might get that big prediction error is because there's just this asymmetry in the state space there's just many more ways in which things can be seen I'm not seeing and so when you do say you've seen something you're kind of resolving you're resolving content lots of different levels of the hierarchy on the on the kind of seen side whereas when you say you didn't see something there's just really one thing that gets resolved right is I didn't see anything and

37:44 that's the same didn't see anything for everything so that's the general idea that the kind of global ignition responses reflect this asymmetry in the prediction error that you get on scene versus unseen trial so is that when you say it reflects the differences in the prediction error and that basically is taken that there's all of this levels of content that are being updated as these right yeah as the state is the deciding which would report to issue I guess but it's nice yes I guess I kind of maybe

38:17 it's useful to just go back a step and kind of fill in what I mean by prediction error there so um so I think one one important difference between the model that I'm proposing and model like global workspace is that the model I'm proposing is situated with its brought within this broader class of models known as generative models or predictive models so in in cognitive science you kind of have this broad distinction

38:50 between models that are feed-forward or bottom-up so you know they take in input they transform it in a particular way and then they spit out the the answer so you know my laptop's like that and a lot of new and networks are like that kind of the process information in there in a bottom-up way and but a generative model is quite different right so the generative model starts with specifying how the data are produced so it has this top-down aspect of it and that means you can then when you see some data when you when you feed the model some data it can

39:24 then invert do what's called invert the model which so it goes from the data to be inferred cause it so you know these are these mothers now really popular in explaining various perceptual phenomena so the idea is that in the hospital I'm kind of like bolting on a higher or a stay to standard predictive coding architectures and this high-roller state is like the highest state possible it's just when it's on it's just generates the possibility of seeing something

39:59 rather than nothing so it's like a kind of really Swiss it's not usually put in models of perception because they just start with the assumption that the model is going to perceive something right whereas when you start to can think about what is what do you need to add to the model to get it to report isn't seeing anything at all then you get into this idea that you need to add in these higher-order states and so that's what I mean by a generative model and then prediction error is basically this idea that you

40:32 know you the model starts in some state but then when you feed it input it changes the state it's in and it does so in different ways depending on you know the level of the model that you're looking at and so how much it changes from that state it was in before it got the data to the state was in after it got the data is known as prediction error and there's you know lots of theories about how prediction error is in those kind of models are tightly linked to neural activation so yeah you

41:07 know one of the most well works out of these proposals is Kalfas turns predictive coding ideas so yeah so that's the general idea so you need this a when you have this asymmetry in the state space between seeing something or not then when you say you see something then there's lots of possible things that could have been so that generates lots of changes in the system on that side of the graph and that's your hat you're suggesting possibly why we see this kind of ignition in all these prefrontal areas and in other areas as

41:40 well actually because that's where although all these levels are being updated that's why the activity that's right that's right so that's the idea in terms of like the ignition multiple levels of the system but one important difference between our our model and the kind of ignition like responses is that at the very highest level at this you know in this higher order state it's because it's so simple there shouldn't actually be much difference between

42:12 saying seen and not seen or aware because it the prediction error at that higher-order level is relatively symmetric and so that's what we show in the simulations in the paper so that then generates an interesting prediction that if you then go and actually study using brain imaging and these kind of you know these kind of awareness reports then you should be able to discover [Music] brain brain signatures of bits of both the asymmetries in this global ignition

42:47 but also the symmetry of the higher-order state right so and so if you had to I know that your model doesn't make any anatomical predictions or anything like that but if you want if you were looking at the prefrontal cortex and wondering where you might find this so I know that you know we're finding that that your your work suggests and a lot of work I think but suggest that maybe the more dorsal areas the prefrontal cortex are involved in making judgments about past things we've seen in the past whereas more medial

43:20 areas are involved in predicting how we're going to do so does then your that's your idea suggests that we would find the symmetry in a particular area because of what we of the differences here I mean I think yeah I I would be [Music] well I I probably want to may remain relatively agnostic about where we see I think I think it's reasonable to say that kind of higher-order Association

43:57 cortex prefrontal parietal areas would be would be candidate regions but that's obviously very you know broad areas of course that I think that there are a few data points out there that suggest that there are more symmetric responses in the dorsal lateral and front Cipolla cortex in studies of perception so there's a couple of papers from Andrea

44:28 his group looking at recordings in monkey just last one prefrontal cortex and they see these symmetric responses for and decisions about yes things were seen on no things were not seen and they seem relatively abstract in terms of you get the same symmetric response regardless of whether it was a decision about a square or a triangle or a circle and so on so you say that that's one data point but there's also you know there's a couple of others out there that suggests that maybe the front of polar cortex is engaged in creating

45:05 particularly abstract representations so the extent to which you think the prefrontal cortex is generally involved in abstracting away from the concrete aspects of the input then you might expect that the most abstract representations are found in these kind of hierarchically higher areas of the people and for gossip that's it yeah I think it's difficult yeah and we do have a couple of imaging datasets now that we're analyzing we're hopefully fairly soon we're gonna have some more solid

45:39 answers on this that's very cool and I mean not to jump the gun on this but how do you think you could differentiate which I mean so I assume you think there is global broadcasting I guess I should ask why do you think that this model of the that there is so if even if you're a fan of predictive coding which I think you know predict the only question in my mind is whether predictive coding is anything new or exciting or whether it's just something we've already known about for a long time without they're basically inferences at all these levels

46:11 anyway so that's a the math of it is interesting and I think that's a different question but do you think that there is global broadcast and yeah I mean either works related to consciousness but just that insurers in the brain sure I mean yeah I mean I think what predictive coding brings to the table is the emphasis on top-down connections so I know from just kind of basic anatomy that there are many many feedback connections in say the visual hierarchy there's been

46:46 more numerous than the fee for connections so you makes kind of dominant model that I think was you know a lot of the way a lot of us think about the brain intuitively is you know this idea of taking an input and then feeding it through two different regions that just anatomically you can't be the whole story because we know there's these massive feedback connections so I think predictive coding provides a really strong framework for thinking about why those are there but then if you know clearly there is also widespread

47:19 information sharing we know that a lot of higher cognition relies on what looks like relatively widespread broadcasts things like working memory you know it's very flexible and you can get lots to different information content into these into these prefrontal circuits so clearly there must be relatively widespread broadcast I guess I guess the my reluctance endorsing broadcasts is this this idea that you know it just seems quite feed-forward it's you know

47:54 if you if you think about it instead in terms of these predictive models then almost you get glued the broadcast for free because the higher levels of the hierarchy have just a wider conceptual purview compared to the lower levels so in the sense like yeah you're gonna get something that looks like broadcast in those kind of models because the information gets aggregated as you go up the tree right okay so I mean the reason I was asking that because I was wondering how you might differentiate experimentally between activity that is

