Latest / Mahon McCann / Colin Deyoung & Adam Safron - Integrating CB5T and The Free Energy Principle

Transcript

Colin Deyoung & Adam Safron - Integrating CB5T and The Free Energy Principle · Mahon McCann

0:00
SPEAKER_02: So the goal basically would be to better understand both people and also understand
0:07
SPEAKER_02: what could be done in AI on what time horizons and could you eventually make
0:12
SPEAKER_02: systems like us and would you want to?
0:21
SPEAKER_00: Colin and Adam,
0:22
SPEAKER_00: thank you very much for joining me today to discuss your integration of Cybernetic
0:27
SPEAKER_00: Big Five,
0:28
SPEAKER_00: the free energy principle,
0:29
SPEAKER_00: active inference framework.
0:31
SPEAKER_00: I'm going to include the paper in the description,
0:33
SPEAKER_00: encourage people to check it out,
0:35
SPEAKER_00: but we're kind of going to go through it a bit today.
0:37
SPEAKER_00: And just,
0:37
SPEAKER_00: I suppose I'd ask you to start off with maybe just to introduce,
0:40
SPEAKER_00: if you want to introduce Adam,
0:42
SPEAKER_00: active inference,
0:42
SPEAKER_00: and then Colin,
0:43
SPEAKER_00: tell us a little bit about Cybernetic Big Five,
0:45
SPEAKER_00: and then we'll jump into the why we're mixing them and what opportunities you see there.
0:50
SPEAKER_00: Absolutely.
0:52
SPEAKER_02: So Active Inference is a framework for understanding agency and intelligence or
1:06
SPEAKER_02: purposefulness of complex adaptive systems developed by
1:12
SPEAKER_02: Largely by the – or Carl Fristen would be the leading figure in this.
1:16
SPEAKER_02: But there's a lot of precedence for this way of thinking in machine learning and cybernetics.
1:22
SPEAKER_02: The original kind of introduction of the ideas was something called the free energy
1:27
SPEAKER_02: principle in terms of taking an analogy from thermodynamics that I guess –
1:38
SPEAKER_02: Diane and Hinton,
1:39
SPEAKER_02: actually,
1:40
SPEAKER_02: maybe some of the origin of the idea with the Helmholtz machine,
1:43
SPEAKER_02: but this idea of free energy,
1:46
SPEAKER_02: you have an entropy term and an enthalpy term.
1:50
SPEAKER_02: The enthalpy term,
1:51
SPEAKER_02: what he thought of as the work capacity of the system,
1:53
SPEAKER_02: and the entropy term is your organized energy that you use for work,
1:57
SPEAKER_02: and then your disorganized energy that will spill out all over the place and create
2:01
SPEAKER_02: friction and not be usable and just lost.
2:04
SPEAKER_02: And then your free energy is how much energy you have left for work.
2:08
SPEAKER_02: In this case,
2:08
SPEAKER_02: this was used as an objective function for training these types of machine learning
2:16
SPEAKER_02: architectures to basically say their empathy term is how well do you fit your data?
2:21
SPEAKER_02: But then you are penalized according to basically how complex your modeling is.
2:30
SPEAKER_02: So it's basically saying you've got to fit it to keep it simple.
2:33
SPEAKER_02: So it's kind of like Occam's razor built into an objective function.
2:37
SPEAKER_02: And so the idea of the free energy principle was that you can think of a broad
2:40
SPEAKER_02: range of systems based on the good regulator theorem in terms of in order to govern
2:45
SPEAKER_02: a system and make it temporarily seem to cheat the second law of thermodynamics
2:54
SPEAKER_02: while serving it in everything it does.
2:56
SPEAKER_02: But in order to do this trick of not just becoming maximally probable and a maximum entropy distribution,
3:06
SPEAKER_02: you're doing something kind of strange.
3:08
SPEAKER_02: Some kind of intelligence might be there.
3:09
SPEAKER_02: You need some sort of like to achieve any goals, hit any mark.
3:12
SPEAKER_02: It needs to be some kind of ability to model.
3:15
SPEAKER_02: And so the idea is like using the free energy principle as considering a broad
3:20
SPEAKER_02: range of systems as basically good models and then thinking of them and the
3:25
SPEAKER_02: goodness of their modeling and the ideal of the modeling
3:28
SPEAKER_02: the idea of being like, how well are they minimizing their informational or variational free energy?
3:34
SPEAKER_02: So the free energy principle is an analogy from thermodynamics for how good you're
3:40
SPEAKER_02: doing at modeling based on this intuition about Occam's razor,
3:43
SPEAKER_02: which is an important one.
3:45
SPEAKER_02: And then
3:46
SPEAKER_02: It's hard to think of a better one.
3:48
SPEAKER_02: And then applying it to systems,
3:51
SPEAKER_02: understanding them as doing modeling in their ability to achieve particular goals
3:57
SPEAKER_02: and hold themselves together,
3:59
SPEAKER_02: even.
3:59
SPEAKER_02: The goal of even being a self-perpetuating, auto-poetic cybernetic system.
4:04
SPEAKER_02: So now active inference comes in,
4:06
SPEAKER_02: in terms of actually trying to extend this to something like autonomy,
4:12
SPEAKER_02: agency,
4:12
SPEAKER_02: if you will.
4:13
SPEAKER_02: I personally think they are...
4:16
SPEAKER_02: I get a little nervous about the way they use terms.
4:18
SPEAKER_02: I think we need more words instead of arguing over single words.
4:26
SPEAKER_02: Now the idea is, OK, you want to achieve goals in the future.
4:29
SPEAKER_02: And so active inference is a normative account of how systems derived from these same principles
4:39
SPEAKER_02: engage in you know action selection and updating of their internal working models
4:44
SPEAKER_02: or policy selection machine learning terms and updating of your priors in bayesian
4:51
SPEAKER_02: terms in order to be good free energy how you minimize your free energy how you
4:57
SPEAKER_02: manage to achieve goals and hit marks and it tends to have
5:02
SPEAKER_02: These systems,
5:03
SPEAKER_02: when you jiggle around the math,
5:05
SPEAKER_02: and it's original for energy function considered in the future,
5:08
SPEAKER_02: where you're modeling yourself achieving goals in the future,
5:13
SPEAKER_02: You then end up getting these two fundamental sources of value of epistemic and pragmatic value.
