Consciousness Live! with Alex Kiefer & Pete Mandik (S3 E13)

0:00 joining me today two guests special then none can be conceived alex kieffer and pete mandick um alex pete uh for those who don't know you can say a little bit about who you are and what you do i guess we'll start with alex alphabetical sure uh so i'm alex alex kiefer um i i guess i study philosophy and i make music sometimes um i'm currently uh a lecturer and researcher

0:33 at monash university and uh at the moment i am in the united states because of that which shall not be named [Laughter] so i don't know how clear that was without context but i'm sure it'll become clear yeah okay so we're not allowed to name cthulhu oh i guess we can now yeah i'm uh pete mandick i uh do philosophy of mind and also

1:08 um am technically a professional cartoonist uh which is fun and uh also um the uh the head of a large media empire so i you know i think everyone knows who i am at this point because what media empire are we talking about here that was a joke okay cool so you're talking about the cartoons for this for the daily news are they they paying you for that yeah yeah it's paying gig

1:42 wow that's amazing yeah i love those i love those things by the way the mine chunks they're awesome thank you thank you and you got merch too right i guess uh you that's a mask that you did recently yeah i just kind of did that as a bit of a joke uh but also because my wife wanted one and then they started selling um i'm like oh that's cool it's gratifying to know that there's people out there i don't know who they are but they you know they're they like it cool sharing my art yeah no that's

2:15 awesome yeah i ordered a mask a while back and it still hasn't got here so you know you make you write a philosophy paper and it takes years and years to write and then like three people read it but you make like a little cartoon or something like that takes you about three or four hours and then thousands and thousands and thousands and thousands of people sometimes it's easy to prefer the one over the other yeah yeah that's an interesting thing about social media like i could post a paper i've been working on for 10 years

2:48 and it gets like five likes or whatever and then you know a famous person posted supposed like i farted and it's like 40 000 retweets um but you know that's just that's the the waters we've got to yeah navigate talking about that is this based on a real life event like you're recently yeah that's like a condensation of tons of like real life events it's not specifically that one no i mean i don't know maybe a middle ground like you gotta you gotta write a paper called i farted and get it right exactly logical review yeah yeah um more interesting titles

3:26 um yeah okay so so alex you said you work on consciousness is that correct oh man i hope i i didn't mean to say that i do a little say that you do a little bit i don't know i do a little bit i it was more it's i haven't done much work in consciousness specifically lately but it's mainly because like i'm interested in understanding what the hell mental representation is and of course you can have a non-reference not non-representationalist view of consciousness you're some kind of weirdo but

3:58 um you know i basically i'm trying i feel like there's a there's a things under the hood that haven't been figured out yet and a lot of people are doing consciousness so for totally pragmatic reasons i don't think much about it that's good yeah but i'm happy to talk about it but you do so you're in more interested in representation mental representation specifically well yeah that's one that's one thing i'm interested in for sure can you tell us about one because i mean i want to ask you about i know that you're interested in connectionism

4:32 uh you wrote an awesome dissertation on that actually that i happen to know about but also you've written a couple things that have come out so you try to connect it to like the free energy principle and a lot of this stuff that's going on over there yeah which i have to admit that i'm not like super up on a lot of that predictive coding stuff and i mean i know about it but i'm not really hip to the cutting edge of it so can can you give us like an overview of that approach for those who may not be initiated yeah sure so um i guess i'll start with predictive coding so i mean i think

5:04 i think a lot of people who might be listening to this probably know the basics and i'm not going to do any better job it seems like every paper in this area introduces the framework again again so i feel like i have a lot to compete with here but essentially it's just the view that the top down connections in the brain are implement a generative model which is a statistical model that tries to predict and sort of generate um states of the lower order representations in the model and also the sensory input like the peripheral sensory input

5:37 channels um and then the the feed forward connections in the brain implement feed they provide feedback to the model that corrects it over short time scales like you know inference being sort of uh inferring the causes that would have generated the input within the model and then over longer time scales um adjusting synaptic weights to come up with a better statistical model and then even things like this starts to shade into the free energy principle even things like the structure of your brain counts as like a structural prior that constrains the kinds of

6:10 interpretations um that you will favor um so can you say a little bit more about that the shape of the brain constrains that oh well i don't know exactly so an example is like convolutional neural networks so like yeah okay we have small um neurons that have small receptive fields that are sort of sensitive to things like edge orientation right at the lower levels um repeating a shape like that across the brain is sort of like a prior sorry across the across the visual field it's like a prior because it's

6:42 it's making the assumption that similar shapes will show up uh at various places in the visual field so by replicating a set of weights across the um the span of the image like that or whatever input you're processing it's like um building a prior into the structure of the network also just the basic organization of the brain into cortical processing layers um you know that could be viewed as a as a prior um so i don't know prior tends to be like a magic term it seems functions that way but but it's it's it's got a pretty precise meaning right in terms of

7:15 inference and such so uh what are i have a question alex yeah um [Laughter] like you didn't raise your hand richard i have an answer i can't say it's i can't vouch for its quality but i'm not i'm not hip to this stuff um or in other words like i don't understand it very much you guys are more hip than you're getting yourself i blame the math i see oh there's a bunch of math that is a pain in the ass yeah i used to know math um but i learned math just for this i literally um several years ago was

7:50 helping my oldest daughter with some math homework that involved long division and i had to google how to do long division because i have literally completely forgot right um anyway so my my question is uh with respect to things like predictive coding what what are the like uh architectural um minimum requirements for this to apply like if you had just a you know old-school

8:23 feed forward three-layer neural network with that can you apply predictive coding to that or does do you have to have at a minimum some kind of recurrence like i'm just wondering from in terms of like because you know for me i was like the connectionism was the 90s and i haven't haven't kept up since then man and stuff got really interesting like right well i was gonna say after that but really during like right in the middle of the 90s is when these cool really cool models started coming out but i don't think many people i don't think enough people paid

8:56 attention to them at the time but like i think it's what you said i mean it's a kind of a hard question for me to answer because i tend to you can sort of i can tend to view any dichotomy as you can probably find a way around it by taking something to be a degenerate case or a trivial case so i bet you could tell a story according to which predictive feed forward counts as predictive coding but that's kind of silly so really i think it's what you said i think you need some kind of recurrence because you you need at each at any given layer that's part of this predictive coding model you need top-down and bottom-up signals to sort

9:28 of conspire to determine the state of the layer i think in terms of just like a neural representation uh and then philosophers will argue about the cognitive significance of that and i try to argue in this new paper that you should assume as sort of trivial mapping as possible from vehicles to contents um based on this idea of structural representation but anyway that's going on a tangent um and what is uh literally what is the prediction is the is the prediction like a state at some higher

10:00 layer or is it i something i mean i think it's relational yeah i think of it i think it was just the top down signal uh that's arriving at a given layer so like if you have layer three um the prediction would be the signal sent from layer three layer two let's say okay if layer zero is input um and so that's that's the prediction in like vehicular terms kind of right it's the signal that's sent down um and i think that has a semantic content

10:33 and that you can assign and however you know we could we could argue about that but like that's i'm familiar with some stuff that's kind of like this through the work of rick grush which was um you know part of the 90s uh neural philosophy thing um and while he was he was interested in um connectionist models he was he seemed to me to be more directly influenced by control theory right um which wasn't you know didn't necessarily require that you're using

11:06 distributed you know parallel distributed processing total in in in the in the forward modeling uh that he was drawing on there there was like see like pretty specific content to the claims that this this one thing is like the plants you've got the controller you've got a forward model and an inverse model and some of these things are like elaborated in terms of control systems like for example you're trying to do some thermal regulation in a in an engine um or you're

11:39 you've got a controller you've got a controller that is trying to like regulate um a chemical processing plant excuse me and there's these huge these huge delays in the in the in the time of the signal right so imagine trying to like control from earth uh a drone that's on the moon and you've got what is it eight seconds between the earth and the moon light there's an eight second delay or something like that yeah something like that so imagine trying to

12:11 like you know do surgery on somebody remotely but there's an eight-second delay and there's all these specific sorts of like um uh bad effects that that can settle in you you know you you over correct in one direction but by the time the overcorrection gets like it's gone too far and then when you get the feedback so there's interesting like mathematics that that you could do to kind of smooth all that out and allow you to control something remotely and you've got these intermediaries that

12:45 are doing predictions correcting for those predictions and it seems to me that you know some of this some of this stuff sounds like that that can control theoretic stuff where you have um in those toy examples there's like very clearly a um a set of signals that you need to to send to a robotic arm there's some feedback coming from the robotic arm or maybe there's also a camera and because there's a big delay between the

13:17 the command and then any feedback you would get you need some kind of representation in a very robust sense of representation you need a model of what's going on and you need this this system that maybe is on the moon it isn't in the drone but it's like closer to earth than where the drone is and it is doing some modeling and predicting and updating based on feedback in a way that overcomes some of these um problems that you would get into like these tremors in the arm of like all these over corrections and then over

13:49 corrections for the overcorrections because the feedback didn't get back to earth enough um so anyway the this is kind of a rambling preface to a question um how literal i have a question for you too when they say uh when they say that there's a representation or there's a a prediction how literally true is that well so engineering cases it seems that it's it is quite literal that you've got this one system that is creating very much the same sort of signal

14:23 that you would get from the sensor exactly it is making do with it because it can't wait around for the for the sensor signal right right so i mean i don't tend to think so much about the time uh delays as the reason for doing it i think there are tons of motivations that's one of them sure but i was going to ask you what you mean by robust sense of representation because i i think that like you just said um one key angle on it is that the top down signal is producing a state that's very much like what the sensor would have

14:55 received and in hierarchical predictive coding that just happens at each layer of processing well there's some people i i'm not in a position to name names um but there's some people that i encounter in cognitive science or neuroscience or computer science they'll use representation for like any anything that happens happens to be happening in between the input and output of something that is worth calling a computation right um so you've got some kind of mechanism and you and you you're able to use it as uh able to do

15:29 arithmetic um and the the inputs and the outputs are things that are pretty easily interpretable as representations of numbers but who knows what the heck is going on in the middle um and there's something going on there and they'll just call it a representation without caring like what it would be a representation of yeah you know like back when i mentioned long division um when a person is doing long division they are literally manipulating symbols

16:05 and the symbols like have a there's a semantics that you could apply to them you could see that like okay this numeral is representing tens numerals in this column represent tens multiples of tens numerals in this column represents uh hundreds and so the things that are like the causal intermediaries between the the the inputs and the outputs themselves have a semantics but sometimes it seems like people will label things representations simply because they are causal

16:38 intermediaries between inputs and outputs the inputs and outputs they have a semantics but they don't care what the semantics might possibly be for the for the intermediaries and that seems to me to be not be a representation in the robust sense of representation right they don't care like what it's representing they have no answer to that question they don't care that they don't have an answer but in some of these other sorts of examples like the one i was trying to remember in terms of like the control plant um there is there's a particular signal that you could see is like a command