48:30 doing the broadcasting stuff and activity that is doing the the in terms of the ignition which is doing the prediction error and memorization stuff but it sounds like maybe what the difference is is once you and so what I guess you'd have to look for time differences and how whether the activity comes back or yeah I mean it'd be interesting actually you know I don't know whether you know if we were to axed underhand and say you know would you be happy with

49:03 us saying that global broadcast is like a very high-level prediction error maybe he be go maybe he goes yeah yeah I think there is a deeper conceptual contrast between the kind of approach I was advocating in this paper and global web space and that is just that the idea here is that to be able to report our awareness about a stimulus it's not just enough to have widespread processing of

49:36 information related to that stimulus right because it seems like that wouldn't allow the system to say there is stimulus X but it wouldn't conceptually allow the system to say I am aware or I have seen stimulus at the the idea is that you need to have some kind of high-level state that allows you to create to make that report and that's why I don't think falls out naturally from global works based approaches yeah this is I mean I think this is like

50:09 related to a point that Hawk Juan makes a lot which is that you know you need some to do some signal detection on the signal that's coming in from the lower order area is that that sense there's I mean it's a kind of a general conceptual point since there's always spontaneous neural activity then you need some kind of higher-order monitoring mechanism that decides whether the activity is just one continuous lis generated or driven by some external stimulus and that's what what you're trying to get at is maybe your computational under a

50:42 mechanism which would do that sort of thing yeah I think me yeah I think Hank Rollins recent paper on reality monitoring and you know the earlier ideas that you need some kind of higher order signal detection I think they're very aligned with this proposal because yeah at the core of the proposals is that you need some way of figuring out whether your perceptual is just you know whether there's just noise there or whether there's some kind

51:15 of signal so it's not enough just to kind of shut all those signals forward you need to have some high level inference about whether you know there's lower level signals reflect things in the outside world or not right so what you so if you think that this is a psychologically real process which I mean you could test and you're looking into so if you think it is a psychologically real process yeah would you say would you speculate that the content of these states is something like what Hawk 1 suggests that it's either the contents is like this is veridical or this is noise or this is

51:47 working memory or do you have any thoughts about what that at the psychological level what those kinds of states might be yeah I mean I think at the moment in our so the the current model doesn't really address the distinction that haiquan makes between imagination and perception or working memory and perception so I think that's a somewhat distinct issue and so I think my understanding of what that coin is

52:21 arguing is that you know the system evolved to be able to discriminate those two cases and then after it's evolved that capacity it then almost has the ability to distinguish signal from noise for free that's kind of a byproduct whereas I'm trying to get it more of the signal and noise issue that's the starting point I have in this model and so yeah so in terms of like what psychologically it would mean to know to kind of represent the difference between

52:55 signal and noise I mean I think this is where it connects to ideas about you know reports and phenomenology I think it it does have this flavor of endorsing a particular endorsing perceptual content so if you know if the system is saying yes there is generally signal in my perceptual generative model then that to me seems like very closely connect to ideas about phenomenology sorry I don't think there would necessarily be a kind of independent feeling of or sense

53:30 of that high level representation instead I would be more along the lines of higher order theories of consciousness there like this this higher order state is what constitutes phenomenology in concert with the lower road of perception stays interesting and so it's the in concert part so I mean this is part of why I'm asking the question maybe this goes beyond what you care about and you could always tell me Steve but what what's interesting to me about this is a debate that's in the background between the Hawks someone

54:05 like Hawk 1 and David Rosenthal because some of the stuff that you know Rosenthal obviously is well known for his higher to thought theory of consciousness and according to him the kind of states that we're talking about our thoughts so they haven't touched on contents you did they're described like like you know I'm seeing red it's just or I see red or you know what I'm experiencing pain or I'm in pain or something like that so just kind of normal thoughts but arrived at automatically and so fun and so on and then Hawk one argues well you know there's this mechanism has the states

54:37 it's vertical it's noise or not and then david says oh well that's just you know the same as my theory because that's just a neurological implementation of these personal level psychological states and and so I'm wondering if like where you might fall on it on a debate like that whether you think that it matters that the contents can be described and kind of folk psychological terms whether we say that this is a thought that has content like I am seeing red or whether you whether you care about that or or if this isn't

55:10 something that you think the model is going to ever address her and so I'm just wondering what your feeling about relation to the computational stuff to the folk psychological status yeah yeah no I think it's really interesting set of questions I mean I feel slightly out of my depth on response to it but I I mean I I guess my one issue with thinking about this is I never really know what thought would look like in conversational neural terms like you can kind of start specifying these kind of models and think about either hierarchies and

55:44 models of perceptual processing stuff but when when you get to thoughts like well okay are we talking about things that can be couched in language and then do we need a model of this do we need a model of like kind of symbolic information processing like it just becomes all quite vague and fuzzy to me so I mean I think in that sense no I don't think the model really says anything about the nature of those higher-order representations the one thing I think it does bring to the table is the idea that is you know because

56:20 these because this kind of inference in these hierarchical models can be relatively I guess automatic is one way of thinking about it and only there's any need to invoke a kind of more strategic or cognitive access to this information so I can't remember where I read it but I think maybe Peter Carruthers makes this point that you know higher order theories you can think of them as just having these occasional

56:53 flickers of higher order inference right it doesn't mean you're kind of having this very heavy-duty cognitive representation you know as long as you get some kind of inference at these happening in these abstract States then I would think that would be sufficient so that might be useful for starting to think about well you know could a could a higher-order approach be compatible with data on say overflow so it could be

57:31 that like you know yes because in this kind of model you might have this pretty you know sophisticated higher-order machinery that's doing this inference but it's doing it all the time like it's doing it you know regardless of whether I'm not I'm in an experiment where I'm asked to hold certain items in working memory which might be a completely different type of problem that I'm solving so I think that you know this is complete

58:03 speculation but I would be tempted to say that you know when we start writing down higher-order models of consciousness in these kind of hierarchical models computational models then we might start discovering that actually you know this is that kind of inference is quite distinct to a whole other branch of cognitive psychology which is interested in things like working memory and decision-making and so on interesting okay and that's yeah that's I mean I don't want to put you

58:38 out on a plank speculating too much but it's good because I wonder to what extent what you're talking about is is a well how higher-order of a theory suggesting is so you said that so on on the model that you have you have the kind of lower order stuff driving the higher order states basically and so what does the theory say if anything about cases where you don't have the first-order activity or the lower order stuff but you somehow accidentally