5:20
SPEAKER_02: And so this pragmatic value is like,
5:21
SPEAKER_02: how well are you able to basically achieve a particular end in the world?
5:27
SPEAKER_02: And the epistemic value is how well are you able to learn as you do it?
5:31
SPEAKER_02: And so...
5:33
SPEAKER_02: Um,
5:33
SPEAKER_02: so yeah,
5:34
SPEAKER_02: so the strategy pencil on its own kind of as a count of intelligence runs into all
5:39
SPEAKER_02: sorts of problems.
5:39
SPEAKER_02: One of them being something called like the dark room problem.
5:41
SPEAKER_02: Like why don't you just like lock yourself in a dark room and like
5:45
SPEAKER_02: huddle and just like that's a good way of like you know minimizing your your
5:50
SPEAKER_02: surprise um and then this is often also framed in terms of like systems are trying
5:56
SPEAKER_02: to minimize their overall prediction error they're constantly predicting themselves
6:00
SPEAKER_02: in particular states or goal states and then they are both updating internal models
6:06
SPEAKER_02: uh perceptual inference
6:09
SPEAKER_02: In order to better fit or active inference, they're updating the world through action.
6:13
SPEAKER_02: And it kind of like Chuck Norris style,
6:14
SPEAKER_02: like he does push,
6:16
SPEAKER_02: does he do pushups or does he push the world away?
6:18
SPEAKER_02: It's like, you like, sorry.
6:20
SPEAKER_03: So it's like,
6:22
SPEAKER_02: or in an active inference way,
6:23
SPEAKER_02: you like,
6:23
SPEAKER_02: you minimize the prediction error in an inactive way by changing the world.
6:28
SPEAKER_02: So,
6:29
SPEAKER_02: minimizing pictures so that their presence,
6:31
SPEAKER_02: and this is not this well,
6:33
SPEAKER_02: this has taken a particularly,
6:35
SPEAKER_02: I think,
6:35
SPEAKER_02: elegant,
6:37
SPEAKER_02: although sometimes difficult to grok form and the work work coming from a Carl
6:41
SPEAKER_02: Frist in his community.
6:42
SPEAKER_02: um within machine learning much of this has precedence in the work of like um uh
6:47
SPEAKER_02: lots of people like a like schmidt huber like in the early 90s was thinking in
6:51
SPEAKER_02: these terms um perceptual control theory um you know even older um a lot of this
6:56
SPEAKER_02: was in like discussion cybernetics um which problems about reversing denial but
7:01
SPEAKER_02: like and also perceptual control theory um but you know there's a lot of it there
7:04
SPEAKER_02: in terms of to achieve a goal since you're like you're like
7:07
SPEAKER_02: You're trying to adjust the world to attain a particular perceptual state,
7:12
SPEAKER_02: which is the states you think you ought to be in.
7:14
SPEAKER_02: And this is very much like a lot of the friendship principles.
7:17
SPEAKER_02: It's in there.
7:19
SPEAKER_02: And then within mainstream machine learning,
7:22
SPEAKER_02: the ideas,
7:22
SPEAKER_02: the history going back to Schoenhofer of the act of learning,
7:27
SPEAKER_02: where the system is basically trying to adjust its behavior to learn better.
7:31
SPEAKER_02: And planning is inference,
7:34
SPEAKER_02: where a system is basically,
7:38
SPEAKER_02: it does things,
7:39
SPEAKER_02: it finds a thing that works,
7:41
SPEAKER_02: and then it figures out,
7:41
SPEAKER_02: how did I do the thing?
7:43
SPEAKER_02: And so it's a kind of system that comes from something kind of like,
7:47
SPEAKER_02: like even like motor battling to eventually something like skillful action and then eventually planning.
7:52
SPEAKER_02: So a lot of these ideas are, you know, they're not unique to clean energy principle, active inference.
7:59
SPEAKER_02: They just have a kind of a package for both better and worse in terms of like
8:04
SPEAKER_02: there's an elegance,
8:05
SPEAKER_02: but also I would say sometimes a more engagement with mainstream machine learning
8:14
SPEAKER_02: and other fields
8:16
SPEAKER_02: formulations would be salubrious.
8:18
SPEAKER_02: So that's my encapsulation of active inference.
8:22
SPEAKER_01: So, yeah, I mean, so cybernetics is older than the
8:30
SPEAKER_01: but I would generally see the latter as a subset of the former.
8:38
SPEAKER_01: So cybernetics was a term that was coined by Norbert Wiener back in the late 1990s.
8:44
SPEAKER_01: He was thinking about the similarities between artificial control systems and natural control systems.
8:53
SPEAKER_01: And so the way I would describe tyrannics is that it's the study of the principles
8:57
SPEAKER_01: of goal-directing systems that self-regulate via feedback processes.
9:03
SPEAKER_01: And so when you think about a cybernetic system,
9:06
SPEAKER_01: at a minimum,
9:06
SPEAKER_01: what you have is some kind of a representation of a goal within the system.
9:11
SPEAKER_01: It's not like a symbolic representation, so we don't need to get into the representationalism debate yet.
9:17
SPEAKER_01: This is just some kind of physical variable instantiated in the system that the
9:21
SPEAKER_01: system acts in such a way as to bring into a particular range or toward a
9:25
SPEAKER_01: particular value.
9:28
SPEAKER_01: not serving just a goal.
9:30
SPEAKER_01: And then there has to be some kind of feedback process that allows the system to
9:33
SPEAKER_01: compare its current state with that goal state and to close the distance between
9:39
SPEAKER_01: those two things.
9:39
SPEAKER_01: So one way to think about a cybernetic system is that at a minimum,
9:42
SPEAKER_01: it has to have some kind of a goal representation,
9:45
SPEAKER_01: some kind of a way of representing or monitoring the present state,
9:49
SPEAKER_01: and then some kind of operator or some kind of capacity for action that can
9:56
SPEAKER_01: transform the present state
9:58
SPEAKER_01: into the goal state.