17:14 to an arm and you could see that like the different magnitudes in the command signal are are isomorphic and causally coupled with um joint angles in a robotic arm and it seems like pretty fair to say that that's a representation and you could see what it's a representation of it might not satisfy jon searle that it's a representation but there's a lot of people that would call it a representation it doesn't seem like a howler to do that but with some of these like you know massively distributed uh systems like artificial

17:48 neural networks it's you know famously um at least in you know the 90s type examples that that we had from like smolensky and and some other the big characters that it's really hard to see what the heck the the the representations would be so you take for example um like the sorts of face recognition feed forward networks that pat and paul churchill like to talk about um the there's nothing going on in like the

18:21 states at the middle level that you that you could say like well this represents an edge or this represents a face so this represents that it's a female face it's just like i want to go with faces the output is like yes or no or male or female and there's nothing there's no semantics you could give to the stuff that's in the middle and it seems kind of like cheating to say like it's distributed representations that doesn't really is that cheating i mean alex is going to say that it's the property of the whole system somehow that represents it well that's that's one thing i would say i mean

18:54 there's a number of angles to take here one is like i mean you can for example you can look at what the preferred stimulus of these artificial neurons is that's one way that's to me that's not sufficient for representation in an interesting sense um because of my weird views on well not weird but maybe uh specific views on um what representation is which is like richard saying it's a property of the whole system because i think it's a functional role kind of property um but like for example if you wanted to convince someone hey this is a

19:26 representation i mean especially at the lower at the lower layers of the network where um these processing units you know which i take it they each stands for kind of like a population of neurons or each represents that in terms of modeling um you know you can look at the the image that would most excite that neuron and it will look like an edge or what i thought was exciting about this stuff in the beginning was if you train uh i don't know you know five-layer network on youtube videos of cats then you'll get

19:58 neurons in the higher layers whose preferred stimuli look like cats right or things like cat faces so um i would argue that the projective fields are just as important as the receptive fields here because you've got this bi-directional thing going on um but so i didn't exactly just give you like a necessary and sufficient condition for being a representation but like i think that's the same kind of methodology that's used that has been used in neuroscience at least to determine what the representation uh what the content of a of a given population or you know population

20:31 code is or something um i mean here's a kind of concern that i have about this stuff and and maybe this can help get it across so if you were to um if you were to sit down and write an algorithm in you know using you'd like using c plus plus or python or something like that you're writing you're writing like an old school ai program to do some kind of like visual recognition task you want to you want to make a thing that could classify two-dimensional uh objects as whether

21:04 they're triangles or um or squares and um you could you could imagine sitting down and writing some code in which there's going to be some early step in the whole process where there's some edge detection going on and uh or and then after you've done the edge detection or you've you've identified things like vertices right or you think of like uh just to make a slightly different example um old-school um um vision computer vision

21:40 where you've got these things trying to extract a 3d form and do some object recognition and it's looking for things like t junctions uh versus arrow uh junctions there's a certain sorts of ways in which the vertices of a three-dimensional solid will show up in a two-dimensional image if there's t-junctions you can infer from that that like the cube is in front of the edge of a table like stuff like that there's going to be points in the code where you're literally have things coding for or representing

22:13 edges and they're using those representations to drive something that is edge detection it is possible um that there could be a machine that accomplishes the same output namely correctly identifying the presence of a cube versus a pyramid without actually running that algorithm it does it in some other way um i guess it seems possible to me and i've done a little bit of programming um and it also seems possible that

22:48 it would uh it would it would have things that um in it that kind of look like what neuroscientists call edge detectors even though they aren't in really edge detectors there might be things in there um that are selectively responsible for the the little neurons that are happiest when when they see an edge but but that's not what they're actually doing they're not doing what the edge detector program did i'm partially inspired by

23:21 this old thing that uh patchlin uh pat churchlin co-authored in like the 80s or 90s where they had um they trained up these networks to do a shape from shading they were able to extract three-dimensional shape from just like shading gradients um and making and drawing conclusions about like well was this convex or concave this three-dimensional surface and it no and then after these networks were all trained up they could show that there were individual

23:53 nodes in the network that were selectively uh responsive to edges yeah but it's an edge detector they argued that it wasn't an edge detector um okay i'm curious about that argument yeah what was the argument uh well the it had nothing to do with detecting edges it was all like all the now i'm going on a 20 or 30 year old memory so i might be [ __ ] this up but um i don't the the training stimulus didn't

24:24 have any edges in it it wasn't like the the imaginary computer program that i was describing it wasn't like looking for the edges and trying to extract from the edges whether it was a cube versus a pyramid it was just looking at various kinds of gradients and making um drawing conclusions about whether it was convex or concave or something like that and as a weird um side effect you got things that behaved the way the so-called edge detectors behaved

24:57 that if you did like a single cell recording from it and then you showed it at edge you would get a maximal response just because your maximal response doesn't yeah so let me just detect let me just say like so in connection you know in like multi-layered convolutional vision networks or whatever you'll get things that are like that that you could describe as edge detectors because of how they behave just to in response to stimuli but also they feed into the next processing stages in the right way so like you know as a functional role guy i would say well it's because those edge detectors then get input into a shape detector which gets

25:30 input into a 3d object detector in the right way that that the representation of edges but just to just to drive on the point about functional role like i think your example of calculating with numbers is perfect because yes they have a semantics but like what is our grasp on the semantics of those symbols consistent other than the fact that we can add and subtract and divide and such in the right way right so like i mean yeah there are maybe platonic entities or whatever you think numbers are but like the only reason we think these symbols refer the numbers is because we can do all this manipulation of them and so i would say in the same way

26:04 the states the intermediate processing states between input and output and human being are representations with semantic content for the same kind as a reason um i do think you mentioned a lot of people might talk about representations without being specific about their content and such and i'm definitely guilty of that because i think oftentimes you can get to give a good story about the specific content of representation but like um there's plenty of cases in which it's just cool to show that there's reason to think that representations would arise in these sorts of systems

26:36 for very like broad general reasons um i mean i think that's partly what's going on in the case of of deep learning right people accuse these networks of being black boxes which i mean they are to some extent currently given our epistemics right but um but uh i don't know i forget what the hell i was gonna say next so i'll stop talking ali can i get you to say more generally what you mean by the word representation since we're throwing it around oh man well what do you mean by it no i mean yeah i mean i i feel like it's a pretty

27:13 i i like to talk about or think about a folk sense of representation so um uh to be kind of an idiot about it it's representation or some kind of presentation right okay how how okay so i happen to like etymology more than an analytic style philosopher should but like why is that terrible um it's it's a it's okay it's a stand-in if you want me to get fancier a little more theoretical it's something that stands in for something else in a really weird sense

27:45 of stand-in though that might be totally proprietary and not very useful to illuminate what the thing is but it's something like a proxy i think um i like alan newell's definition from the some paper in the 80s where he says basically well okay he defines designation as a technical term but um i think he says a state of the system designates some other thing um relative to a process um just in case um something like the process uh

28:19 depends a process that's that involves the state or the internal variable ends up depending on the thing that's that's designated ends up depending on the other thing so that sounds like sort of a co-variation definition i think there are different ways he doesn't specifically say that dependence is you know statistical dependent dependence or causal dependence um but something like that so you asked me okay i think i've i guess i've answered the question sufficiently for now i mean i don't it's i mean yeah i just i'm the only

28:51 reason i brought it up was because it's like discussion you're having with pete right now yeah you know whether this system that pete described which you know from the churches or whatever wherever came from whether it represents edges or not that was one of the questions under under consideration and you were saying yeah well i think there's good reason to think it does represent edges and p was saying we have good reason to think it doesn't represent edges so this may end up to be a verbal uh issue is what i was thinking depending on what you mean by representation because maybe people's using it in a kind of stronger sense than

29:24 you were using it uh so that's what i that's the only reason i brought it up i mean i wasn't trying to trap you in thank you i don't know anything deep or profound about representations but yeah i i do wonder if that is in the background of this exchange i think it might be that's why i was wondering what what pete what you meant by like a more robust sense of representation um like the thing is i guess i've avoided getting involved in debates over this specific topic because i mean i think minds are important and like the

29:57 difference between minds and non-minds is important and there's a temptation to try to like given representationalism as a as a background assumption to try to like make what's special about minds be a thing that's special about mental representation and i'm just not sure i think there's there are many very good reasons to use terms like representation um where you don't have to at the same time be trying to define a special sense of representation in that way but i do think it would be interesting to figure out if there's anything

30:30 more special here than just like so like for example why would why would the states of artificial neurons in a three-way or feed-forward network be different in terms of their representational status from the states of a predictive coding type model um i mean i really i guess my my thinking on this is that the content of represent a representation is something that plays a role in a system of representations that's analogous to the role that the represented thing plays in a system of represented things right so um

31:05 you could have really trivial examples of that i don't think the word representation itself is gonna is gonna distinguish the trivial boring cases from the more interesting ones so i mean because one thing i'm not hearing from from you and maybe on purpose is something about misrepresentation or you know getting things wrong oh yeah accuracy conditions things of that nature uh yeah because some people like i guess like me maybe would want to distinguish what a tree that rings in a tree do i'd say they

31:37 don't really represent the age of the tree they carry information about the age of the tree but they don't represent it because there's there's no real sense in which it can make a mistake same for like a thermometer an old school like um you know mercury thermometer can't misrepresent the temperature yeah so it can't get it wrong is anything like that figuring in in your account of it or was that um yeah sorry i'm done you want to take i do have an answer but go ahead no no i was going to ask um like so i tend not to rely on the

32:09 misrepresentation criteria and so maybe there's nothing to this thought i've never really tried to voice it publicly so i'll try it so my thinking is a lot of philosophers say the possibility of misrepresentation is essential to representation and i think it's not because i think misrepresentation is a type of representation so um i want to have an account of representation first and then say misrepresentation is when you represent the wrong thing but that that presupposes representation so i don't tend to think of

32:41 misrepresentation as uh as a constitutive condition on representation um that said you easily could go that way in predictive coding models because this whole prediction prediction error mechanism is built in so that would be a nice easy answer i could give but i don't think it's the right one okay great and so what about what's your take on this pete are you using the word representation in a more loaded sense or yeah it's a little bit i think a little bit more loaded than um alex um but not as loaded as some other

33:15 people a little bit earlier i said this wouldn't make john searle happy there's some some people they have such a loaded view of what representation needs to be that for example it needs to be phenomenally conscious or needs to at least be in a system that it could become conscious and i just say um [ __ ] that noise no thank you um just you know uh no wait so you see the overreaction that was me being polite

33:50 i'm inclined to think that um representation is kind of a messy topic because it's not a folk term it's like uh it's a theoretical term but it doesn't live in any particular theory there's not like a full-blown theory like for example einstein's theory of uh relativity that's a full-blown theory and there's certain terms that have their meaning because of that theory representation gets kicked around by a whole bunch of people that have a whole bunch of partial theories they kind of get it from the folk and i do think that it's important to keep in