59:11 arrive at the at the higher state I mean I think this is the nice thing about adopting generative models is that you can easily have that you can easily get that from changing the prize you know I can like simulate the model where I have a strong prior that things are gonna be seen right and then it can start telling me things are seen right in this really tiny tiny architecture that we've got so I don't think there's any yeah problem

59:44 with that I think maybe the problem would be like what it so if if that's what happens then do you count it as having phenomenology or I mean not in the toy model maybe you don't want to say that although I'm not but in our case you know yeah no I don't want to say that two dimensional maybe we can come answer this because this is an interesting question and I was pushed on this by Ned Clark at the end of my talk right and maybe we should come onto this actually but just to address the

1:00:19 the point about kind of self generated kind of illusory higher-order state so I think that yeah in terms of the way a generative model is set up there is really no distinction between the products of inference that are generated by priors kind of endogenously and the products of inference that are generated by bona fide the input right so in that sense in this particular model it couldn't tell these things apart but

1:00:52 this is exactly where I think it connects to these interesting proposals about reality monitoring and this is a debate that you know I've been having with Wow I also saw that some Gershman wrote a paper on these issues recently you know relating them to these generative so and I we had we had this conversation on Twitter between like me some and I think you know it was revolving around this issue saying you

1:01:26 know is reality monitoring enough to give you phenomenology because or give you an explanation of lighter and things happen um not not because even in the cases of top-down endogenously generated imagination when those are strong enough presumably those would also have phenomenology so so in a sense it's like yes you could tell whether my phenomenology is internally generated or not but it doesn't seem to tackle the you know the issue of whether things are there and not in your perceptual

1:02:00 experience and so I think you know and I still don't really um kind of thought this through properly in my own head but it feels like if you can combine some elements of that with some elements of the kind of models were working with where you can start addressing directly why the model concludes it is seeing something that concludes things are there or not then you might be able to have the best of both worlds all right what are the best of both worlds the best breakers will be a model in which you could have an

1:02:35 explanation for why it says seeing something or not and but also orthogonal to that it has the ability to say the thing that it's seeing is in imagination or reality I say yeah yeah so the so not only can it distinguish whether something is there or not but it can distinguish whether the thing that or not is internally generated that's right but so so but you're it sounds like so you the question that I was asking about

1:03:08 how high order is a theory sounds like it's higher order enough I mean basically ideas if the system can tell the difference no no don't but I mean so that if it's a higher order theory of

1:03:42 consciousness I think one of the key claims that we want to know about is whether in the absence of the first order state you still get conscious experience or not and if it really is a higher theory of consciousness then the answer should tend to be yes at least asunder whether it's the same or it impoverished or whatever it I think that's an interesting question but there should be some kind of experience especially if you going so far as to say is that the system can't tell the difference between a judgment generated by just the priors versus job driven by the actual lower order content then

1:04:16 there's nearly there are plenty of cases where we can't tell the difference because we know the prize play a strong role in just regular perception yeah to take the kind of lumen case I wonder you had the thought nice thought experiment in your in your paper about a darkroom versus with a person with her eyes closed versus with her eyes open and so you had the same kind of level of neural activity but the person with her eyes closed is curved going to be certain that they didn't see anything whereas the person with their eyes open is going to be well they're not gonna be certain because he's always going to be

1:04:50 a possibility that maybe they you know they're getting some some of that weak activity is representing us a signal as opposed to noise and the only difference is whether the eyes are open or the eyes are closed and exactly oh they're actually that example was almost introduced to explain why you know the limitations of the model the Eisley at the moment the current proposal couldn't distinguish between those two things that I was making actually was that if in the closed case so if you have a case where you don't know which one it is

1:05:27 yeah if you're told you're eyes are oh I mean the eyelids are opening and closing like you know that there's a shutter screen in front of you opening your TL so some of those some of that noise you may interpret as a signal and you may actually come to experience that you saw something okay and or at least that's what the model seems to predict yep yep no absolutely limitation of it I mean this is one of the themes that I've been trying to push a lot in my when I forced people to think about this stuff is that you know the kinds of stuff that and

1:05:59 that men talks about and a lot of people think is an objection the higher theory is actually an empirical prediction of the theory it predicts that you should have these kind of conscious experiences when the first-order stuff is missing or weak or you know great news in various ways and that that's something that we can try and check that we can check and see whether that ever happens but I guess so I always get slightly confused about that I probably from thinking about it from the in view of these kind of generative neural models because presumably you can say okay the first

1:06:34 day of state is missing but as soon as I make a change the highroad state but we'll have consequences right so you in a sense like maybe the middle ground is really about whether things are endogenously generated right not necessarily whether the first-order state is kind of missing or different it's like is the system working in a mode where the perceptual experiences being generated more endogenously or more bottom-up yeah yeah I see what you're saying because as soon as you make a change in system the

1:07:07 there's back propagation and immediately it's influencing what's going on at the lower outer level yeah exactly yeah I mean that's why in the in the in the cleanest kinds of cases whether whether you could get something generated just by priors is really the best way of thinking about it which is what you're saying basically yeah yeah no I think that's exactly right no but but I mean it's still you know predicts that if you have an anatomical lesion or inactivity through some other reason and and Nan redundancy if you can somehow nail all that down which yeah

1:07:40 right but you know if you if you don't have the original system you don't have a backup of the visual system and yet use them and yeah you still have visual experience then and if that's related to these kinds of higher-order metacognitive Xand that would be pretty good yeah yeah yeah I mean I think that is a I mean I think I mentioned that somewhere in the paper that there is an empirical question it feels that you know the way the way this is currently shut up he's very much saying like it really turns on these very higher-order

1:08:13 states and they in terms of the this very abstract judgment about whether things are seen or not seen right but in terms of like all the heavy lifting in terms of the content is being done by all the lower order stuff so but it feels like and I think you and a coin in your paper you made this point as well though it seems like a kind of empirical question where the balance of power lies in terms of what determines the concept it could be that just this you in both cases you need highroad of states but it could be that the higher order states

1:08:46 are also quite rich in terms of what they can represent or it could be that the highroad relatively impoverished in terms of their their content and they just index yeah things happening lower down which is exactly what this model is doing but that seems like an empirical question and you could start maybe testing it using brain imaging and you know finding cases where there have been legions to to visual areas I mean yeah I think these are important things that we're gonna have to look in I mean part of me is I mean I want to get too pessimistic

1:09:19 but you know I mean I think you wrote a paper about this I'm not a paper but a blog article about this about this paper recently about the good the microchip oh god yeah donkey kong' head the donkey I'm sorry depressing and that's that's something that does tend to depress me a little bit is that you're something we understand basically a hundred percent and then they basically show half of the things we do and see it's like C elegans right we know that circuitry 100% and that's a real biological system and yet it still has like yeah yeah it's so