10:00
SPEAKER_01: And when I first started talking to Adam about the pre-energy principle and actual inference,
10:07
SPEAKER_01: it really helped me to understand the way in which this von Helmholtz idea of the
10:13
SPEAKER_01: pre-energy incentive engine for prediction was really
10:19
SPEAKER_01: universal in the sense that it's not only about the way in which the brain predicts
10:22
SPEAKER_01: its inputs,
10:24
SPEAKER_01: which is sort of,
10:25
SPEAKER_01: I think,
10:25
SPEAKER_01: the way we classically tend to think about it and the way that the free energy
10:28
SPEAKER_01: principle was originally focused,
10:31
SPEAKER_01: but it's also about the way in which the brain predicts the results of its outputs.
10:35
SPEAKER_01: And so that is the action for the goal.
10:38
SPEAKER_01: So in a sense,
10:39
SPEAKER_01: a goal is just another type of prediction,
10:42
SPEAKER_01: because it's a prediction about how the world will be after you engage in certain
10:48
SPEAKER_01: kinds of actions.
10:51
SPEAKER_01: And so,
10:53
SPEAKER_01: as I said,
10:54
SPEAKER_01: I really think of the FEP AI as just one potential approach to cybernetics and to
11:02
SPEAKER_01: modeling the way in which the brain or other cybernetic systems go about doing what
11:08
SPEAKER_01: they need to do.
11:10
SPEAKER_01: Now, the big five part comes in from personality psychology.
11:14
SPEAKER_01: That's my background as a personality psychologist primarily.
11:19
SPEAKER_01: And there,
11:21
SPEAKER_01: the big five are a set of dimensions of variation and personality that are
11:26
SPEAKER_01: important because they represent the major dimensions of co-variation among a whole
11:31
SPEAKER_01: bunch of,
11:32
SPEAKER_01: essentially all of the more specific personality traits that people can have.
11:35
SPEAKER_01: So you heard me about how you could describe people as
11:39
SPEAKER_01: You know,
11:40
SPEAKER_01: irritable or impulsive or intelligent or extroverted or whatever,
11:45
SPEAKER_01: you can come up with hundreds of different ways to describe people.
11:48
SPEAKER_01: But if you measure a large enough collection of those specific ways in which people
11:52
SPEAKER_01: differ from each other,
11:53
SPEAKER_01: and you look for patterns of co-variation,
11:56
SPEAKER_01: which traits tend to appear together with each other,
11:58
SPEAKER_01: you end up with these five dimensions that are the big five.
12:03
SPEAKER_01: That's sort of just like a basic concurrent observation and any understanding of
12:06
SPEAKER_01: people's personalities needs to understand or explain why there might be those
12:12
SPEAKER_01: particular five dimensions of variation in human behavior and human experience.
12:18
SPEAKER_01: And so cybernetic big five theory is my attempt to basically to start from these
12:25
SPEAKER_01: general principles of cybernetics because a human being is a cybernetic system and
12:31
SPEAKER_01: to try to understand the big five as variation in a set of very centrally important
12:38
SPEAKER_01: parameters for a cybernetic system.
12:42
SPEAKER_01: And, you know, specifically for the human cybernetic system, right?
12:44
SPEAKER_01: Because even I described earlier what a minimal cybernetic system might have to include.
12:49
SPEAKER_01: But of course, there are much more complex cybernetic systems, organisms,
12:54
SPEAKER_01: human beings being especially complex organisms.
12:58
SPEAKER_01: And so,
12:58
SPEAKER_01: you know,
12:59
SPEAKER_01: in addition to negative feedback,
13:00
SPEAKER_01: they employ positive feedback,
13:02
SPEAKER_01: they employ feedforward,
13:03
SPEAKER_01: they employ a whole bunch of different control processes.
13:05
SPEAKER_01: But at the core, you have this basic imperative to be able to pursue a core set of goals, some of which
13:12
SPEAKER_01: are evolved.
13:13
SPEAKER_01: And that's part of the reason,
13:14
SPEAKER_01: that's part of the answer to the dark room problem,
13:18
SPEAKER_01: that we don't just have a tendency to predict dark inputs,
13:22
SPEAKER_01: we also have an evolved set of goals and motivations that are driving the system
13:27
SPEAKER_01: from the beginning.
13:29
SPEAKER_01: And so we can think of using the big five,
13:31
SPEAKER_01: as I said,
13:32
SPEAKER_01: as essentially variation in a particular kind of control parameter.
13:36
SPEAKER_01: Extroversion, for example, seems to be a reflection of people's basic sensitivity to reward.
13:41
SPEAKER_01: So how strongly are they driven to pursue the possibility of good things and how
13:47
SPEAKER_01: strongly do they react when they get good things?
13:51
SPEAKER_01: Any kind of organism is trying to have that kind of mechanism built into it.
13:56
SPEAKER_01: But then there's something like agreeableness,
13:58
SPEAKER_01: another one of the big five that basically covers all various ways in which humans
14:02
SPEAKER_01: can be more or less altruistic and cooperative with each other.
14:05
SPEAKER_01: And that's a kind of a control parameter that you really only need in a social organism,
14:10
SPEAKER_01: one that has to coordinate its goals with the goals of other individuals in order
14:14
SPEAKER_01: to be successful.
14:17
SPEAKER_01: So basically,
14:19
SPEAKER_01: I think that probably serves as an adequate introduction and also shows the
14:22
SPEAKER_01: motivation for Adam to try to integrate these different viewpoints together because
14:29
SPEAKER_01: we've got this very highly popular and technically sophisticated approach to
14:35
SPEAKER_01: modeling the way in which cybernetic systems can
14:43
SPEAKER_01: keep track of the world and pursue their goals effectively uh and then we've got
14:47
SPEAKER_01: this kind of more uh applied effort to try to understand how this relates to human
14:52
SPEAKER_01: beings and their individual differences from each other thank you very much for
15:01
SPEAKER_00: that colin that was a really good introduction and i suppose yeah
15:05
SPEAKER_00: You kind of clarified there,
15:06
SPEAKER_00: Colin,
15:06
SPEAKER_00: in terms of why you want to integrate,
15:10
SPEAKER_00: why you think an integration between Cybernetics Big Five and Active Inference is important.
15:15
SPEAKER_00: What about yourself, Adam?
15:16
SPEAKER_00: What was your... I know you're very interested in control as well and the control architecture.
15:20
SPEAKER_00: I'm wondering, does that factor into it?