34:24 mind the connection to the folk i think that that story is something like um if we do a little bit of reflection on folk psychology we noticed that even though the folk might not put it this way that there's a distinction between attitude and content you can um you could believe something or you could desire something and there's some sense in which it's the same thing um i i you know i hope that i'm gonna go to the movies first as i remember that i went to the movies um and there are

34:58 in in your kind of basic folk psychology there's certain sorts of like causal or functional relations between these various attitudes um so if you if you you know if you desire not to be bit by dogs and you predict that a dog is gonna bite you if you don't leave the room then you're that's going to cause you caterers parabus to leave the room so i think that's kind of like the the semi-theoretic or pre-theoretic model for for um the kind of representation talk that i think is good representation talk

35:31 and in those cases you are able as a theorist to say well what is it a representation of and i don't think it's always going to be the case that the the representation that you're dealing with is a sensory representation sometimes it will be something worth calling a sensory representation but there are other sorts of things that are much more closely modeled to thinking in terms of like intentions or desires or um or what you know gets called intention in action and in a more theoretical

36:05 frame we might call it a command signal or a motor signal um so i think of in this more theoretical sense of representations representations are are such that they're you could say what the content is of the representation you could distinguish between the content and the vehicle of that content that's kind of like distinguishing between the attitude and the content in the in the folk theoretic case and i wouldn't want to uh at least off the bat privilege the kind of empiricist story that i hear when people say like alex said um

36:41 representation is representation that really biases the whole enterprise toward a model whereby first and foremost the content-bearing states or sensory states and everything else in the cognitive system that has content gets it by courtesy of the sensory states that might be true i think there's some good reasons to think it is true there's other reasons to think it isn't true but i wouldn't want to bake that into the initial characterization of what a representation is supposed to be um and then just just to point out some agreement with alex

37:13 the the general kind of view that he described uh i would i would say i'd agree with it i'd describe it as one of like uh functional isomorphism um that if you've got a system that behaves in certain ways that um he didn't say isomorphic but um i would have something like the structure of that which is in the environment or something like that that that seems about right i would add some some caveats there's um i think at least some some kinds of

37:46 representation that are worth uh calling what um galistol calls nominal representations um so for example the gallisdale's uh example of nominal representations are are the numbers that are used for team members like on a football team um the uh you know so like player number 32 or whatever and player number 24 the 32 is just working as like a pure name

38:18 um the the fact that 32 is a greater number than 24 isn't supposed to represent that this player is better at basketball or taller than the other player so there's no isomorphism there's no isomorphisms there um that right it's just what uh galistal calls pure uh or nominal uh representation um and i think that in plausibly like some things worth calling representation are like that uh and maybe the right semantics to give for those

38:51 representations is something more like an indicator semantics as opposed to something that would involve multiple pieces that are jointly isomorphic to some environmental features um so i'm i i think that you're going to need a lot more heterogeneity you're going to need to allow a lot more heterogeneity in like types of representations and what story you tell about how those types get those contents i also um i don't know whether i agree or not with alex's claim about um teleology or i mean

39:28 or misrepresentation sometimes i'm inclined to think like yeah once you've entered the the representation room you better be committed to saying when when it can misrepresent i sometimes do think that that's the right view but i also find uh i mean if you look at a lot of if you look in the field if you look at scientists using this terminology i don't see them putting much effort into providing for that making sure that that so-called constraint is being honored and and that makes me think that like

40:01 maybe it doesn't matter um well that you know so um but there's so there's a lot there and i don't want to cut pizza sorry i don't want to cut everyone off but so in what sense does the things you're talking about here not apply to the edge detector in this earlier primitive system sounds like what you just said is you agree that it represents edges no i think there's still would be an open question that represents edges so um just because it it's it's identifying um cubes and cubes happen to have edges

40:37 it doesn't matter it does that doesn't necessarily mean that at some point there's a thing who's uh you know and this maybe is pushing me in more in a teleology direction there's nothing in here whose purpose it is to to carry edge content yeah it sounds like what you're saying though is that and what maybe the churchill were saying that someone had to put that in there the purpose but why couldn't out a purpose have evolved like it's a consequence of the way the system started working and whatever that it requires these states whose purpose carry information about edges i mean i think so um i'm trying to think of a

41:11 good example to illustrate the the fact the fact that it carries this uh informational content where information is described in terms of like this causal relation doesn't matter for its performing the function so there there might be um you know there might be a part of your car that is is hottest you know it lights up the the most uh in in response to uranium um but it wasn't built to do that um it's actually built

41:46 its optimal performances when it's not lighting all the way up um and in fact every description of its optimal performance is going to emit uranium you don't want uranium anywhere near the car right so it's just a weird accident that that's what makes this thing maximally light up and now if you wanted a uranium detector this mechanism would would be a good one um but uh but that's not it what it's for um and uh but if they played a certain role in

42:19 the system then you would be happy saying that's what it's for um in order to like if it turns out that the only way you could get the car to run is to put this thing in there that that as a matter of fact would be maximally hot or maximally bright in the presence of uranium i don't think it follows then that you've got your uranium detector or detecting uranium is how it works like if you wanted to know like how does it make the car go it's just irrelevant that it has this particular relationship

42:53 to uranium i mean this is how we get into cement as just semantic issue like you're right now when you say this thing wouldn't be a uranium detector it reminds me of when when um milliken says um you know swamp swamp mans wouldn't have a heart or wouldn't have a brain like that's a way you can use language i just don't deeply see the motivation for it but um i don't want to i don't want to harp on that too much um um let me just volunteer that i don't i

43:26 don't think representation needs to be understood in a normative way at all so like i don't want to go at all in a teleo direction yeah i don't um you um i mean i i i'm very ambivalent about the the the telio thing on the one hand like of course swamp man has a heart i mean look at i mean his pump in the blood and uh it's keeping him alive uh in the way that it keeps me alive and um

44:00 and maybe you could do some like fancy telio story there like the the um you know 10 minutes later he's still alive how is that possible well it's because his heart is pumping the goddamn blood that's how it's possible um and so that allows us to say that that's what the his heart is for but then you know like a armchair metaphysician is going to want to know like well what about at time t 0 when he first lightning bolt strikes a log is his heart for pumping blood then

44:34 maybe the right answer is like who gives a [ __ ] um like i don't i mean just uh just a defense of millikin here you know her idea is sort of i thought is similar to what alex was saying uh not in the content but in the spirit so what she wants to do is provide a theory of representations first and then see what follows from it so if you have a theory which is really useful for like organisms like us and gives you good predictions then who i mean if it has this consequence about swat man what why does that matter he's not you

45:06 can't make these kind of generalizations you know we don't know if he'll be alive 20 minutes from now we don't know if he you know so it's not anyway so i don't really see the problem saying he doesn't have a heart uh nothing really hangs on it it's not important in any real sense if you have a theory which explains other things and is useful for organisms like us which do have a history um and you can explain the functioning of normal hearts i think that's what and when they go and when they malfunction why they're still hearts etc that's what's important so these other

45:39 kinds of you know anyway so i just wanted to i mean that's what i thought was in common with what was alex was saying was that you want the theory first of what a representation is and then you see what it tells you about these weird cases there's a lot i agree with there um and maybe that puts me in large agreement with millikin but um the thing i would want to focus on from what you were just saying richard is you know what is it helping within the theory so in the in the if we're if we've got some cortex and we're trying to figure out like what is it doing with vis-a-vis vision

46:12 um and uh you might ask questions about like well is it representing color or is it just representing luminance is it representing edges or or not um i it might turn out that it like calling it an edge detector isn't helping your theory at all and then right the mere fact so the mere fact that it could be used as an edge detector like if you wanted to detect edges like one way you could do it is hook an electrode up

46:44 to a cat's brain and then point the cat at edges and see would make like you could use it as a as an edge detector if you wanted to but is the cat's brain using it for an edge detector well that's the question explanation for how the cat is able to tell the food from the bowl that at some point there's something in there representing edges were you just about to say so that is do you think that is the right question yeah it has to happen using it yeah like what

47:17 what role does it play and not just what information does it carry and i mean there are questions i get into about what we mean by that too like about carrying information but not just what information does it carry but you know how is the information exploited etc i mean that's not my original point by any means a lot of people no but how is that not normative ooh uh how is it normative i'm talking about how the system in fact functions right so like if it's malfunctioning then you'll be talking about how it ought to function sorry i'm talking about like what the kinds of things the system does so like

47:49 i think you could take a you know instantaneous temporal slice uh of a cat's brain and you know assuming you could infer the dispositional properties from the categorical properties maybe that's a big assumption but i'm assuming that you could figure out the roles that these different uh parts and signals would play um in the broader system and then you could tell what whether this thing functions as an edge detector in a sense of function that just means actually does that not that it has that purpose in any more robust sense right but so in the i mean one of the questions i mean i i

48:22 you know i i feel outnumbered here a little bit because i like teleology i like theological theories i like causal theories of reference i'm against connections and generally so i know i'm on the other side of you guys in a lot of different ways but one of the things that you know teleological theories are are aimed at trying to answer is why is a broken toaster still a toaster that why why is a malformed heart still a heart it doesn't actually do that job or are they so so you don't think a broken toaster is a toaster no i think it's a i think

48:55 it's a broken toaster right but so what in what sense does it qualify as a toaster if it's not that well it's a broken toaster therefore it's a toaster the other day we're listening to wfmu they got this kids show on saturdays and there was a song about toast and in the song there's a mythical uh telling of the tale of the origin of toast and uh so this guy uh had some bread and and uh and a toaster and he was the

49:30 first guy to think of putting bread in the toaster and posting the bread so that's where toast comes from [Laughter] brilliant excellent okay so so if you take a a smashed brain um there are still edge detectors in the smashed brain well or do you not think that i don't know i think there's definitely a brain uh i don't know that they're there's nothing that would now function as an edge detector so i kind of would want to say there isn't any more an edge

50:03 detector wow okay i mean just that's honest intuition that's not like yeah cooking the books no yeah i find that interesting because i i mean i guess i just my yeah i find this really i mean hard to uh well right and why you would think that well because i think that's going to function as an education there currently but there are things in there which the reason why they are there is because they've detected edges right that's why those systems are there they're broken now they're smashed but that what explains their

50:37 prolification through time is that they have this function well right i think probably there's reasons that ultimately like at the end of time i might want to be less sort of uh i don't know i might have to walk this back a bit but i kind of think that that's a story about how the thing got there not not about what it is in terms that are relevant to his status as representation like i mean i'm not saying a priyari represent you can't define representation in terms of some historical diachronic property i just don't see the motivation to do that and i

51:10 don't tend to think of it that way so like the the story about how it got to be the way it is for me is you know at least a priori distinct from the question of what it is and what it does and things like that um but i'm it's it's interesting to talk to you about this i mean i love um ruth milken's work um i happen to disagree on like almost every axis not completely though is really into the whole you know it's it's not just that one thing co-varies with another right it's it's that a whole structure kind of maps onto a whole other