1:09:54 nonlinear in the way all the bits interact that making predictions is incredibly tough exactly I mean but then the reason I you start getting I often oscillate through depression and optimism about this issue I should like like the I think the optimism is that you know it's almost remarkable that cognitive neuroscience is work athol yeah exactly if that's the case like you know people do experiments all the time

1:10:28 using brain imaging and they get a lot of the time very cool and interpretable results now there's obviously lots of you know things you need to think about when you do those kind of experiments and you don't want to fish around just to find things but yeah there's it's almost more remarkable that we've got anything rather than nothing when you think about the kind of problems with interpreting even simple system it just highlights how careful you need to be in drawing inferences about function and in lesion studies and stuff I think yes yeah it does because I know also maybe it tells you that you know maybe we'll

1:11:03 never have an explanation at the level of kind of the individual neurons when it comes you know transistors in that microchip right when it comes to kind of more software level functions but maybe we don't need to maybe staying at the rest of the core systems level is enough for kind of explaining higher cognition I mean yeah so this is a debate that's come up in you know a different context in deep dream and those sorts of things when they're when they're looking at trying

1:11:36 to decompose these AI systems and figure out what the levels of representation what do they actually do it's sometimes it's opaque but sometimes they find things that look similar to you know I guess there's one play with that so there was things like grid cells which sort of man effect manifested which I was surprised by but that was something that passing them on the tree passing on that to me but so it does seem like maybe there's yeah I some hope some hope for this kind of methodology but also that you know power the optimism have is

1:12:10 that like if things are more like a generative model so you know the kind of classical microchips the con recording and was looking at in that paper ya know generative models right whereas if the brains a generative model and if some of the computation was we can start developing a generative models then maybe things will just magically start magic exactly that's that's our hope I don't want to keep you forever but I

1:12:43 want to ask about consciousness and prefrontal cortex because some of your work has been used in this debate so I want to I want to kind of get it from your direct from your perspective directly there's a tremendous to be I mean this is kind of related to a lot of stuff women talking about with overflow and higher theories and you know because there's a specific hypothesis namely that a lot of these things that hired a serious are interested in may be found in the personal cortex and if at content has a more broader a general argument that maybe the system's underlying metacognition are

1:13:16 are serving both explicit and also this kind of thing here but they're shared in a way so coming with that is this idea that we should look at prefrontal cortex for for maybe the part that's important for whether there's consciousness or not and then when one kind of major response to that is that well there's been lots of lesions to the prefrontal cortex and we don't really find deficits and conscious experience so your work has often been taking the suggest that we do

1:13:50 but you have to look carefully so I'm wondering what exactly what the results that you've had or do you think it bears on this debate in the way that some people think it does yeah so I think I think so there's lots of things circling around this debate so I guess one is the state of Silesian studies the other is the methodology used you know could all the prefrontal ats they should be about report and not about experience so the kind of rise of these no report right

1:14:24 methods and I think the other you know the other one is it goes back to what we're talking about earlier like maybe just definitions that maybe you know one group of studying what they define as conscious awareness but actually it's more to do with metacognitive sensitivity you know ability to reflect on performance whereas another group is just looking at you know first-order discrimination and classifying that as conscious and so I think those three things often get all mixed up in this debate I think there's two things you

1:14:59 know that we often forget so the first is that the prefrontal cortex is pretty big supremely brain region especially in humans if Leslie yeah otherwise yourself like you I think there's some fantastic word that's been it's being done in terms of both comparative and after me you've been studying the functions of the human prefrontal cortex in other

1:15:35 areas of cognitive neuroscience like decision-making and and so on that could be quite useful I think for as in consciousness science so for instance work from my from from Matthew rush West lab in Oxford has shown that the paths and the connections in the human brain is very similar to the pattern of connections in them macaque monkey brain in terms of prefrontal cortex but there are really interestingly the kind of needs one or two extra regions or extra

1:16:09 kind of areas that have these connection or fingerprints that just don't appear in the macaque brain right but they're super specific in the sense that like they have a particular passive connections going to areas the parietal cortex going to other areas within the frontal lobes itself and you know we're really interested in kind of starting to use some of these sub regional analyses in our imaging experiments to think about you know what is the contribution of this relatively fine you know this finer scale enough to me to things like

1:16:45 metacognition in the lab so I think often you know the debates turn in on you you really need to think about the details of how different the different contribution the different reasons the prefrontal cortex might be making to higher cognition in general but also into conscious awareness so I kind of think it's too premature to say is the prefrontal cortex involved in consciousness because you're in a sense like that's a very broad it's a very

1:17:18 broad question and really we need to be thinking it'd be a bit like saying you know is half the brain involved in conscious as well you know probably yes but you know we need to nail down the question a little bit more clearly but but I think that you know that perspective then also sheds new light on debates surrounding lesion studies you know so I know you know where are these articles you know debating between Melanie boldly and her colleagues and and hat qualms loud and that side of things and and I think you know it's difficult because lesion studies our are

1:17:53 messy and you know you in when you're doing human lesion studies you know you're kind of dealing with often quite a few injuries that and you don't really know what connections have been severed or not without unless you've got access to kind of like high-resolution diffusion-weighted imaging where you can actually see white matter tracks and all that kind of stuff so I think it's just it's just a coarser we know that like we're dealing with a when we talk about prefrontal cortex we're dealing with a

1:18:25 system with many different sub regions with many different computational functions and it's interconnectivity is really complicated and so when you've just got single lesions it's really hard to make a prediction for what you're what you're gonna see and then also you've got the issue of compensation so on so I think with all those caveats in mind yeah there probably are we've seen in our allegiance studies where when we do selects patient so this was a study that I did with when I was in New York actually with my so when we when we selected patients who had had tumor

1:19:01 resections in the anterior prefilter course that so the the criteria was that they the surgery had sort of affected the Brahmin area ten and from like a human brain atlas but it could have been that the surgery also affected loads of other reasons it was just like that that was the criteria for selection when we selected people with those kind of legions and then we invited them to participate in these perceptual tasks then we saw a pretty massive drop in

1:19:35 metacognitive sensitivity compared some control group who had lesions elsewhere in the brain right so I mean yeah it's not you know if you were to me the you know obviously me and my RA were talking to these people coming in for the test and it was really hard to tell you know who is in which group it wasn't as if you're dealing with someone who's unconscious and sitting there in the chair and you just like press it but it's not that at all it's a super subtle change but you can pick it up using these measures of metacognition so I but