15:23
SPEAKER_00: I'm interested on the motivation for bringing these together.
15:31
SPEAKER_03: I guess it's...
15:34
SPEAKER_03: part of a extension of, I guess, an approach of... I guess you could call it maybe...
15:56
SPEAKER_02: it's kind of like related to like computational neuro phenomenology,
16:00
SPEAKER_02: or if you will,
16:01
SPEAKER_02: like Marian neuro phenomenology,
16:02
SPEAKER_02: where it's like,
16:03
SPEAKER_02: how can you understand something as complex as a person?
16:07
SPEAKER_02: And so like,
16:07
SPEAKER_02: like David Mar proposed,
16:08
SPEAKER_02: like you have these like three supervenient levels of analysis with which you can
16:12
SPEAKER_02: understand a person,
16:14
SPEAKER_02: like a functional slash computational levels,
16:16
SPEAKER_02: or like what are the abstract description of like the functions that the system is
16:21
SPEAKER_02: doing a,
16:21
SPEAKER_02: um,
16:24
SPEAKER_02: lower level implementational or mechanistic level of like,
16:26
SPEAKER_02: what are the mechanics of the system doing that lets it do it?
16:30
SPEAKER_02: And then some sort of algorithmic abstraction layer in between them that lets you bridge between them.
16:36
SPEAKER_02: And that can be populated in a number of ways.
16:39
SPEAKER_02: So for instance,
16:40
SPEAKER_02: Some people might almost try to do an end run around them,
16:45
SPEAKER_02: and the dynamic systems theory is all you need to go between implementation and function,
16:50
SPEAKER_02: or dynamic systems theory is the algorithmic layer.
16:54
SPEAKER_02: Other people might... And what I've been using is basically...
16:57
SPEAKER_02: principles of like ai and like machine learning and trying to get that as a source
17:02
SPEAKER_02: of coherent algorithmic descriptions as a kind of like language to talk about this
17:06
SPEAKER_02: abstraction layer for like a kind of multi like this multi-stack understanding i
17:11
SPEAKER_02: guess there's other ways you could do it that might be like kind of like the most
17:14
SPEAKER_02: like a human could hope to grok you know and even that's like a big lift
17:18
SPEAKER_02: and then the hope is that basically um in doing this he has like two kind of um
17:23
SPEAKER_02: things that help with understanding one is it's both separable and that while you
17:28
SPEAKER_02: have ignorance on one level you can kind of chip away at another level so it's like
17:32
SPEAKER_02: i'm just going to work to like characterize it at this like functional level
17:35
SPEAKER_02: they're they're supervenient you know the noble in their own right and like you can
17:38
SPEAKER_02: or i'm going to work at this like
17:40
SPEAKER_02: I'm going to work at the mechanistic level more.
17:43
SPEAKER_02: And yeah, I'm not sure why, but I'm just going to do more detailed data collection.
17:47
SPEAKER_02: And there's different literatures of people with different approaches.
17:49
SPEAKER_02: You can then just make progress in the case of your ignorance.
17:52
SPEAKER_02: But also you could do potentially cross-referencing,
17:55
SPEAKER_02: like filling in a multidimensional crossword puzzle and then using that to
17:59
SPEAKER_02: constrain the inference space.
18:01
SPEAKER_02: And so let's take another way to, I guess, of thinking of it.
18:04
SPEAKER_02: And so then,
18:04
SPEAKER_02: like,
18:05
SPEAKER_02: active inference would be one way of using something like the language of
18:09
SPEAKER_02: probability theory and machine learning-like concepts to basically try to get this
18:15
SPEAKER_02: full staff understanding of a mind and its different characteristics.
18:19
SPEAKER_02: And I said neurophenomenology of that, like,
18:21
SPEAKER_02: um i actually take uh as part of the the goal is to understand like a mind and i
18:26
SPEAKER_02: actually think like consciousness um but i'm non-illusionist and that like i think
18:30
SPEAKER_02: it's actually like or at the very least we need to take phenomenology seriously in
18:35
SPEAKER_02: trying to understand minds like ours it's like part of the core explanando and um
18:38
SPEAKER_02: and i think also minds and their conscious character have functional significances
18:44
SPEAKER_02: and uh also if you will uh the other i guess reason of interest is um
18:50
SPEAKER_02: I guess there's an approach that Demis Hassabis described it well in terms of the
18:56
SPEAKER_02: fruitful marriage of neuroscience and machine learning,
18:59
SPEAKER_02: where it's like you go back and forth between here's this base domain of trying to
19:06
SPEAKER_02: reverse engineer systems that do what you want to do and systems who you actually
19:11
SPEAKER_02: want to recapitulate their properties.
19:12
SPEAKER_02: Systems like us, we are remarkably generally intelligent.
19:16
SPEAKER_02: It's amazing how we can learn so much from so little.
19:20
SPEAKER_02: It's amazing what we can do.
19:22
SPEAKER_02: And only recently have we begun to have systems which can remotely approach us and
19:27
SPEAKER_02: in some ways surpass us in their prowess along these dimensions.
19:31
SPEAKER_02: And so you have this base demand of what you're trying to reverse engineer.
19:34
SPEAKER_02: And then you have basically...
19:36
SPEAKER_02: uh, you then use that to inform the, the, the building of technology.
19:40
SPEAKER_02: And so I've,
19:41
SPEAKER_02: in some ways,
19:41
SPEAKER_02: like I got into cognitive science because like two concerns or three concerns,
19:45
SPEAKER_02: uh,
19:46
SPEAKER_02: a very long freak out about free will.
19:49
SPEAKER_02: Um, that was really quite terrible.
19:51
SPEAKER_02: Um, uh, I read the singularities in the air, uh, when I came out in the year 2000 and I was like, wait,
19:57
SPEAKER_02: could this wait, could that even be possible?
19:59
SPEAKER_02: Is that insane?
20:00
SPEAKER_02: Like,
20:00
SPEAKER_02: so I tried to figure that out and,
20:02
SPEAKER_02: uh,
20:03
SPEAKER_02: uh,
20:03
SPEAKER_02: research research,
20:05
uh,
20:05
SPEAKER_02: my own,
20:06
SPEAKER_02: uh,
20:06
SPEAKER_02: mental hangups is a variety of ways.