51:41 structure which is totally what i think is that the essence a lot of people think is the essence of representation um but i also think we can talk for a second like when you when you guys say carry information um i i take it you mean that in kind of a dressy sense where like um well actually drezky's officially his account is in terms of like nomic co-variation right yeah um so that's interesting um i think a lot of philosophers have taken have have have kind of latched on to a concept of representation as like

52:15 objective statistical covariation and that's the thing that's often held up as the thing that's like well obviously that's not good enough for representation so you need something more um maybe there's something more interesting built into dretzki's actual idea of gnomic covariation but um i was going to get into a whole thing about how i think statistical co-variation is is not really what information is either like if you're if you look at like shannon's theory i think it's really an articulation of a folk notion of somebody's being informed about something so i actually think a reductive account in terms of

52:48 information reductive account representation in terms of information probably isn't going to work either because information is essentially an epistemic notion um and i think this the rise of the re the rise the uh continuing ascendancy of bayesianism uh you know points that up like we're talking about subjective probability distributions here even even when we're talking about like simple systems oftentimes um yeah so i don't know i think um

53:23 i think you can easily get into a trivial kind of account of representation when you pursue these lines of thought but um i i never finished talking about the free energy principle i'm not going to insist i do so now but i think one possible concept what one possible consequence of that theory if you read carl fristan's new kind of book thing uh the free energy principle for a particular physics is that everything is essentially a representation so like pan representationalism um as like a well i was going to say deliberate

53:56 commitment more like just as a a conscious commitment right but not not because as tristan would say like this is just a mathematical description of how things are he's not trying to to push this because it's controversial but that's that's one of the things that falls out of it is that literally anything anything that's a thing that you can distinguish from its surroundings is going to in effect um mirror its surroundings in a certain way it's going to act as though it's performing bayesian inference um i think there are trivial cases like rocks and stuff that

54:29 are not very interesting but pretty clearly like anything uh even like really low level uh you know things like individual cells are gonna have it's gonna those things are going to have representational properties for the same reasons that that human beings would yeah so can you say mar i mean i you sidetracked me from what i was thinking so that's fine to this what is this free energy how does this free energy principle work your way into pan representation does that trivialize the whole theory i

55:02 mean it seems well so it's yeah it's it's dicey because um it's not really dicey um it depends on your commitments like um some defenders of the principal have said explicitly that it's not an empirically falsifiable theory it's meant to be a principle um it's often compared to uh hamilton's principle of least action it's like something that you that that you can use to um describe systems and to frame hypotheses and guide research in various ways but it's not a falsifiable principle um so why should one adopt it then well

55:38 okay one thought about that and please shoot me down if this is wrong but it could be that it's not falsifiable because it's it's like true so you can't find any reasons going against it because as hard as you look you'll not find anything that goes against it because it's the truth and you can't um find it you can't actually find you could no i think that's that's you could uh conceive oh right so if you can even think of some so if you can't even think of a thing that would falsify it um that's you know that's um a different level of unfalsifiability rather than

56:11 like de facto and falsifiability right so i don't know i mean it's probably it's not really fair me to introduce the free energy principle by pointing to one of its most controversial implications i mean it is interesting though that like for the free energy principle is one of those things that you have people who are either like total adherents of it i think it's the best thing that's reorganizing our whole worldview and there are people who think it's total quackery and that you should dismiss it like out of hand and right

56:43 you know people make jokes oh though if you bring up the free energy principle then we're going to do this conversation it's not serious anymore and it is weird when something has that kind of really dichotomous effect on people and this may be one of those one of the reasons for that yeah that has this really radical consonant so to me like i'm kind of stuck in the middle of this to me the task is to understand why because i find it very interesting and uh potentially informative and like a useful tool for building models so the question is how can that be true even though it's potentially trivial and that it applies to anything i mean one thing i'll throw out

57:16 here and then i want to i don't know step back from it a bit because i'm i'm by no means an authority on this by the way i've done some work in this area recently but i'm still trying to understand the free energy principle i think a lot of people are um yeah well factors in your identity theory paper yeah yeah but um i mean one comparison i'd make is just to like so like what if physicalism is true then um in a sense physics describes everything there is and there's nothing it doesn't apply to but that doesn't mean it's not any good as a theory right

57:50 but it's like falsifiable like you know you can imagine at least uh you know in the worst case scenario you can imagine that when we started doing neuroscience we opened up our heads and there was nothing in there right now and yeah okay so and then you say well must be some thing which is i mean outside of the realm of physics which is accounting for this you could you could also have found that uh certain active potentials start generating spontaneously without any input from other action potentials so there's lots of ways you could find evidence against physicalism in that sense

58:21 well right but i mean i can imagine a physical theory being so like you can falsify the theory in the sense that you might find things that don't conform to what the theory says happens um i mean i could try giving you a really quick like substantive answer as to why the free energy principle is unfalsifiable which is it defines a system in terms of what's called a markov blanket which is like this set of states that separates off the internal states from the external states essentially in terms of um statistical dependence

58:52 and so once you make that partition um the it just follows mathematically that there is a you know the distribution of over the internal states is distinct from this distribution over external states so you can on that basis say that there's the internal states map on to probabilistically map onto external states so like you have a representational apparatus of a certain sort as soon as you draw the boundaries of the system in the terms of the theory right so the theory says this is what a boundary is

59:25 it's a markov blanket that by definition means there are certain statistical uh independencies in play and that from there it's just a matter of derivation to get um this the formal description that's involved in free energy minimization which is which is basically just an approximation to bayesian inference um so i don't know it's weird that that's true if it's true you know um i guess the question would be is it useful to describe systems in terms of these markov blankets and things yeah by the way i wanted to um

1:00:00 point out that there's two objections that seem to be getting to slam together over the past few minutes um that i think it's probably helpful to distinguish these objections so one objection is like the pan representationalism objection and then the other one is the paparian um unfalsifiability objection the the kind of like the pan representational objection i might be wrong about this but i think that that's

1:00:31 specifically like a philosophy type objection or even more specifically a type of philosophy of mind type objection so there's discussions in the literature of consciousness or our representation and someone will try to naturalize those and then someone else comes along and says well if that's true everything has got representations um but but outside of philosophy there's lots of things that are ubiquitous but but we celebrate so like equals mc square like uh

1:01:05 or just the solarity itself speed of light that's ubiquitous everything is bound by that everything is located in space and time or made out of matter slash energy so you ubiquity doesn't seem to be that big of a deal outside of these specific philosophical concerns which is not to say that they're bad concerns if i think you know philosophy isn't total garbage um but the but the the pomperion one is slightly different and right so even if something is ubiquitous

1:01:36 like um i'm offering uh e equals mt square seems to be ubiquitous like everything in the universe has to obey that um it still might nonetheless be falsifiable or surprising the kind of similar sort of thing like it's not something that like we just had been saying forever and ever it took a long time for people to to come around to um equals mc squared and whenever it gets there's some new experiment that is able to verify it

1:02:11 with even more uh precise detail that that's celebrated too like it's often offered up as one of the most well-confirmed uh scientific theories ever um but also i think it's like super surprising like we could have easily imagined it being false um so anyway i just wanted to point out those seem to be two different um objections and yeah we're i don't want to say one objection is more important than the other um but uh alex

1:02:44 some of the stuff you were saying kind of made it sound like you were treating them as the same objection um i may not have been distinguishing very clearly but i agree with you totally that those are yeah those are two different things um so ubiquity might be might be true like you might be happy with um pan representationalism but still worry about falsifiability or maybe you might be happy with both like you might say like yeah so what if it's unfalsifiable like there's certain things that um i mean people say the dark like the basic idea

1:03:18 of darwinism you know that you're the origin of species comes or or you know wherever you've got variable inheritance of fitness you're going to have darwinian selection you're going to have novelty of of species a lot of people have said that that you know once you really get your head around that it's it's like it has to be true like it it just couldn't be a system that doesn't conform to that but nonetheless it's super illuminating to be able to look at the origin of species through that lens of um

1:03:53 yes you've got these various things and you could specify a fitness metric for them and the ones that are gonna and if there's variation in in their ability to like pass that fitness on to their offspring then yes of course you're gonna get right yes of course we don't need god yay so yeah illuminating um so so so the question that that raises for me is why is it illuminating if it doesn't rule out if there's no way that it could be false and therefore

1:04:26 it doesn't really rule out that alternative right so if this is kind of a point about like what what are the meanings of concepts and terms like one account is that you know the meaning of a term at least should should be that you know it selects certain things as consistent with the meaning and other things is not it should rule out you know or a claim it should rule out certain possible states of affairs so like why is why is darwinism or the free energy principle why are these theories illuminating if if they actually if there's no possible state of affairs that's not consistent with them

1:04:59 i mean maybe the right response just suck it up and say like we put too much on the i mean the right response is that in darwinism they go to a lower level of mechanism they start talking about dna uh so if you just describe it this general level then you might get this sort of problem but if you go down a level lower then there are falsifiable things there um what's the unit of her edibility what's the thing which is selected the gene or the organism the trait or whatnot so there are things that are which are falsifiable but it's only described at a very general level

1:05:30 uh that this the notion of um natural selection becomes unavoidable but there's no lower level you can go here so say a thing that maybe turns out to be the same thing richard but what i was going to say is that you know there's a lot of math that we accept as awesome so take for example um the infinitesimal calculus and that was just a huge achievement for the human race to be able to do calculus and um but since it's math like we're not worried about its pan

1:06:03 applicability or its lack of falsifiability it wouldn't be super useful to use calculus to compute the area under a straight line it's more useful when you're trying to compute the area under a curve and map that on to like um accelerating objects and and and then you know now you're better at knocking knocking down fortress walls with cannonballs um but there's a but there's a sense and we're comfortable with this in mathematics of like well it couldn't be wrong like like once you really get your head

1:06:38 around it uh why it's true you you're not you're not bothered by the paparyan kind of objection it may be so the right way then to think of say the free energy principle or or maybe darwinism fits into this too is that it's math and um and math is great math is useful uh math is illuminating um but it it shouldn't be this regard as the same sort of

1:07:11 theory as for example like um the the germ theory of disease transmission right um yeah now that sounds reasonable to me i mean there are definitely people who say that the free energy principle what like i said earlier it's a principle it's not it's not an empirical theory so there's a sense in which it's more like math um you you could if you could say it's falsified in a sense if you found a system that um i don't know is an individual system but doesn't behave as those performing bayesian inference in any way or

1:07:43 something but you know you in fact wouldn't be able to find such a thing if you look closer at it so like um maybe that's that's a you know a prima facie kind of conceivability kind of like a zombie a philosophical zombie um but yeah i think there's a sense in which that's right it's like it's it's more like a mathematical description and there's questions about whether it applies to actual systems um although i think there's very good reasons to think it does yeah i guess my question is why is it

1:08:18 useful for cognitive science so it may be it's these things you're describing it's that's fine but uh i don't i don't really see its applicability of course i'm not deep into the steep race culture but i mean why isn't what can cognitive science perceive without it yeah so i think it's going to be useful for building better ai and robots like i wasn't convinced that at the beginning like i i think free energy minimization as an objective function for learning neural networks is absolutely important because it's been behind a lot of the good i mean there's an argument really that any kind