1:20:09 then there's also been you know there's a recent paper from a group in france who have found a null result with allegiance to the anterior people to cortex and metacognition and that's not published so no interesting in it yeah I think yeah leaders whose messy prefrontal cortex is a poorly understood and differentiated system and if the guys with donkey kong can't work it out with a microchip then I mean I don't think we should be having too strong a debate yes the argument from Donkey Kong with a microchip by the way yes but also

1:20:50 a year old is now in love a donkey kong atom okay yeah it lives to this day so it's just to remind me one thing which is I think and I think I mentioned this in my blog post about the donkey kong paper is that again you know once we start taking these details into account I think there is also reason for cautious optimism because there's a really beautiful series of studies a long-running programme of work by kathy

1:21:25 price he runs a group in our centre she's our center director studying language function and there's been this is not something I know much about but there's as far as I know there's been all these debates about where the different regions atmosphere contributes a language based on often patterns of brain damage because like people have strokes they can't speak but another person might have another stroke very close to that and somehow they're able to recover language and what she's shown is that you can get really good predictive power in terms of predicting

1:21:58 whether someone's going to recover their language or not following a stroke but only if you take into account the extent to which the lesion has damaged connections between regions not just individual reason right and these models are now so good that they've built the kind of a big machine learning model and they're now being used in the UK National Health Service to actually you know you get a kind of personalized prediction if you have a stroke they can now say well you've got an 80% chance of regaining language function so you know it makes sense for you to go along to all your appointments and try and have their army

1:22:30 so it's a really cool kind of practical result but it only only was obtained by recognizing that you can't just focus on kind of brain damage here okay knowledge here you have to you have to understand the connections between other regions and how they're sharing the information yeah it's it's kind of a simple point that the brain is massively interconnected but yeah people get corralled it's so easy to get corralled into thinking about like this area and that's thinking about the brain it has to be yeah yeah I mean it is it is like

1:23:05 you know I think it it's something that affects everybody in the field I think it requires a kind of concerted effort to often get out of that way of thinking because we're so used to kind of think about great imaging as areas of lighting up subtractive method is is so far in the backgrounds now of our just way of doing things that it's not questioned anymore and and or not could not not question but it's just it we don't have to remind ourselves actually but it's a yeah it's yeah exactly

1:23:40 so but so one of the cuz one of the objections that just to bring us back to the pattern of lesion stuff and with all the caveats new thing accepted and well noted so one of the kind of objections that you get to the kyoto theory is that we should expect some visual deficits when you have a large enough damage to the prefrontal area and not just subtle one so one of the criticisms you know this is Benjy Kasich has made this point but a couple other you know and then makes this point as well I think in a couple places but the idea is that yeah

1:24:13 the effects that you find are kind of subtle effects and you said so yourself and I think it's an undeniable that good to really look for them and they're there they're subtle effects and in their judgments about how well they're doing they're either overconfident under-confident but that if this is really the area that's important for consciousness then losing it we would expect something more like more more dramatic like blank sites where they just know part of it's gone or something like that yeah and and we don't seem to find anything like that so that that doesn't bother you about higher order theories

1:24:45 or how do you respond to that kind of objection yeah yeah I am really thought this through enough but I guess I mean the blindside case is interesting because in a sense the strong you know the strong dissociation there is that in a sense is the reverse dissociation where you don't have kind of sufficient input potentially to these how would you know if you take a higher order view of

1:25:17 blindsight then you can think of the reason that that might have occurred is based on some kind of like you know process that sets the criterion for whether you consider something as seen or not seen and you know that could be very higher-order along the lines of how Conlon's models but in a sense the you know everything else about the blindside subject remains intact in terms of the his ability to abstract away from

1:25:52 experience to comment on what it is like to not have that experience right so it's kind of like in a sense that how it's not as if he's unconscious right that highroad state space is there and the the most salient aspects of blindsight is that he says you know patient gy he says to you that they he has no experience you know it comes back to this idea of like absence being really important when we think about you know these heroine states so like but in the case of the the kind of prefrontal

1:26:29 lesions then I in in a way it would be surprising if the deficit would be so salient to the subject who has the deficit so you might well be able to do lots of you know I I don't know I think this is again it's an open question and maybe it comes back to that point we were debating earlier about the kind of relative division of labor between more abstract States and more first-order like states in terms of

1:27:05 the contents of experience but it seems to me like that if you did have lesions to higher-order States they would it be hard to find out because I wouldn't be able to tell you about it yeah exactly I still maybe be able to do lots of different other tasks because you know many of the systems remain intact but it'd be pretty hard to but then then it's a really slippery slope is they know what are you saying he's saying like that person is not conscious for you know a variety of important purposes

1:27:38 yeah it's really tough to work up it in a way it's kind of reminiscent of this there is this you know early set of studies by a colleague of mine at UCL Paul Burgess on the role of the anterior prefrontal cortex in higher cognition and he recognized in the 1990s that like it seemed that people would lead into this region just didn't have any problem you know they could pass all these cognitive tests like they were fine on the card sorting test and on IQ tests no it's kind of thick and so it turned out

1:28:13 that you know to find what to find out why you know to kind of find a deficit following those lesions you had to devise some really subtle tests which turned out to involve you know very abstract monitoring of tax switching and attention so the idea is that is like a really abstract level of processing it's just not really tapped into by Stanley Collins and Tess so from the outside they look absolutely fine but it's a really intriguing question like if you have these really high-level

1:28:46 abstract states that are adapted that are that are impaired the you know you might seem fine but you know does that have consequences for your conscious experience right it would seem reasonable to me that it does have big consequences for your conscious experience but it'd be hard for you to tell other people about it right and so that I mean it's it seems sort of seems like I started to push on the whole awareness reports as a kind of source for knowledge about consciousness because we're running up to the limits of that here everywhere well one of the

1:29:19 most important things that we would want to know about his way of making sure is unconscious experience in this case and it seems like maybe maybe maybe maybe we're hitting the limits of kind of narrative discussions about my experience like the kind of blind side case it's like no it really does seem you know and there's obviously been very lots of interest in interviewing people with blindside about what does it feel like to have blindside but I think and maybe that is the kind of you know very abstract level that would be impaired if

1:29:54 you have problems with forming you know abstract state spaces about your perceptual experience right but I still think like if we maybe reduce it down to kind of more task based measures like just saying yes or no on a did I see something or not on a second visit stuff then my hope is and you know we just need to do the empirical work here is that there will be features those data that will tell us that high real estate is knocked out right right so it's not that like the reports the reports are

1:30:30 critical but it's just that the kind of more kind of narrative level my might not be very helpful right now yeah I see what you're saying I guess what I was saying was that if what you're trying to do is test empirically the claim that consciousness depends on the prefrontal cortex or is it that the prefrontal cortex is involved in an important way or something like that then it doesn't look like you're gonna be able to use reports in this paradigm where you're looking