20:08
SPEAKER_02: Uh, but, um, anyway, so in this approach, it's going back and forth.
20:11
SPEAKER_02: So then he was like, you want to, uh,
20:14
SPEAKER_02: So you want to understand how to build better technologies.
20:17
SPEAKER_02: You take inspiration,
20:18
SPEAKER_02: but then you can go backwards and say,
20:19
SPEAKER_02: okay,
20:20
SPEAKER_02: you now have this abstract description,
20:22
SPEAKER_02: and now you can better understand the space domain too.
20:24
SPEAKER_02: You can kind of go back and forth that way.
20:26
SPEAKER_02: And so the goal basically would be to better understand both people and also
20:31
SPEAKER_02: understand what could be done in AI on what time horizons,
20:39
SPEAKER_02: and could you eventually make systems
20:42
SPEAKER_02: like us, and would you want to?
20:44
SPEAKER_02: For which reasons?
20:46
SPEAKER_02: With me generally viewing a kind of long-term goal of,
20:50
SPEAKER_02: it's a separate conversation,
20:53
SPEAKER_02: but viewing this as like something,
20:54
SPEAKER_02: if you could do it,
20:55
SPEAKER_02: one might argue it's even like a moral imperative to like,
21:00
SPEAKER_02: they could be our worthy successors or at least beings that could help us flourish
21:05
SPEAKER_02: and not die if they don't kill us.
21:08
SPEAKER_01: We argue it was a moral imperative not to do it.
21:13
SPEAKER_01: Many people are arguing exactly that.
21:18
SPEAKER_01: You've opened up several cans of worms there.
21:20
SPEAKER_01: We've got AI, we've got free will.
21:23
SPEAKER_00: And also this kind of reverse engineering project,
21:28
SPEAKER_00: it seems,
21:28
SPEAKER_00: that if you could account for a person,
21:31
SPEAKER_00: in some sense,
21:32
SPEAKER_00: yes,
21:32
SPEAKER_00: you could make a person then,
21:34
SPEAKER_00: if you knew fully all of the levels that make us up,
21:38
SPEAKER_00: whether we should do that or should not.
21:39
SPEAKER_00: But it would also seem to be...
21:41
SPEAKER_00: that it's possible to use it then for therapeutic purposes on some level of being like,
21:46
SPEAKER_00: okay,
21:47
SPEAKER_01: this is like...
21:48
SPEAKER_01: As a psychologist,
21:49
SPEAKER_01: right,
21:49
SPEAKER_01: is that it's not so much...
21:51
SPEAKER_01: I mean,
21:51
SPEAKER_01: you know,
21:52
SPEAKER_01: I think the prospects for AI are fascinating,
21:55
SPEAKER_01: exciting,
21:55
SPEAKER_01: frightening,
21:56
SPEAKER_01: you know,
21:56
SPEAKER_01: all these things.
21:58
SPEAKER_01: But my primary interest is in using these kinds of approaches to better understand people.
22:04
SPEAKER_01: And that, you know, as I'm not a clinician by training, so it's not purely to...
22:11
SPEAKER_01: help people.
22:12
SPEAKER_01: It's partly just that I want to understand what people are and how they work.
22:17
SPEAKER_01: And I do think,
22:17
SPEAKER_01: however,
22:18
SPEAKER_01: that the better our understanding of that is,
22:20
SPEAKER_01: the better we're likely to be able to help people using that knowledge.
22:25
SPEAKER_00: And where do you see then the cybernetic big five fitting in with active inference?
22:31
SPEAKER_00: Like I was trying to picture,
22:33
SPEAKER_00: you know,
22:33
SPEAKER_00: because I think about generative models and that we are kind of models of our environment.
22:38
SPEAKER_00: And is a person's personality factoring in and not like is that contextual?
22:44
SPEAKER_00: Like they talk about the generative model as being like hierarchical.
22:47
SPEAKER_00: increasing levels of more abstract control being higher up the hierarchy i mean it
22:53
SPEAKER_00: is to the personality traits factor into that i'm trying to get like an image of
22:56
SPEAKER_01: yeah so personality traits are certainly contributing to the ways in which the old
23:01
SPEAKER_01: model of the world and to the uh which might all this style of operation of the
23:08
SPEAKER_01: cyber addict system or the active inference system which i
23:16
SPEAKER_01: I think in order to really explain this,
23:18
SPEAKER_01: I have to introduce this distinction that I make in psychiatric bio-theory between
23:22
SPEAKER_01: personality traits and characteristic adaptations.
23:25
SPEAKER_01: So, the personality traits are these variations in parameters of mechanisms that everyone has.
23:32
SPEAKER_01: Like, every person has the capacity to respond to some things as rewarding.
23:37
SPEAKER_01: If you didn't have that, you don't move, right?
23:41
SPEAKER_01: You're literally inert.
23:43
SPEAKER_01: Clinically, you're very depressed.
23:45
SPEAKER_01: So we all have a reward system, right?
23:47
SPEAKER_01: We all have a punishment system to try to avoid having things that we don't want to happen.
23:54
SPEAKER_01: And there are a number of these mechanisms that everybody has by virtue of our
23:59
SPEAKER_01: human nature,
24:00
SPEAKER_01: our biological nature.
24:02
SPEAKER_01: And those provide us with some certain kind of broad foundational goals, like I want to
24:10
SPEAKER_01: achieve a certain amount of,
24:12
SPEAKER_01: you know,
24:13
SPEAKER_01: feeling of reward,
24:15
SPEAKER_01: or I want to,
24:16
SPEAKER_01: or I'm,
24:17
SPEAKER_01: you know,
24:17
SPEAKER_01: curious about things,
24:18
SPEAKER_01: I have an inherent motive to expand my knowledge,
24:21
SPEAKER_01: right?
24:21
SPEAKER_01: I guess back to this distinction that Adam made between the, um,
24:27
SPEAKER_01: I can't remember the terms used.
24:29
SPEAKER_01: There's so many terms that are used for this,
24:30
SPEAKER_01: but the epistemic reward...
24:32
SPEAKER_01: The epistemic and pragmatic value.
24:35
SPEAKER_01: The epistemic reward of getting what you want and the pragmatic...