1:08:50 of unsupervised learning cost function kind of can be understood in those terms but that's another one of these you know it's a blanket here that covers everything kind of arguments but like it's it's it's guided the the construction of actual useful algorithms for unsupervised learning but i think one one spot that even people who you know were at the sort of foundations of deep learning people like yahshua bengio say we possibly need to advance to the next level we need uh virtual agents that that build models that are that are sort

1:09:27 of tailored to action um and so what the free energy principle and active inference type stuff does is build that in um in a very explicit way uh you know action and designing models that are specifically geared this actually relates to what you were saying about control theory being a motivation for well we were talking about predictive coding then but um anyway the the fact is the generative model although it's a representation of the environment also needs to function as a control system um and so one of the things that modeling uh free energy minimization

1:10:02 through both perceptual inference and action does is ensure that you uh that you have a model that's sort of action oriented um which might actually make for better ai and um and there's a whole connection then to like all the embodied cognition type literature which isn't really my jam so much but um but that's there if you want it so it so you're so is this going to carry over to a claim about because there's two claims i mean that i hear a lot one is that it's good for

1:10:34 building ai the other is that it's good for understanding the human brain um so do you think it's going to carry over into the realm of understanding neural functioning or is it just good for ar do you see a bridge between the two or i mean i think those two have to well they don't have to go together right i guess artificial intelligence doesn't have to i think we're still at the stage where to get better ai we're going to want to better model human intelligence but clearly there's possible divergences so like you know all this whole um to do about

1:11:07 adversarial examples and deep learning is interesting because it shows that maybe these things are reasoning or drawing inferences in a different way than people do um so i mean if anyone's not familiar with that it's like you can basically through sort of back propagation type tricks you can figure out how to perturb the input to a neural network so that it ends up classifying something that's like clearly a koala as gwyneth paltrow or something just by like changing a few pixels by a couple of you know fractions of a decimal or whatever so um

1:11:40 that seems weird like obviously people aren't um susceptible to those kinds of um mistakes so i do think you're very good at degenerative damages actually yeah yeah but i don't know this gets us into the singularity type territory like maybe there are ways of developing intelligence that are going to totally diverge from um what we're like um i mean there's also i guess the octopus right it's also the they're not exactly like us but so i mean it sounds like you're i don't know

1:12:13 how to interpret what you're saying but it does it sounds like what you're saying is that these ideas may not necessarily be good for understanding neural functioning works i think they will i i i think they will um i just i don't want to be on a soapbox about free energy principle but so i'll say i mean i'll say one more thing like there's um some new work that um casper hesp and maxwell ramstad and colleagues are working on that tries to model population coding in terms of individual neurons

1:12:47 performing you know what's called active inference so like basically a free energy minimization description of the components of the system at their level at their scale gives rise to behavior that's uh function functionally uh useful at the level of the super system and there are many cases where that kind of thing happens so also in terms of like multi-scale modeling like if you're trying to understand how a component of a system plays a role in a larger system which plays a role in a larger system etc um this kind of formalism can be useful for that too um

1:13:23 but i i don't know i like sort of standard vanilla neural unsupervised neural networks um i'm just starting to get sucked into the free energy principle stuff because it's just it's super interesting and i didn't mean to get sucked into it though i wish they'd had a better name because it really sounds like it sounds like a cult or something free energy principle is like uh yes if you just like believe really hard in your dream you're gonna get rich or something no i know i know people hate it it's oh god ever since i brought the

1:13:57 free energy principle into my life i've lost 50 pounds and um but you have to minimize your free energy okay so that's all yeah yeah there's the rub right you have to minimize it why would you want to minimize free energy that's an awesome thing you want to maximize it exactly i'm pure free energy okay so you'd like vanilla unsupervised learning um are there any good examples of unsupervised learning that's that does anything interesting

1:14:30 lately um sure uh all the interesting stuff i know of is totally supervised and back propagation plays a huge role and which always sounds seemed like cheating to me not not very interesting right i know i mean it's interesting but it doesn't really you know it's not going to explain what's going on yeah um so i mean i think if you look at any impressive like neural generators like uh so there have been also i mean there's been many examples

1:15:02 of language models um that are okay of that some of that is supervised in a way it's like supervised in the sense that you're trying to predict the next word given the previous one or the next set of words or whatever or context um but that's one example that arguably you can understand in terms of unsupervised learning um just just image generation um i don't know if i wanted i could in the background try to find an example to show you i don't know if i can share my screen can we do we have that technology here yes we do so you you guys like keep

1:15:35 talking i'm looking for something real quick in the background okay um but yeah i think i mean i think there have been there have been plenty of examples of tremendous examples tremendous amazing examples um i know it's not convincing um uh yeah and because i hear rumors lately too that people are talking about mixed systems

1:16:06 lately involving uh connectionism plus some kind of old-fashioned symbolic manipulation as maybe being a gateway to some better functionality yeah i mean i'm not convinced that that that's the way we want to go but um you're not no i i mean maybe but i don't think we need to build in more classical assumptions i think we're doing great um oh one other example is like so all these like style transfer type

1:16:39 algorithms um now i can't commit to saying that i know that every one of those algorithms is an unsupervised learning algorithm or is based on one but like um you know this algorithms that allow you to input any image and then it spits out like a you know a version of the image that looks like it was painted by rembrandt or something or that's the same sort of principle that's behind deep fakes and like placing someone some celebrities face on someone else's face or switching switching people's roles in films and stuff that's another interesting example um potentially

1:17:15 and those are unsupervised um like i said i can't it would be irresponsible for me to say they're all unsupervised but i know that you could do that by unsupervised learning and that many algorithms do um can i share my screen here yeah so here's an example of here's an example of just like roaming through the um hidden representation of a generative adversarial network so like the latent you know weight and representation as they call it and this is how the output changes as

1:17:48 you just wander through the latent space um so these are all fake people you know um are you just seeing me no we can see okay i still see me so i was like that would be pretty funny um so this is just a system that's wandering through the possible parameters yes right so it's a learned generative model and this is what i'm sorry yeah what the [ __ ] are we looking at i mean i know all right there's a bunch

1:18:20 of faces oh yeah so this is changing but uh what was the network so it's a generative adversarial network we could get into that if you want that's something that that hawk one's been talking about a bit for a while network's like supposed to do or what was its outputs so so okay so there's there's two sort of sub networks there's one that generates generates things that are supposed to look like the input that's received in another network that tries to tries to distinguish between the fake things that the other network generated and real images and so that's there's sort of a joint optimization function that learns both

1:18:53 components but basically the thing is just learning in a really with a nice kind of cool unsupervised learning algorithm it's learning how to generate faces you know right based on the statistics of the input and generating faces from what though what are the inputs the inputs are just a bunch of pictures of real people but after it's trained what what is it supposed to do this will generate fake people right so these are all just these are all just you know fantasies as they say um that the network is generating and so like

1:19:26 here you're just like if you think of the the hidden layer representation i don't know how many hidden layers there are here but take you know whatever the highest hidden layer um and you know you can map the all right can you stop sharing that while you're explaining this because i'm feeling nauseated sorry sorry these imaginary people oh i showed that you know what actually i showed that to students one time and one of those especially freaky faces came up and i i was like i'm sorry i i thought this was safe um but it's a little freaky

1:19:58 yeah so what so what happens you have the system you input a bunch of actual pictures of people right and it learns yeah so so it learns to generate those images um based on inputs of actual you know thousands millions of actual people um now so i think that's that's cool um i don't know if it's useful um i like art so i'd like to see you know a lot more art like that but that's you're saying you're using that as an example of

1:20:30 something an unsupervised learning can do so there's no there's no back propagation there's no adjusting of the weights by a person in there anywhere i mean not by a person i mean really that really back propagation isn't that different from unsupervised learning it's just sort of like you're using the input channel as the signal from which the error is back propagated in one way or another so like i don't see a deep difference there um but it's more much more what we call what we think of as unsupervised learning is a lot more plausible because yeah because you don't need a teacher

1:21:02 telling you you've performed correctly on some downstream task like you've classified the thing correctly um in terms of the processing it's not necessarily all that different you're still um i don't know jeff penton has some work that shows that certain um ways of learning an auto encoder which is an unsupervised network end up being equivalent to back propagation um like locally so uh so as algorithms i don't think they're that different but yes so in this case you don't have any there's no teacher telling the network uh you've done a good job on this um

1:21:37 i guess actually in the case of generative adversarial networks there is the discriminator network is told whether the uh whether it got whether it gets right about whether the image that's classified as fake or as real is actually fake or real um so that's that's kind of a sort of supervised learning i mean yeah but one of those one of the problems well it's just that's just one way of optimizing generative networks it's not essential to what i just showed you but it is kind of a really good trick to help

1:22:10 them perform better anyway one of the problems with with backprop uh is the the cheating one right that you know you've got the right answer but in the wild we don't have the right answer we have to learn we have to figure out and there's not a teacher telling us and this we're able to learn things um based on a pain signal or something like that um so that's one objection but another another problem with backprop has to do just with with some other feature

1:22:41 about whether it's biologically plausible or not and that is that you've got to run this thousands and thousands of times so if you like you're you're like old-school three-layer face recognition network from some old paul churchlin book to get to just train a thing to be able to say like yes or no this is a human face or to categorize human faces as male faces or female faces you're running the back prop you're doing the like right it's like thousands and thousands and thousands of times and we think that

1:23:12 that's not biologically plausible that any of the sort of learning that we're doing you know like i don't know about you guys but my kids i only have to tell them once yeah there's no one-shot learning in these systems there's one shot learning in some of these isn't there um yes there is uh but by and large all of the most successful networks do tend to be very you know data hungry as they say that that's generally true it's like thousands of

1:23:46 like you're showing the test of the training stimulus yeah letting that percolate up through the network making some measurement of how close it is to the correct answer and you're going through that process literally thousands and thousands and thousands of times yeah yeah so true of a back prop training regime i think i don't think it's essential to backfall but i think that is how it tends to be implemented in like most of these networks like richard said there are there are networks that do um you know one-shot learning um but i think i mean i think there's something to that objection this is why

1:24:18 i'm sort of attracted to this newer work that jeff hinton's been doing on capsules which is kind of like connectionism 2.0 or something it's like let's make each thing that used to be a unit into a set of units and have these sets of units connected by weight matrices and then do a more fancy um sort of updating updating algorithm so i think i think that there's probably more to the story than what we've got now um so i would concede that like as a fan of connectionism that like i think

1:24:50 there are probably complexities that we still need to build into the model but but the important thing for me is that that's going to be at the level it's still going to be at the level of like domain general and hopefully unsupervised types of learning algorithms and once we come up with a better one of those we're not going to need all sorts of special purpose things like language modules and all that crap that's a big bet that sounds falsifiable yes i would bet you three at least three american dollars [Laughter] that's a big bag by the time we'll be