1:31:03 its absence but normally we do use that when people say I don't see it that's you know good good evidence that they didn't see it and blah blah blah and yet I mean I think I just like I guess you know I guess I guess what I'm saying is to take our experiment on the you know on the front Cipolla lesion patients yeah we could have asked in detail you know after the experiment did you notice anything strange they're like you know and they you know everyone

1:31:36 might have said the same thing we didn't do that but everyone might have said the same thing but because we had these child by child ratings we could then fit our models and find a pretty drastic drop in metacognitive sensitivity so I guess that's what I'm getting at that like it might not seem just interacting with each other that there's major differences but when we actually quantify things like criterion sensitivity metacognitive awareness all those kind of things then you might see fairly big differences but that's not

1:32:10 typically done yeah so so that so there's there's an empirically answerable question here but the question is whether that's the same as the question about consciousness so is there a conscious experience impaired in the way that you would predict given that you know important areas are not working of generating the Jim representations or is it just you know they have the normal conscious

1:32:41 experience and just they're not very good at paying attention to it or judge or monitoring others yeah and so how could you test that if if what we're using the as alright I mean we'll take these I'm believe like neglect and say yeah neglect is another case yeah that is really interesting case because you asked me at patient's anything wrong they're like no it's cool I can see the whole visual field but then you're going to do a plot face is half the clock so that's really then you can make a really strong inference that they're you know experiencing a you know they don't

1:33:17 experience about half the visual field but you know as we know there's been all these debates about maybe kind of first order processing of that is intact and so so I agree with you there comes a point at which you've got to kind of put your hands up and say well you know if you want to claim that in the black first or the phenomenology is intact and the neglected hemifield them you know like I can't really get much better on that but like you can say there's a difference between you know someone claiming that everything is okay and

1:33:53 when you actually probe there are constant experience using a series of finer grain measures like tasks like you know tests and so on then it looks oh yeah I completely agree with everything you're saying it's just that I don't want to belabor this by me already am so who knows I mean it just seems like if there if there's ever gonna be any kind of progress as a field then we're gonna have to kind of just and with this some we're talking more earlier but we're

1:34:24 gonna have to at some point as a field decide what counts as an indication when there is raising consciousness and a subject that we're looking at and you know there's a lot of debate about whether it should be subjective reports or some objective measures and you know some people are saying you know and in binocular rivalry you can look at high pupil dilation and that's an indication and then other people say no no you can have that unconsciously and so no one knows what counts as evidence for or against a theory of consciousness at this point because no one agrees what

1:34:57 counts as an indicator that there is consciousness around and it just seems and so one side saying look you know you hire people if this you're saying this and so if that part of the brain is missing we should expect large deficits and we say no you're thinking in terms of a first-order theory and we'd only expect that if the you know there was a tight correlation between the conscious experience and the first-order States driving the behavior and then well right so what do you do after that I mean it just it just starts to seem like until

1:35:31 all sides can kind of agree about what we're measuring and when it's there it's gonna be a state of chaos in the Wild West is gonna continue to ensue yeah yeah no I thought yeah and I think that you know these are all reasonable critiques against like yeah the current state of the field but I think you know in a sense we've come a long way as well if you think that like we now do have you know things like the development of

1:36:05 solid tools for quantifying you separating out sensitivity from where you set your criterion using single section theory and metacognitive sensitivity yeah so in a sense like it it'd be a bit like the field of you know memory research saying well you know we have all these forgetting curves and stuff but still we want to know what's happening on that particular trial right you know it feels like consciousness

1:36:39 research is not happy with having metrics that maybe a derived from 50 models and things we always still want to somehow go back to the experience on kind of individual trials I think that will or you know when we question people about individual instances of their experience and I think that we might have to remain unsatisfied about that we might as a field and divorcing from the folk psychology on that point yeah that's very I think I think maybe maybe

1:37:15 the pressure is just much stronger in the field of consciousness because it feels like you know no we want to get to the what this is like on this one particular child I mean you think we've already done there like with with attention and memory like divorce ourselves from the folk psychology I mean or yes I think we did I think we have to a large extent I think often when you go to you know big neuroscience conferences and you sit in talks and people talk about mechanisms for memory and attention they might have an opening slide where they say we you'll know what attention is like once they go beyond that they're off right

1:37:51 and they're dealing with measures and mechanism right and I think maybe that's just you know it feels like consciousness tugging at the reins of being able to do that but maybe haven't quite got there right yeah that's that I mean that's a really interesting that's an interesting pony I mean okay so I want there's a million things I want but I won't try to stay focused I mean I have a bad trouble doing that but I wanted to ask about this it's good non my a lot of just realizing that my my headphones are only about should be this is not a signal that we need to finish I just need to unbelief myself

1:38:27 just see if that works I wanted to ask about consciousness and non-biological systems which i think is what Ned was pressing you about that you alluded to earlier so do you think that it's a sophisticated enough implementation of your model is going to be conscious yeah sure answer yes and with a major caveat that so who's interesting say maybe I'll just say what

1:39:02 Ned was asking me so I gave a talk on this high roller state space model SSC and at the end of the tone and Knobloch said you know you haven't said anything about why this model you know whether this model would be conscious in the first place what is it about the model that makes it conscious and if you're not going to say that by that then we should just assume that the kind of little MATLAB simulation you have in your computer you know absolutely that's not the case but then it then then is

1:39:38 responsible as well it seems just really easy to make it conscious because this awareness state is so simple right I think that is a really important point that it seems that if you have a very kind of abstract state that because it is so abstract and you know in in the in the model that I wrote down the paper is just binary it's just kind of stuff there or not so if you just think about that state itself then yes it's super simple but the the fact that to get that's to do inference

1:40:14 to that level in a very asymmetric Network is actually really hard so one lesson we've learned from these kind of probabilistic generative model is they're doing inference so kind of asking what each level in the network is telling you is really hard when the state space is big which is presumably the case of a regular perception right so the model I have in the paper it just has the ability to represent things that are tilted left or right it doesn't have the ability to represent all the possible

1:40:47 things we could see so seem to have a really big state space you can't be exact inference anymore because doing that means you have to kind of do what's called marginalize over all possible values of all the states you couldn't see so an alternative is doing something approximating doing approximation to gauging influence and so there's various proposals about this and it's you know it's a fast-moving field but you know two of the dominant proposals are some

1:41:21 kind of form of sampling like Markov chain Monte Carlo where you draw a samples from the distributions and that you know once you draw enough samples then you kind of get inference for free falling out of that sampling algorithm another way of doing it is something called variational bayes where you make assumptions simplifying assumptions about how the variables interact so there's lots of people working on this but I you know and I don't have a proof of this because I haven't like explored more complicated versions of this