24:39
SPEAKER_01: Sorry,
24:40
SPEAKER_01: the epistemic reward of increasing your knowledge,
24:43
SPEAKER_01: just fundamentally,
24:44
SPEAKER_01: improving your model of the world,
24:47
SPEAKER_01: and the pragmatic reward of achieving the goals that you have.
24:52
SPEAKER_01: We've all got this basic machinery that allows us to
24:58
SPEAKER_01: model the world.
25:01
SPEAKER_01: carry out behaviors in order to approach our goals,
25:04
SPEAKER_01: that obviously only touches the surface of what we actually do.
25:09
SPEAKER_01: Because what we actually do is create an incredibly detailed world model and
25:14
SPEAKER_01: develop an incredibly wide range of strategies or policies for action in the world.
25:22
SPEAKER_01: And none of that is given.
25:23
SPEAKER_01: All of that has to be learned and it has to be essentially a set of adaptations to
25:29
SPEAKER_01: the world as we find it.
25:31
SPEAKER_01: And so those are the characteristic adaptations.
25:33
SPEAKER_01: And a lot of what we think about an individual person's personality,
25:36
SPEAKER_01: like their habits,
25:38
SPEAKER_01: their tendencies,
25:39
SPEAKER_01: the things they want in life,
25:41
SPEAKER_01: their ideals,
25:41
SPEAKER_01: their values,
25:43
SPEAKER_01: all these things that are specific to their individual lives and influenced heavily
25:47
SPEAKER_01: by their cultures,
25:48
SPEAKER_01: these are not things that everyone has.
25:50
SPEAKER_01: These are things that people in particular circumstances and particular cultures have.
25:54
SPEAKER_01: Whereas personality traits are things that everyone has.
25:57
SPEAKER_01: Everyone has some level of extroversion.
26:00
SPEAKER_01: But let's just give an example.
26:02
SPEAKER_01: If you're introverted in American culture, you might join a fraternity when you go to university.
26:10
SPEAKER_01: And then that might become a very strong part of your identity that you're a fraternity brother.
26:15
SPEAKER_01: You know, those are networking organizations fundamentally, right?
26:20
SPEAKER_01: That could last you the rest of your life.
26:22
SPEAKER_01: You're part of this fraternity.
26:23
SPEAKER_01: It's important to you.
26:24
SPEAKER_01: Part of your personality.
26:25
SPEAKER_01: It's part of who you are.
26:27
SPEAKER_01: But the extroversion is the underlying basic trait.
26:29
SPEAKER_01: And in this particular case,
26:30
SPEAKER_01: it's driven you to do something and to develop a particular kind of habit,
26:35
SPEAKER_01: being a fraternity member.
26:37
SPEAKER_01: And that is a characteristic adaptation.
26:41
SPEAKER_01: And so these generative models,
26:45
SPEAKER_01: right,
26:46
SPEAKER_01: the generative models are in some sense,
26:49
SPEAKER_01: like the models that the system builds are these very complex and detailed sets of
26:54
SPEAKER_01: predictions about what the world is like and how certain actions will function in
27:00
SPEAKER_01: the world.
27:01
SPEAKER_01: And
27:05
SPEAKER_01: The traits are more like the underlying processes,
27:09
SPEAKER_01: the underlying tendencies of the mechanisms that drive you toward certain policies,
27:14
SPEAKER_01: rather than being specific policies themselves,
27:16
SPEAKER_01: which are the characteristic adaptations.
27:20
SPEAKER_00: Yeah, and do you think they're contextual in a sense?
27:23
SPEAKER_00: Like they contextualize action selection?
27:26
SPEAKER_00: Like if I'm an extroverted person,
27:29
SPEAKER_00: I'll perceive or interpret situations or opportunities for me will appear in the
27:37
SPEAKER_00: form of more social or more rewarding opportunities.
27:40
SPEAKER_00: Like I'll ignore maybe an opportunity to sit down and read a book.
27:45
SPEAKER_00: Because of this tendency that I have.
27:48
SPEAKER_00: So it constrains my interpretation of situations.
27:51
SPEAKER_01: Yeah, absolutely.
27:54
SPEAKER_01: Traits constrain both your interpretations and your choice of actions.
27:59
SPEAKER_00: yes okay and so in that sense yeah it's um part of that control because i suppose
28:06
SPEAKER_00: in active inference as far as i understand it control is really contextual it's
28:11
SPEAKER_00: like the levels of the hierarchy the higher up levels contextualize the lower
28:15
SPEAKER_00: levels and that's how control is working so would would the personality traits be a
28:21
SPEAKER_00: part of that or um
28:25
SPEAKER_01: Yeah, I mean, you can model them as higher-order infill parameters, right, or metagrammets.
28:33
SPEAKER_03: Yes.
28:37
SPEAKER_03: This might not be the case.
28:38
SPEAKER_02: I'm also kind of reminded of the search for the engram, and then you conclude the engram is nowhere.
28:44
SPEAKER_02: In terms of the trace of memory in the brain,
28:47
SPEAKER_02: it's also somewhere,
28:50
SPEAKER_02: in addition to not being anywhere in particular.
28:52
SPEAKER_02: And so there's some places where it's traits that would be particularly impactful
28:56
SPEAKER_02: for traits,
28:57
SPEAKER_02: but you can't reduce them to any one part of the brain or particular process.
29:02
SPEAKER_02: So it's like...
29:05
SPEAKER_02: in a way both actually in terms of like your policies even like there might be like
29:08
SPEAKER_02: a particular like nugget of straight or cortical loops and like their sequencing
29:13
SPEAKER_02: it's like that's the thing I do and that's like that's the thing I do when that
29:16
SPEAKER_02: happens and that's the things range of things I tend to do but still even that it's
29:21
SPEAKER_02: kind of like
29:23
SPEAKER_02: it's still going to be kind of distributed as like a data structure and maybe like
29:26
SPEAKER_02: multiply realizable but there might be like more of like a concentrated like basis
29:31
SPEAKER_02: you could point to this particular thing the traces are still that they've seen
29:34
SPEAKER_02: places where like as you go deeper in the hierarchy you might say like you might
29:38
SPEAKER_02: get more like trade-ish type things in terms of like things that might end up being
29:42
SPEAKER_02: more deeper and during however maybe not like there's or other areas of brain like
29:46
SPEAKER_02: the orbital frontal cortex or something like that where it's like you have like
29:49
SPEAKER_02: complex like
29:50
SPEAKER_02: like the nature of like your valuing and the way that golf conversations,
29:54
SPEAKER_02: but like,
29:54
SPEAKER_02: it's like,
29:55
SPEAKER_02: you can't point to any one thing and say that is like the locus of the traits.