1:25:22 able to settle this ballot three dollars would be worth a lot more yeah with inflation um yeah that's really interesting so you think i mean you think that this is kind of like a universal solvent or something like that once you get it right it can do any any task probably i mean you know what what is it a two-layer neural network can already do a compute any function in principle but that doesn't mean you can actually train it to do so but um there are all sorts of

1:25:55 you know proofs of um universal capabilities of of even simple neural networks if you if you have enough hidden units right right how many is enough for like yes depends it depends i don't i don't know [Music] what's what alex what's cortical specialization for then what's it for yeah i mean or am i mishearing like you a sentence of the universal acid thing and yeah yeah one way of hearing that is that you know

1:26:27 it's against modularity it's a that would be against um special purpose uh modules that have special purpose learning algorithms um but there seems to be a lot of evidence for specialization in the brain well they're certainly screwed by this simple fact or uh probably you're not no no why then do we see um the kinds of specialization that we see across individuals

1:26:59 right i mean so okay so one one answer is you would expect to see specialization arise as a result of doing unsupervised learning in an initially unspecialized kind of network so like you would expect that lower layers of the network would come to encode things like edges because of their proximity to the input and then that the next layer would come to encode collections of edges because of their proximity to the first layer etc so you'd expect to see and you could you could

1:27:30 run the same kind of argument with respect to areas of the brain that are proximal to certain sensory input channels you know so like you would expect to see specialization arise i think there are a couple papers arguing that you know the dorsal versus ventral streams um just because of their separation end up taking on these two different functions so just to to make sure i'm understanding this if so like you you richard you and richard and i we all pretty much use our occipital lobes for roughly the same

1:28:02 things in our frontal lobes for roughly the same things as each other but if um if we took like that my next child and at birth like change the sensory wiring so that like the inputs so that for example um occipital cortex is no longer relatively early and frontal relatively late in visual processing stream yeah they would be fine and they would wind up with occipital lobes occipital lobes in the front i mean i don't know in the back i don't

1:28:36 know if they'd be fine but i do know that's what you're committed to some of the some of the research that motivated unsupervised learning in the first place was research you know barbaric experiments on animals of course things like cats that show that if you yeah if you rewire inputs that are supposed to go to visual cortex to audio cortex then the auditory cortex ends up doing vision so like that is one yeah that's one reason to think that there's an unsupervised learning algorithm i mean i think it's a lot of good reasons but so i don't know if they'd be fine right because

1:29:08 you need to have i don't know it's probably the probably differences in efficiency and you know how many neurons are available in certain areas so it's not like you can just scramble the whole thing and it'll settle into a state that works fine but um i think that's that's one kind of answer but also you can have like you can have innate stuff i mean the whole structure of convolution i was talking about earlier you could think of that as like an innate prior um and as far as modularity goes like you could even argue that layer wise separation is a kind of modularity um it's a weird kind of modularity there's some interesting

1:29:40 papers on this actually like zoe drayson has a paper on modularity in the predictive mind um and there's some there's some more recent stuff but um you it's not like it's just gonna be a homogenous field and neurons like yeah um gotcha i don't know yeah i feel like i'm trying that's interesting well yeah we all yeah you're doing a good job um so can i i want to ask about i mean we're running out of time here i don't want to keep you forever i know you have a real life to attend to but i do want to ask you about some of the more specific things that you're working on lately um like your account of inference for example

1:30:14 uh where you defend the idea that what's going on here are literal inferences in the perceptual system right maybe this will connect to some of the stuff we were talking about earlier because it sounded like in a way pete the way pete was talking about representation kind of defined it as intentional something conceptual maybe i know p has predilections for conceptualism things of that nature but there's nothing like that going on in your account of inferences there there's not a relationship between intentional states by which i mean conceptual um um i don't know what are we what do you mean by conceptual

1:30:49 uh well that's a good question but i guess i would mean has um uh is well i don't i can't say without going a circle but kind of thing it figures an inference okay um so in that inference paper you're talking about i have kind i start out with kind of a very minimal view of like what propositional content is which is this this is something that divides people and uh i probably should need to come up with a better term but i mean like in i guess stallnecker's sense like proposition being just a set of possible worlds or something or

1:31:22 possibilities be less metaphysically heavy-handed about it so like i think you can understand almost any representation any representational content really in terms of that so like especially in a probabilistic system like if each representation is associated with a probability well what has probabilities well states of affairs events you know sentence size things um so i think you can there's a sense in which any content if you want to know what the thing is representing you can specify in terms of an exhaustive enough um proposition propositional description so like um

1:31:58 if you have it that propositions sort of propositions all the way down then uh then there's no obstacle to saying that at least that's that's not an obstacle to saying that inferences are literally taking place at um you know lower levels of processing for example so can you i can i'm not sure i'm following so take the level of the retina right so if you have some um some bipolar cells in there some uh connected to some retinal ganglion cells you know some some light hits the

1:32:31 the photoreceptors and these guys start firing so there you're gonna say there's propositional content at that level um i think relative to the whole system in my in not my sense in this minimal sense of propositional sure i realize that sounds crazy to some people because some people are bringing to the table that propositions are these structured things that have linguistic structure obviously it's not what i'm talking about yeah so you're saying just just in the sense that they they divide the realm of possible space so

1:33:04 well yeah i think they represent to the organism that something is is the case okay in what sense do they represent it to the organism because why don't they represent it to the retinal ganglion cells one layer back as opposed to to the organism well because i'm taking the content of these things to depend on their functional role within the overall system um so look you might be able to say something actually i mentioned that work earlier on individual neurons performing active inference so there is going to be a description of the contents of the states of the

1:33:37 individual neurons from their point of view so to speak but that's not going to be the same description of that you give of the entire system um so i mean i'd take this just to be like once you've defined the system then the representational content of the system is given in by the structure um functional architecture of the representations and the way they hook up with each other so like i would say that these you know what retinal gang galilean cells represent what they do um in virtue of what they sort of inform the system of

1:34:13 okay and so in what sense can you just tell us more about what this what sense of inference you i mean i've read the paper and i sort of yeah you're gonna say it but can you just tell us about what how this is an inference on your account well so i yeah so i'm not sure if that paper was the best strategically because i kind of go for a uh a philosophical kind of inference that a lot of people wouldn't be into anyway and then try to say predictive processing influence implements that so it's inference but it's kind of a harman's sense of inference so the idea is that you just you take a collection of beliefs or belief like states

1:34:44 as input and it yields an entire entirely new collection of belief-like states as output and it's um you know um a thing that sort of happens in parallel for all the things you believe but i mean the core definition of entrance i just i just take an inference to be a state a transition between states that's in fact um truth preserving like as a matter of fact it tends to be truth preserving that doesn't have to be like with the strength of you know of logical entailment it can be

1:35:17 like uh probabilistic uh notion so like you know tends to preserve the truth and what's the sense of preserve here like if you give given true inputs you'll get true outputs okay so it's just it could be just so wired into the system that this happens yeah i mean it does it's not there's no commitment to a specific like way of implementing this at this level i'm just saying like what what i said an inference was in that paper i said it's just a truth a de facto

1:35:50 truth preserving transition i suppose it has to be truth preserving somewhat reliably like over more than one case at least at least like a couple cases but it could be possible right they don't have to be actual uh so yeah yeah so i mean i guess what i'm trying to work my way towards to is something that connects back to the beginning of our discussion because it sounds like from the things that we were talking

1:36:23 about previously um about what representations are on your account um and how you want to reject uh well yeah so can you tell me why this isn't involving like a teleological notion of inference because it sounds to me like that's sort of what it is so i must be confused somewhere i think again i'm saying that inferences are transitions that as a matter of fact preserve truth i'm not saying that they need to be designed to do that or that they they have the function of doing that in

1:36:55 a more robust sense than just that's something that they actually do and i think that used to but they don't have to actually do it because you said they could be possible i guess i had in mind some hypothetical i don't think this would occur in a real brain but like some state you know transition between states that only ever gets used once and it in fact gives yields a true output given true input um but suppose there was a system that just popped into existence and it performed there was one state that led to another

1:37:27 state it was in fact truth preserving and then it ceased to exist like you know yeah i think that was known as an inference that would count as an inference because of the possible other transitions that could have occurred well yeah yeah yeah so like um right it's not that you need like more and more of them to make the case more compelling but it's yeah i think that kind of state that kind of transition counts as it as an inference just because it's sort of a non-accidental sort of truth preservation

1:38:00 but not accidental i think you can cash out just in terms of the categorical properties of the system and what it's thereby disposed to do i don't think you need to talk about functions okay i want to try a slightly different tack on you alex and i am i'm enjoying the game of trying to hoist alex on his own puts hard wait all right fine so um so one thing that you might say uh which i've mostly been hearing you say is like you don't really have to make any appeal to functions as a theorist trying to understand

1:38:33 how this thing works there's no normativity at all you don't have to ask what it's for well so that's one kind of view our distinction i suspect i'm eventually going to have to back off on that but that's what it is okay but so here's a different kind of uh choice instead of the choice between some normativity and zero normativity here's a different sort of choice and that is a choice between um like really real normativity versus uh kind of den it as if design stance normativity as i

1:39:06 understand the den it point there's certain systems for which we're going to get maximal like explanatory or predictive power if we adopt at least tentatively some kind of view about what it would be for but then it is explicitly um denies that there's any like real answer to the question what is it for he he doesn't go off looking for some kind of specific evolutionary story that's going to deliver determinate norms he thinks that all that it's like semi um real

1:39:43 and kind of going to be viewpoint dependent and there could be two different theorists that give completely two different uh sets of normative assignments and two different sets of like decompositions of of a system and um end up making different predictions but equally um successful in terms of how many predictions that are accurate um so so the the different choice that is instead of like zero normativity versus some normativity a choice between two

1:40:17 different um kinds of normativity or two different metaphysical views of of the of normativity one being what i see is like the dretzki or milliken sort of project where you know it's got to be it's got to be rooted in darwinism real intentionality comes from real normativity comes from real evolutionary histories and then this on the other hand this more genetian thing which is like i mean we'll never know what those evolutionary histories are we're fine not knowing

1:40:50 because all arguments from tone of voice slightly made up anyway um so alex if given that choice uh would you go along with like the more donetian thing yeah i would if you can't tell i'd be happiest thing like i could tell you know really yeah you do have to make some kind of like propositional content assignments and that's going to be against the backdrop of some kind of tacit assumption of normativity like what is this part of the brain for what's it doing for the organism uh how is it serving the organism's

1:41:24 survival um but why is there is this obsession with survival we're talking about representation like what else is your brain for um well it's very good at helping me survive and that probably is a necessary condition for all the stuff that's really fun like imagining stuff and writing music and so on but like i really don't see that it needs to be tied to that so i would go for the den at option if you give me only those two options like i'll do the then adoption doesn't have to necessarily be about survival you might for example um you might be someone who is