1:41:54 particular network but I would bet that even with approximate inference the computations required for doing inference in this room would be super heavy and it would be computationally costly to evolve this highroad estate so I think that you know a we're not going to be able to write down a or simulate a model with these kind of realistic perceptual spaces that does does this kind of higher-order inference anytime soon and

1:42:30 even if we could it be super computationally costly said we would kind of need a good reason for doing it and so that's also where I think you know an interesting evolutionary story might come into play right but but why does any of that stuff matter to answering Ned's question I wonder well there's lots of answer in the second point which is the misperception that this is somehow easy to do just because it's so simple but why isn't it why isn't the answer that you're giving that it's hard to do in an asymmetric state space yeah but that's not what's

1:43:02 important to the model at the top is just these two states that's what's important how they're arrived at is what's interesting and new in your model but just getting to the top abstract state doesn't require having all that a sum a symmetry in place it does because in these abstracts away from the the lower-level state to but if you just have a really simple isn't a chance to reverse the to the important part of the important aspect is then it's a factorize space so you know it has the ability to report the presence or

1:43:36 absence of anything rather than just having like you could imagine building a simple Network that kind of can do presence of red things absence of red things presence of blue things absence of blue things right but it can't abstract a general quantity of general property which is presents or things being there right so it does need this you know the various ways of setting up but it does at least need this ability to factorize the space okay so in the simple system that you mentioned with it's just the presence of red absence of

1:44:09 red yeah when you set the system up so it will it detects that then there's a monitoring of that and then if it detects presence of red it's connected to this switch that says aware it's ever something and if it says non presence of red is connected to this which says awareness of absence of red so yeah same not basically the way that the current model is set which is incredibly simple so I think what you need basically basically my kind of rich answer to this question is and that in order to be role is to characterize human awareness

1:44:47 reports you need a rich perceptual state space and then you need the ability to factorize out a representation of presence and absence from that rich state space right then debate about like when you make the state space simple and simple and simpler but is there a point at which it doesn't qualify anymore for these kind of awareness reports and I kind of dodged that question but I think that's what I mean that's the way I interpret is that's the question is an awareness report right the question had

1:45:23 two parts one is just a straightforward question about conversational cost right and the other thing is yeah there's the more the bigger question about like what makes it aware and yeah I I will put my hands up that I dodge the question I see and I'm gonna dodge it again yeah I think that isn't you know the goal of the model is to characterize human awareness reports and you know now we have a model in place that can do that I think and what we're trying to do you know the next year or two is to try and fit this model to data and I you know so there that's really the goal you when it

1:45:58 but you know the questions about how rich does the state's base need to be and how does it need to be organized to actually connect to a kind of stronger version of the model which is saying this is really a true model of awareness then yeah I wouldn't have a good answer for that at the moment right okay interesting I mean cuz generally the higher order type answer to those sorts of questions is that you know consciousness is associated with a kind of awareness of your own mental functioning and this is how that gets

1:46:32 implemented but I didn't I noticed that noticeably you were not saying any of that sort of stuff and response to the yeah well I think no yeah that's true I know and I think that like one thing I've been thinking about more recently in relatively half-baked sway is you know if it is computationally costly which presumably is if it is set up in this way then it must have some kind of function yeah what's the point of it it was the point and I think I've been quite influenced by Michael Graziano's

1:47:06 thinking on this yeah yeah so I think you know the highroad a state-space idea and the attention scheme and idea share a lot in common you know they're both ideas that is kind of you need a model of the perceptual system but they differ in the graziano the model is of attention whereas here I would view attention as kind of providing a supporting role you know it provides a useful input for this however state space but it's not you know what's doing

1:47:39 the heavy lifting right but but you I think what you're where you like to is like kind of the the explanation of the what's sometimes what Chalmers called the meta problem so why do we think consciousness hard well because there's a model that's kind of describing you know what's going on in a way that's inconsistent with the way we think is really in there's only right right right right exactly exactly so no I am definitely trying to those kind of ways to think about it and I think that you know one interesting proposal that Graziano makes is that in some way

1:48:13 there's a tight connection between conscious awareness of ourselves and social cognition right and I think that you know this the the approach that were working on could have a similar flavor although we haven't yet explored that so one interesting function could be to kind of keep track of perceptual access in general so do things in my environment perceptually do you like it and maybe you know and there's all those cool experiments on chimpanzees where

1:48:46 they kind of know that the trainer can see the food and all that kind of stuff so I think do it it's clearly not the same thing because it's not as if suddenly I have first-person access to what you're saying right the computations that are involved may be quite similar you know it requires me to have an abstract state which tells me something about whether you can see something on an array that has nothing to do with the content of what you're saying it's just that you know tag which tells me do I need to be worried about the fat rich you can see

1:49:17 the food that I can see right you don't you can have it so that to me seems you know to be and then you know there's a whole interesting line of thought you can take on experiments you could do to try and test that by you know integrating some ideas from Tesla theory of mind and social cognition and things but it feels like that that to me could be a really interesting intimacy I mean it's it's a voice I wish that Graziani himself was more open to this but he clearly that that type of you is that

1:49:51 kind of higher-order view yes a description of what attention is doing at a higher level that's what consciousness is he doesn't want it to be like David Rosenthal's version of the theory so it goes around saying it's not but that's I think that's I mean that's part of partially why we wrote the paper with Hawkman and Jolla do which is you know coming awfully someday was just so people sort of know like yeah that's a version of high order theory you're saying the same stuff just implementing in a different way yes I found I found that a really useful page by the way so yeah thank you thanks for doing that yeah yeah thanks it was a lot of fun but also a lot of

1:50:27 work so so I guess what I'm what I'm trying to get so the what 11 let me try to back up and see if I understand what you're saying in answer to this question why isn't the simulation on your laptop cuz your response is what that is not complicated enough there's not enough inference going on there the state faces is I mean so I'm there's not exactly sure what what your response was

1:51:01 I mean because what I hear you saying is yeah it's not like human experience we have a list very complicated asymmetric safe space this one's less complicated but that doesn't seem to really address the question about whether it is conscious I mean in the way that you might say that the logic gates of the IIT you know they might the bullet on that one says yeah no no but you you're not biting imitational I'm wondering why why maybe I missed it and I'm just being deaf I didn't say well I think I am I know you want to dodge it but I so I think there's a few answers I think

1:51:43 basically I want to say that it's just too early in the current characterization that you know this simulation is not conscious because the simulation is wrong it's not doesn't have all the bits that we need in there to to actually do this higher-order inference and there are various things that I you know mentioned in passing in the paper that are I think it going to be critical so the thought experiment on the close in the eyes you know that that is aimed at saying we need more than