30:00
SPEAKER_02: And at the same time,
30:01
SPEAKER_02: there's a lot of things that have more or less contribution to something enduring
30:05
SPEAKER_02: and binding in the causal system unfolding.
30:08
SPEAKER_02: Yeah.
30:09
SPEAKER_01: I think that's true, but let me put a slightly complementary spin on it.
30:14
SPEAKER_01: So, in a sense, the traits are nowhere because the traits are emergent properties, right?
30:19
SPEAKER_01: They're probabilistic dispositions to act and respond in certain ways.
30:26
SPEAKER_01: But at the same time, I don't think we want to think about this only in terms of the specific traits.
30:33
SPEAKER_01: memories or values that are instantiated,
30:36
SPEAKER_01: you know,
30:36
SPEAKER_01: very broadly in these critical systems,
30:40
SPEAKER_01: because then we're starting to get into,
30:42
SPEAKER_01: like,
30:43
SPEAKER_01: as you said,
30:43
SPEAKER_01: the engram,
30:44
SPEAKER_01: right,
30:44
SPEAKER_01: which is the characteristic adaptation,
30:46
SPEAKER_01: which is the particular way the system has adapted to its past experience.
30:51
SPEAKER_01: But if we think about the traits as these underlying inbuilt control parameters,
30:56
SPEAKER_01: then we start to be able to point to slightly more specific parts of the brain that
30:59
SPEAKER_01: are associated with different specific traits.
31:02
SPEAKER_01: So,
31:03
SPEAKER_01: for example,
31:03
SPEAKER_01: if extroversion reflects our general sensitivity to reward,
31:08
SPEAKER_01: then it may emerge out of a lot of
31:11
SPEAKER_01: things about our past experience in part,
31:13
SPEAKER_01: but in part,
31:14
SPEAKER_01: it's going to be very centrally driven by the dopaminergic system and the
31:17
SPEAKER_01: opioidergic system,
31:19
SPEAKER_01: right?
31:19
SPEAKER_01: It's going to be driven by these specific neural circuits that we know are designed
31:24
SPEAKER_01: to carry out particular functions,
31:27
SPEAKER_01: right?
31:27
SPEAKER_01: And that kind of run the whole predictive machine.
31:32
SPEAKER_01: In the sense of organizing these central cybernetic factions that are required to
31:38
SPEAKER_01: carry out any of the more specifically contextualized processes.
31:46
SPEAKER_02: And my thinking just now is being actually like a...
31:50
SPEAKER_02: like a connectomic chauvinist, basically, in a way.
31:54
SPEAKER_02: For instance, as Colin was mentioning, part of the reason there are traits is because we evolved.
31:59
SPEAKER_02: We have these evolutionary priors.
32:01
SPEAKER_02: There's this multi-scale account of cybernetics.
32:07
SPEAKER_02: The systems are getting feedback of different kinds on multiple scales and being shaped,
32:12
SPEAKER_02: and shaped at the level of phylogeny.
32:15
SPEAKER_02: And so we have traits of particular kind because we need to be fitted to the world
32:19
SPEAKER_02: in a general global sense as the kind of cybernet systems we are.
32:24
SPEAKER_02: I guess you could also say,
32:26
SPEAKER_02: I think,
32:26
SPEAKER_02: more fundamentally,
32:27
SPEAKER_02: based on getting the cybernet framework,
32:29
SPEAKER_02: we have to have some computer traits like just...
32:33
SPEAKER_02: to achieve anything it's like we would not exist at all as like we would die pretty
32:38
SPEAKER_02: quickly if we did not have some sort of a dirty but but we also have traits that
32:42
SPEAKER_02: are identifiable in this way not just as kind of like universality classes for
32:47
SPEAKER_02: complex adaptive systems such as us that can be severed systems but also because we
32:52
SPEAKER_02: are the fruits of the tree of life
32:54
SPEAKER_02: and so it's like so i was being like this kind of chauvinist like colin say like
32:58
SPEAKER_02: you like the value signals of the brain and then you have things like genetics and
33:04
SPEAKER_02: epigenetics and like that um basically and that would be like a powerful source of
33:10
SPEAKER_02: um you know particularly dense causal nexuses um how dense people might debate of
33:16
SPEAKER_02: like the sources of traits so yeah i was just like
33:20
SPEAKER_02: trying to do everything we can on topics.
33:22
SPEAKER_02: But no.
33:23
SPEAKER_02: And part of the fruits,
33:24
SPEAKER_02: I think,
33:25
SPEAKER_02: of active inference,
33:25
SPEAKER_02: not just active inference,
33:26
SPEAKER_02: though,
33:27
SPEAKER_02: is it does have this multi-level account where it's considering in a single
33:32
SPEAKER_02: intellectual framework,
33:33
SPEAKER_02: which is gnarly,
33:34
SPEAKER_02: but elegant in its own way.
33:37
SPEAKER_02: You're considering
33:38
SPEAKER_02: it's not unique to active interference but they've been trying to do it of like
33:42
SPEAKER_02: you're considering phylogeny and ontogeny and mechanism and the causal cascades
33:47
SPEAKER_02: that are basically bringing them all together this unfolding of evolutionary
33:51
SPEAKER_00: development so which seems very relevant to your work colin in terms of like the
33:56
SPEAKER_00: neurobiological foundations of personality but then also like measuring the
34:00
SPEAKER_00: phenotype and how it expresses in people um does that dovetail with your own work
34:05
SPEAKER_00: that kind of attempt to bring together the levels
34:08
SPEAKER_01: Yeah, absolutely.
34:09
SPEAKER_01: I mean, I, I've always had a very synthetic interest in these things.
34:13
SPEAKER_01: And so I do pursue a lot of neuroscience as well as the more traditional kind of
34:18
SPEAKER_01: psychometric artists to personality.
34:21
SPEAKER_01: Um,
34:22
SPEAKER_01: You know, Adam was talking about these things as constraints.