1:41:56 um uh a music doctor or a sports uh doctor and and so you're gonna be evaluating brain function relative to like what's gonna help this patient better like uh judge saxophone performances or whatever their job is um and that's not committing you to like saying all human brains have evolved to like listen to the saxophone yeah totally nonetheless it's a temporary passive assumption about what would be better and what would be worse vis-a-vis this particular do you think that you need to

1:42:29 make such assumptions in order to describe usefully describe the system as as as representing things or as having contentful states i mean because i actually like that the account of dennett's that i like is this account in terms of notional worlds i don't know if you guys that's i don't know it seems to be a somewhat neglected view of his but i was working on something actually developing that idea but like the idea is you look at this physical structure right you look at the at this organism or the system you're considering and you try to imagine the kinds of worlds in which it would be useful to be this kind of thing in various ways

1:43:01 and so to me that doesn't necessarily bring in anything about any particular kind of normative normativity it's just like what sort of world would be a good fit for this kind of thing but i think it smells normative to me but it's yeah you know loosey-goosey right so right so what i wanted to do was fix that idea by taking the normativity completely out by bringing in jeff hinton's stuff and saying uh well just look at the kinds of worlds that this thing would uh generate sort of top down using the generative model so take a

1:43:33 human brain look at the kinds of worlds that it believes in by by you know exploring the uh you know as you'd say the latent space or the hidden layer representations and seeing what sorts of sensory inputs get generated by the thing and then if you do that and you explore the entire space and you you see all the different peripheral states that could be generated by this thing then that constrains down to as much uniqueness as the organism knows about which world the thing lives in and thinks that or thinks it lives in so i take it that's a totally potentially totally descriptive account the thing that philosophers won't like about that is that it's

1:44:06 massively indeterminate probably with respect to certain things but i don't know i don't know if skepticism is false i don't know if the thing i'm looking at is a tree or a hologram or what so um i don't so why i don't know i feel like there's got to be pushback like why do you why why do people i can't say why why do you why if you guys think that normativity uh so the so far the argument the the argument that i remember currently for normativity as a constitutive aspect of representation is that you need to be able to say why like um a malfunctioning belief is still a

1:44:40 belief but is there a deeper reason that we need normativity well that's one reason there's another reason what you could say you need to be able to uh say to well it would be nice to be able to say what the content of representations are and so like you have crimean kinds of worries um is this river is this state representing a dog a temple slice of a dog is it representing um a doggeth which is you know part dog and parts um oh yeah logical function if and then you

1:45:14 can say well if this state has the purpose of representing dogs then that's that answers that question whereas if you don't have something about normativity in there what what the state is for is representing dogs not dog slices temple slices or uh you know weird combinations of dogs and logical functions um and if that it seems like every theory of content eventually runs into this problem um antelogical theories at least have they can wave their hands around and try to give some kind of answer what your kinds of theories doesn't seem like i can't um can i sneak something in

1:45:49 please so just to build on what richard was saying one sort of thing in this ballpark that you might be worried on about is intentionality with an s um yeah and i'm happy with indeterminacy like i'm not on a quest for determinate content um and so i wouldn't be motivated for to bring in normativity based on that but i do think you might nonetheless you know if you're trying to model um certain aspects of folk psychology one

1:46:23 aspect of folk psychology is that propositional attitude descriptions are intentional with an s and i'm wondering if you in your so-called purely descriptive project project we're all friends ladies and gentlemen whether you can get that intentionality with it with an s i i don't think i can get anything else extensionality um i think it's harder to make the argument this targets back to

1:46:55 like some taco gave um savannah or something many years ago but i don't the way i think of it is this this whole system right and the hypothetical worlds that can sort of generate right as evidenced by the sensory fake sensory inputs the system can generate that gives you the sort of the worlds that the organism believes in um it's an extra piece of theory that you need to link that to any actual state of affairs so like i think it's harder to get but you know this is kind of like the you know the unicorn problem that

1:47:28 some of us have talked about at length before um i think it's harder to you need additional machinery to provide a convincing link to the actual world in this kind of account that i'm giving but maybe i haven't explained it well enough if that wasn't obvious already so um right but i guess i'm assuming that the content is somehow determined by the um structure of representations and the way they interact um and then you can ask the question does that structure match any actual structure in the environment that the

1:48:00 you know animal is engaging with why are you guys so happy with indeterminate content uh i don't like it i'm not happy with it there's there's like facts hard facts of life you know and one of them is that i don't know if skepticism is false i i i don't know i think skepticism in this context is a red herring sense if you're looking at the tree um it's a tree whether it's a hologram or not that's interesting so yeah okay it's a tree even if it's hologram trees are holograms kind of thing

1:48:32 yeah exactly all right well that's depends anyway so i mean i don't think you need to answer metaphysical questions like that to talk about uh but then isn't that kind of what the determinate content is of of a particular state you need to find out i mean that's what the good of geological theory is you say what the what the function of that state is what its purpose is what is it therefore but on your account you can't give an answer really to that question it sounds like what is it there for i don't think that problem comes up so much like so at least maybe i misunderstood but i thought you were talking about cases like

1:49:06 cases like sort of disjunction problem type cases which definitely come up for causal type theories but i think a functional role account doesn't face that problem in the first place because like why why does this state represent a dog well it's because it's connected to all the doggy things if there's anything sorry it's this representation is connected to all the representations in the right way that um that represent things that have relations to dogs and in a way that mirrors the relations among the states so like the time slices of dogs

1:49:38 well no no no because okay okay actually as you know i also think quinn was right about a lot of that stuff so exactly i'm not worried about this are you guys worried about those cases like if there's anything more in the box whiny in case that you you can't if you really if there really are indistinguishable alternatives then yes i think those are beyond our representational reach like we can't just we can't decide between genuinely functionally indistinguishable or you know alternative contents um to which we bear the exact same

1:50:12 functional relations yeah so that's where the teleological story can come in handy though is because it probably is not useful the reason the state is there is not because it was representing time slices of dogs but rather that it carries information about dogs as organisms and that's useful for us getting around so that we can give an answer why the state has that kind of content as opposed to the time slice content but wouldn't representing wouldn't time slices have done just as well for getting the system to survive and you know have the thing states that it currently has

1:50:45 no time slices have weird properties and things are true of dogs that aren't true of time slices well right but the time slices insofar as we interface with them they seem to have the right properties by hypothesis they have the right properties to get us the representations we have and they've right and unless there really is a again unless there's a difference in the functional interaction with the thing but if there isn't then i think this kind of is related to voters you know kind of crusade against darwinism in the later years yeah right about natural selection being intentional or not

1:51:20 an s with an s yeah it's definitely intentional with the t right i mean all most all persons who are knowledgeable about darwinism think that's like a joke kind of so it might i mean i took it more seriously than a lot of people i know did but uh anyway so i i really see we're running out of time there's one other thing i wanted to ask you about um i mean there's a lot of things i want to ask you about but uh um time is finite sadly uh so i wanted to ask about this sort of came up when we were talking about the

1:51:53 anniversarial networks um and you mentioned haqqan uh and so one of the this is relates sort of to some of your earlier earlier work since i've known you for a long time i remember how you kind of started off on this area and one of the things that's um that you were explored i don't know why it changed but i kind of want to get you to talk about it a little bit is is how these like in predictive processing models and in general adversarial networks in general why isn't the fee the top-down signal a representation of what's happening at the lower levels of that system oh yeah

1:52:24 you can map a map on a kind of higher order i mean part of hawkwind's motivation for talking about gans is because he thinks it sort of lends support to the higher order view that he wants to defend um and so i wonder if you think that that's kind of uh the right path or if there's uh a reason for thinking that the representational contents are not going to be about the first order states or about the lower levels sorry um no it's okay i'll wrap up soon um yeah yeah so i think uh i mean this is one this is one thing

1:52:57 that i'm still that i'm still working on very much but um i do think so i think in part of what you said was why are the top-down signals not representations of what's going on to the lower level and my potentially half-baked answer is that they are representations of what's happening at the lower level maybe maybe in terms of the contents of those representations not in terms of the vehicles um so like and i think that's consid i think that's consistent with

1:53:29 them having the correct sort of distal content as well um so this goes back again to alan newell's work which i'm still trying to sort of plumb for like philosophical you know purposes but um his view is basically that when you do represent an external thing um it's essentially a matter of simulation and it's in virtue of a chain of sort of designation relations these sort of dependence relations of some sort obtaining within the system so uh for me the holy grail would sort

1:54:04 of be an account of representation that has it that you represent distal things in virtue of your internal states representing one another in the right way um but you're probably gonna have to place some constraints on that although i i can't do justice to it right now but i did work out a few cases where um it would reasons for thinking that it's not crazy basically to describe things in that way so there's a there's there's a worry that i always have about this stuff and i first got turned onto the worry through

1:54:36 uh guffan gazelda ray at duke university um when you're when you're trying to give an account of a so-called higher order representation you want to know well in virtue of what is it a representation of a representation and if it turns out that it just has the content of the first order representation if it has the same content as the first order representation then itself would be merely your first order representation

1:55:10 so if i've got some first order representation of a tree and then there's this other thing in my brain that's triggered by that first order representation and what it represents is simply a tree then the fact that it's later in the processing stream doesn't make it higher order in order for it to be higher order it would have to in some sense represent the first order of representation and simply right have to represent that you represent a tree something like that yeah um and you know just representing the tree

1:55:42 won't be good enough it's tempting to me to say that it has to represent the tree and the first order vehicle because i don't know what else there is um alex yeah solve this problem i mean i'm gonna need some time on that honestly um so your argument was that if you're only representing the lower order

1:56:13 representation in terms of its content then it's going to collapse into just the representation of the thing that was represented by the first order state and therefore not higher order so so i can't promise that i have a coherent thought to give about this because it's you know i'm working this out as a this isn't something i've written about yet um i've taken some notes but um i think one thing that might do it maybe it doesn't end up being sort of a vehicular representation so what i'm thinking is if if some kind of noise or

1:56:45 you know um if some kind of noise gets into the picture then you might be able to make a case that lower order states are being represented so what started me thinking about this actually was um so here i don't know when you guys close your eyes or when you look at sort of a you know a blank expanse of a certain color do you see any kind of like little fuzziness yes right so that i assume is some kind of neural noise or something and by the to my knowledge david rosenthal maintains that he does not right so i look at this he denies that he

1:57:17 he sees any evidence yeah yeah i look this first i first noticed it on lsd his brain's not noisy late 80s and and i never stopped noticing it ever since so cool right did you just say you noticed it on lsd that's when i first like noticed visual noise well you don't need osd to notice it but i'm sure yeah exactly kids don't do lsd no don't it's seriously um but um i think that would happen alex's background yeah exactly anyway but go ahead um i like it so so i um this up and then people people

1:57:53 call it visual snow um and i guess maybe there's some controversy about whether it's a normal feature of human perception or whether it's a sort of you know unusual thing i think it's pretty normal based on this conversation but like that's an example of like arguably of you sort of seeing the external world through these layers of representation at least one layer of neural representation like you could take a photograph of a photograph it could be indistinguishable from just the first order photograph but there could be a scratch on on the photograph that you photographed and then you might get something that's more like visual snow