1:52:16 just kind of bomb a pimper to do it we need a way of the system also knowing something about the stay of attention so this can be modeled as the precision of the input and at the moment that in the model is just kind of floating there without any it's not connected up to anything and so I'm you know Carl Kristen has these very powerful proposals about how we not only model

1:52:49 content but we also model precision you know and that is really what attention is it's like do I expect and good signal over there and if I do then I end up paying attention to it so all these things like how it connects to action and expectations about attention and precision and then if you've got all those things in and you've got very rich state space and you're doing this marginalization and creating this abstract property and then yeah okay fine you could do all that pretty much like the human brain so I

1:53:27 mean yeah and so I mean that's there's nothing in principle that you would object to to an AI being able to do that no I think I think it would have to be quite and yeah I think it would it couldn't be a kind of brain in the back I I it would need to be quite active and it would need to be doing kind of this you active search in the world and to have me approaching human consciousness I mean I always wonder about that brain in the bat stuff

1:53:59 because nothing at too distracted but in one sense I understand why you're talking about that if all the if the brain is kind of fundamentally passive and just fed a projector Joe it's in a virtual environment or body then it should be the same right yeah bye-bye brain if I didn't nestle mean yeah something fundamental about yeah you could simulate all those things like it's in the matrix that's fine by me yeah it's more about it needs to act in its in its world right exactly yeah right it can't be passively fed a movie of something yeah okay because I mean I

1:54:36 I always secretly hold out hope that that that that's gonna be wrong and that a eyes won't be conscious but you know so I'm kind of persuaded by some of Ned's intuition Marv going in in this area but I'm not really I'm not as confident as I used to be but yeah have some hope that there's something distinctly biological which is I say look I see I see yeah well maybe generative models rich generous and most could you know go well I mean we know that we can implement

1:55:09 rich generative models in software as well right but I think you know if you want to contrast it with currents I mean a lot of the successes in machine learning are fundamentally feed-forward neural nets doing like massive backprop and may be coupled to some value learning system right like really many you know that is very distinct to the kind of Genesis model I do right exactly yeah I know there Dino they said it's a different it's a different computational architecture definitely for sure right I

1:55:43 haven't really had a chance to ask you about any of the practical results that year is that your lab is uncovered which is so interesting but I mean I want I wonder so how important just as a general way of swapping over all that stuff how important do you think metacognition is for normal functioning in society is it is given that you find such a variance and in performance yeah is this this is a problem isn't it I mean yeah well no I mean I think well I think as a general concepts I think is important and I think there have been

1:56:19 interesting efforts to kind of put method cognition in to things like education that main mindfulness and stuff like that not so much mindfulness more like actual using the word metacognition and saying like you know encouraging kids to check their thinking on a particular topic and kind of you know filling in little questionnaire saying what do I know what do I don't know what don't I know and so on so I think those are interesting efforts but I think part of the problem is that you know they're kind of quite the the more

1:56:55 the more quantitative approaches that have been developed for studying math acog mission in the lab haven't yet been translated into these more applied contexts so we don't really know whether they work at all in terms of like improving people's metacognition and so on but that is something we're trying to work on it as well we're completely different line of work on metacognitive training the extent to which this training is domain-general so if I train you your medical mission on one SATs do you see generalized benefits and we have some encouraging encouraging results on that and so I

1:57:28 think that you know to the extent which is a relatively domain-general thing I think there are exciting opportunities to potentially think well if I could boost your awareness and performance more generally maybe that would have benefits in different areas of life but wait there's some real suggestive results like this Radek political radical in this result which got a lot of attention from here now that people on both sides tend to have impaired or meta bility to be sensitive to how well they're doing exactly

1:58:01 exactly so you know in a sense it's like one hypothesis where the series could boosting domain general method cognition help you in being more reflective in other aspects of daily life like decision making like learning and so on and so yeah I think we're just starting that line of work I'm pretty optimistic that that could be the case because you know in a sense like we've been quite you know there's been a focus on

1:58:35 improving just general ability you know this is all the brain trigger boost your IQ and education is all about like academic success and knowing is you know knowing all the capitals the world and all that kind of thing right yeah I feel like there's maybe a opportunity to complement that I'm not saying well that's not important but to complement it right systematic exploration of whether training up self-awareness in a similar way as formal education trains of intelligence would also be helpful

1:59:08 exactly no I think that's really important especially when you start thinking about you know the criminal justice system and I mean I think it's only a matter of time till we have the metacognition defense and in the courtroom not good until I was doing something bad metacognition I didn't know I mean definitely there's interesting stuff from trying to use mastic emission tests to qualify a witness's eye an eye witness statement because I think that is a whole field of research on confidence and you know it's

1:59:43 amazing I see like when think about like the threshold for kind of new evidence being introduced like new techniques like DNA guess a into courtrooms because it's so reliable yet we're still with eyewitness confidence which is just like completely unreliable like change buying this test then you should not be I mean there's no probably an argument to be said I mean you know not that one it create like a kind of tear into tears societies to you know

2:00:16 I'd be very happy if someone told me hang on Steve we've tested your memory metacognition for this particular you know set of faces and you're terrible so we'd rather not you not you know give this confident statement in court I'd be like yeah fair enough to be honest like you've told me you need that I'm not very good at you know evaluating confidence in my memory so I shouldn't be what I want it's like yeah don't see a problem with that I think there's a resistance to that kind of classification because it sounds like we're kind of saying well let's divide people into the good ones and the bad ones but but it's just what we're taking seriously the idea that this is an

2:00:50 ability and you have any other ability there's variance in how good you can do it so we don't have to treat everyone the same in this respect yeah but but they sit but the criminal justice system does it just assumes eyewitness testimony yeah I need look into this again I know there were a couple of papers on trying and I think there was a couple of jurisdictions New Jersey apparently is very good on you know integrating the latest science on eyewitness memory and that was the last stuff I read about it it was been a while interesting yeah okay well I mean I've kept you all ready for very very much time so I wonder if is there

2:01:24 anything that hasn't come up with that you were hoping would come up or anything they haven't talked about yet he's wanted to bring anything that has come up I was happy yes no I think another thing this has been really fun yeah it's been great talking with you I think you're doing great work I really I mean it's not only doing great empirical work but it's just always really nice to have someone who cares as much about doing good science having good models but also thinking hard about the kind of

2:01:57 big theoretical questions that some people call philosophical but really I just two big theoretical questions so it's nice it's rare that you find someone who really cares about all of those things I think you really do those I think it's good that there are people like you out there well thank you that's very kind and Aaron Mexican ice for saying this up and it was really fun and yeah yeah all right well stay on board I'll say bye after