1:58:26 so i don't know if this is actually an argument it's a cool thought that i want to develop but it's very cool but but in in the general i mean in the general predictive processing framework there is no representation of the vehicles right i mean so like the way i think of a so you have a seeing you see red then higher representation says i am seeing red so you represent the seeing of red but is there anything like that on the predictive coding account or is it just literally a kind of regeneration of the signal i'm

1:59:00 not sure if i'm i know what you're asking i mean one distinction is between explicit a con a state that explicitly has the content like i see x or i am i am in a mental state and i think you know i think hot people higher order thought people can can if they want to dissociate themselves from the kind of thing that's going on at every level of these predictive hierarchies by saying you need this robust sort of self-involving content which maybe isn't present at every level but um i'm not sure if i'm all i was asking is in the in the top-down uh predictive signal

1:59:35 is there a prediction of the so this is just pinging back up pete's worry is there just a prediction of what the content is is there a prediction that there's a state with that content um so are we are we talking at the level of content here like is part of the content of let's say layer three which is i know so sorry to be so generic it's part of the content of layer three that there's a state at level two with so yeah that's the question i kind of wanna say no

2:00:07 just because i don't think that's i don't think it depends on the level we're talking about maybe maybe at the levels at which high order thoughts proper occur maybe that's the case and i need to explain why and so i don't really i'm not just thinking about higher order thoughts i understand about i was just thinking about in this framework is there a representation of just the content of the lower levels or is it representation that there's a state with that content at the lower right i think i think i i think that it's the the former i think that each layer sends a prediction that really causally determines the state in part

2:00:39 the state of the layer below and that state that layer below is itself a representation so right the effect of the higher order state is just to induce a representation at the lower level that also isn't but it's not representing that state error so that's just kind of the p point it's not it's not representing that state um i mean it depends right it depends what you mean by representation and such but um i mean you could argue that it's not representing that state because there's nothing about your right so there's nothing about the overall representation

2:01:14 that that suggests that you're representing another representation or a state that's if you just looked at that representation you would never know that it was like but so i constructed it out and just looked at the content of it it would be like tree it wouldn't be like there's a state down there that says tree so i guess what i'm trying to argue at least in the case of a vision is that um maybe to the extent that things are going well or to the extent that there's teleology in there to the extent that um you're clearly just perceiving the external world without

2:01:47 hallucinating or seeing visual snow or whatever you're just representing lower level states in a transparent way um but i don't see why i can't say that you're still representing the lower level states in terms of their content it's just that that doesn't really again unless there's noise uh unless you notice some property of the representation that shows that it's a thing in its own right rather than just a conduit for uh sort of revealing i guess what i the way i was interpreting the

2:02:19 kind of challenge here was that so we want to one one thing you might mean by representation you want to be able to say what the representation of is of but it sounds like in this model it's even hard to say is that a representation just of the content is the reputation that there's a state that has that content so you can go either way so there's nothing in the theory which really says this is the representational content of these states so it sounds like a more loosey-goosey sense of representation let me ask you i mean do you think this is connected to a point about holism in general like

2:02:52 semanticalism so there's a sense in which in any kind of holistic theory which i take it that sort of structural representation is one of there's a sense in which the the meaning of any term or any representation is like an encyclopedia you know it's it's the connection of to every other representation so i mean there's a sense in which you can't pin down especially in statistical modeling right where you might i think anyway there's there's a sense in which you could say that you represent some things better than others but

2:03:26 if there's some even like minuscule dependency with some other thing depending on your account it might turn out that that's part of the content so it's hard it's hard to it's hard to give this is something that grand priest gave me a hard time about when i was defending my dissertation my prospectus is it's hard to give you know necessary and sufficient conditions for a vehicle to represent one thing and not another um but i think if you talk about i was there by the way at that i hope i'm not being any in any way like that into the question i'm asking

2:03:58 no i mean you're being more gentle and richard like um so here's the thing i don't know i love grand priest by the way yeah yeah um but i do like bringing the visual noise in into the discussion of this this problem that i attributed to um gavin gazelda um but now they're you know thinking of your like notional world or possible world uh framework for um propositional content we could we could cook up a

2:04:31 world and an organism such that um that world is noiselessly represented as having sparkles in it and this organism represents those sparkles noiselessly okay and then a similar world where that's noise i mean in a world without sparkles it's a perfectly clear world but you you represent it in this sparkly noisy way yeah and you also have this additional

2:05:04 like conceptual apparatus you know about things like visual systems and and the way your world works and how it's uh sans sparkles yeah so since you've got all this um extra conceptual theoretical knowledge and also as a theorist are relatively happy with holism you could say that the way this organism achieves the higher order state is through some kind of like theoretical inference that like well these sparkles these sparkles i'm seeing aren't out there

2:05:37 what's out there is the tree and the sparkles must be due to my seeing of the tree uh and now you've got a higher order representation and your counterpart in sparkle world might just be stuck in the first order they're they're they have like the same first order state that you have and are representing a sparkly world perfectly accurately and it would never be need to be prompted to the like higher order state but so i think this is a good solution and i just want to point out

2:06:10 that like probably you're going to need the holism to pull it off like strictly strictly information like causal informational views of representation i don't see how they get out of the gazelle deray uh problem yeah well no you the detector of a a detector of a trip a state of a tree detector itself turns out to be a great tree detector and so it's really hard to see that you get a first order second order distinction but if you bring in all this holistic stuff then it seems a bit

2:06:44 easier right so i basically i think that's all right wait hold on oh [ __ ] if if you like higher order as a theory of consciousness you might wind up allowing only relatively sophisticated conceptualizers have conscious states because they might have to like have a theory like a full-blown theory of mind or something like that maybe a rat can't be conscious okay maybe maybe um yeah i mean i don't know uh richard

2:07:20 what do you think [Laughter] uh consciousness hey man i don't know no um yeah no i agree with all that pretty much i was gonna one boring quibble make it really fast like i bet the sparkles wouldn't sort of there'd be a lack of dependence with sort of other with properties of the world you know the constituents of the world that you're representing that would make it maybe possible to distinguish at earlier levels whether you know whether you're representing a feature of the world or right like if the sparkles didn't interact with anything in the world

2:07:53 which seems to happen in the case of visual snow i think you can make a sparkle world well i mean you can have an external environment that appears in the same way that your visual field appears to you oh i got you because because your your visual field isn't that impressive you wouldn't you would need like you could do it with an iphone retinal display that's why they call it a retinal display um so by the way there's this artist um i have to look up

2:08:26 uh i have to look up the name of the artist he does a lot of light uh sculptures and he has this permanent exhibit at the knitting factory in pittsburgh in which you you are led into this completely like almost completely black room and you have to you sit in this chair and you just stare ahead and after your eyes become dark adapted you're able to see this like paint this faint like thing maybe it's a violet disc you're not sure like it's so like faint and everything like that and what it really is is like

2:08:58 he measured like exactly what the wavelength of the dark adapted eye would be that would be like just across the detected detectability threshold after like two minutes of dark adaption and so what there is is there's like this purple disc with a slightly lighter purple disc in the middle and it's all measured in exactly the right way so that if you're sitting in the chair you're supposed to be sitting in and looking straight ahead this thing is exactly in your retina and the outer disc is or it's exactly in your fovea

2:09:29 and the outer disc is exactly the outside of the phobia uh but anyway this is and he's using special like technical lighting so that the disc is illuminated with like the precise correct wavelength of of light at the at the at the right level but this this whole thing is cooked up in such a way that um the out in the world stimulus exactly matches like your inner visual response to a dark adapted room in ways that just mess with your head trying to figure out like am i looking at something that's out there or is this just some weird [ __ ]

2:10:02 happening inside him inside of my eyes um but i mean i think that we do know enough about what's going on in the inside that we could cook up an external stimulus that just is perfectly blurry or perfectly noisy and it could be on a video a video display and someone that somehow doesn't have visual noise would be in the first the same uh low level sensory state that you are in with your your noisy your noisy eyeballs um that's really cool i want to see that

2:10:37 piece of art um i'm not sure i'm not sure that that would work because if assuming the visual noise comes from noise in the actual activities of neurons somewhere in the intermediate i mean you could cook up i see what you're saying you could cook up a external stimulus that if perceived without noise right give you the same experience right um i don't know i i think i could object to that but

2:11:08 i don't know that i want to because you know um i have a feeling that ned block and some some people have written about this back in the back in the 90s during the the representationalism wars about consciousness you know so dretski thai and block and and so there's a lot of discussions about whether blurriness indicated the existence of mental paint or whether a pure externalist representationalist could give an account of blurriness so i feel like there's a lot of

2:11:40 literature waiting waiting for you if you want to explore the blurry or the noise solution okay thanks i mean i think it's i think your higher order solution is kind of cool so miner richards who's you yours pete i'll peep i'm sure richards is fine too it's a team effort yeah social holdings

2:12:14 that's that's good because um i i mean we're really going over overtime i know someone's trying to drag you up back into the real world of alex so i want to i want to be sure to to let you go but i do want to ask if there's anything which you're hoping would come up which hasn't come up in the course of this discussion um the answer can be no i just like to check yeah yeah i'm just thinking about it um uh we didn't talk about death nearly

2:12:54 enough but we can do that we can do that next time okay so you're saying there's going to be a next time i love it um say death yeah all right let's talk about death next time that could use some talking about okay cool yeah let's get back together and do that again i mean normally in the old days we would now go into the other room and have a jam session um right but so we're not going to be able to do that but then honestly most of our jam sessions were just arguing about this stuff anyway

2:13:32 we have we can't really do it in person right now we should we should uh jam over one of these um internet based jam applications yeah i'd love to do that absolutely i want to give a big shout out to um a youtube channel called two minutes to to late night do you guys know about this with garcinia hall no no there's this guy who wears like uh corpse paint and then he gets all sorts of musicians um and some of the musicians are like relatively big name metal musicians so like people from cave-in and uh high on fire

2:14:06 um and then they will do like uh death metal or black metal covers of like guns and roses or boston and but they're all each individual is in their apartment or whatever and then anyway it's it's uh two minutes to late night um okay they're pretty pretty great really check that out yeah okay i don't know what software they're using to i mean i don't i don't know if zoom would allow for a decent musical performance but i know they're musicians that are doing

2:14:40 they're doing collaborative stuff in the group alex hook us up with something well we'll definitely all right do you know some software that we can use let's get together and jam but on that note let's go ahead and let let uh alex be free uh i want to thank you for joining us alex and pete as well both of you guys it was a great discussion alex i think your work is really cool i think we barely scratched the surface of the stuff you're doing but i think i learned a lot from reading it's challenging it's thoughtful it's uh it's interesting so i highly recommend people check out what you're doing thanks for stopping by thanks for having me on this on this

2:15:12 program i love these kinds of you know long-form discussion things so i'm really happy to do this with you guys cool and well next time death uh so just hang out for a second i'll say bye off the air but i'll go ahead and end the stream all right peace peace out all right we're clear guys