34:26
SPEAKER_01: You can think of them as sort of like causal bottlenecks, right?
34:30
SPEAKER_01: I mean,
34:30
SPEAKER_01: one of the things that's coming out of genetic research on personality and psychology,
34:34
SPEAKER_01: other individual differences,
34:35
SPEAKER_01: is just the fact that the underlying genetic and biological background that sets
34:41
SPEAKER_01: the levels of these different traits is incredibly complex.
34:45
SPEAKER_01: To the point where some people are essentially arguing there is no genetic basis of
34:49
SPEAKER_01: a specific trait because any particular genetic variation is influencing multiple traits.
34:55
SPEAKER_01: Each trait has thousands of genetic variations that are influencing its levels.
35:03
SPEAKER_01: But I think that the way around that,
35:05
SPEAKER_01: you know,
35:06
SPEAKER_01: and that seems probably to this idea of like,
35:07
SPEAKER_01: you know,
35:08
SPEAKER_01: the brain is sort of undifferentiated mass.
35:10
SPEAKER_01: The brain is not an undifferentiated mass, but it has very distinct mechanisms and circuits within it.
35:16
SPEAKER_01: And I think you have this very complex causal background and,
35:22
SPEAKER_01: you know,
35:23
SPEAKER_01: very complex causal output,
35:24
SPEAKER_01: but it does get funneled through these particular causal bottlenecks.
35:27
SPEAKER_01: such as the adult and minority system for incentive reward.
35:31
SPEAKER_01: And so I think...
35:37
SPEAKER_01: Yeah,
35:38
SPEAKER_01: just to rig off another thing that Adam was saying there and using this word constraints,
35:43
SPEAKER_01: and you brought it up too,
35:44
SPEAKER_01: Lauren,
35:44
SPEAKER_01: this idea that traits impose constraints on people.
35:47
SPEAKER_01: And this is reminding me now of something we talked about earlier,
35:50
SPEAKER_01: Adam,
35:51
SPEAKER_01: about the addition to cybernetic group flag theory made by John Berecki and his collaborator,
36:01
SPEAKER_01: Gary Ovanissian.
36:04
SPEAKER_01: Gary was the lead on this work.
36:06
SPEAKER_01: that I really liked because they were essentially
36:12
SPEAKER_01: thinking about these dynamics in terms of the fundamental constraints that traits
36:18
SPEAKER_01: impose and what are the basic constraints that any organism needs to function.
36:22
SPEAKER_01: And they're talking about this distinction between enabling constraints and selective constraints.
36:27
SPEAKER_01: And so I think a lot of the times when we're thinking about cybernetic systems or
36:33
SPEAKER_01: inference systems,
36:34
SPEAKER_01: we're thinking more about the selective constraints.
36:37
SPEAKER_01: We're thinking about how can the system pursue its goals, whether those are
36:42
SPEAKER_01: The original kind of,
36:43
SPEAKER_01: you know,
36:44
SPEAKER_01: inbuilt or evolved goals or the ones that are adopted as a consequence over a lifetime.
36:51
SPEAKER_01: But if you focus more on that question of how many things to adopt and then how do
36:56
SPEAKER_01: they change,
36:56
SPEAKER_01: then you get to this idea of the enabling constraints,
37:00
SPEAKER_01: right?
37:00
SPEAKER_01: It's like, which almost seems like an oxymoron, right?
37:02
SPEAKER_01: Like an enabling constraint.
37:05
SPEAKER_01: But there are some ways in which the functioning of a system are constrained that
37:09
SPEAKER_01: actually allows it to expand its capacity and to adapt to novel information.
37:17
SPEAKER_01: And so, you know, the cybernetic group tie-through is named after the Big Five because those are the most
37:24
SPEAKER_01: well-known and well-established dimensions of personality,
37:27
SPEAKER_01: but personality dimensions are structured hierarchically.
37:30
SPEAKER_01: So you can break the big five down into smaller pieces,
37:33
SPEAKER_01: and you can also,
37:34
SPEAKER_01: even if it turns out,
37:35
SPEAKER_01: see that the big five have their own set of correlations that reveal these two
37:38
SPEAKER_01: higher order factors
37:41
SPEAKER_01: which I've labeled as stability and plasticity.
37:44
SPEAKER_01: And Gary and John linked those to this idea of enabling constraints and selective constraints.
37:51
SPEAKER_01: So stability is selective, right?
37:52
SPEAKER_01: It's like you need to do what you need to do in order to pursue the goals that are already established.
38:00
SPEAKER_01: Whereas plasticity is about enabling.
38:04
SPEAKER_01: It's about, well, you need to be able to expand and to adapt and to explore new things.
38:08
SPEAKER_01: And that also ties back with this idea of systemic goals, right?
38:12
SPEAKER_01: You explore the world.
38:15
SPEAKER_01: It's just a little broader, right?
38:16
SPEAKER_01: Because you explore the world not only to build a better world model.
38:19
SPEAKER_01: That's part of why we just have this general curiosity,
38:22
SPEAKER_01: because the better your world model is,
38:23
SPEAKER_01: the better you're able to control your behavior and your outcomes in the world.
38:28
SPEAKER_01: But it's also about potentially discovering new goals.
38:31
SPEAKER_01: And then humans are particularly flexible in terms of their goals.
38:37
SPEAKER_01: And,
38:38
SPEAKER_01: you know,
38:38
SPEAKER_01: some of the most important mechanisms have a particular set of evolved goals,
38:41
SPEAKER_01: and they may identify new goals that are very,
38:44
SPEAKER_01: you know,
38:44
SPEAKER_01: specific instances of those evolved goals over time.
38:47
SPEAKER_01: Like, where can I get the best food, right?
38:50
SPEAKER_01: Well, that might be, you might learn that from your environment.
38:52
SPEAKER_01: So in that sense, then the location of the best food is a new goal.
38:56
SPEAKER_01: But it's still very much within this relatively concrete evolved goal of finding adequate nutrition.
39:04
SPEAKER_03: Yeah.
39:05
SPEAKER_01: You know, human beings have this capacity for much more abstract roles.
39:08
SPEAKER_01: Like,
39:09
SPEAKER_01: you know,
39:09
SPEAKER_01: I remember this example that we came up with once with Scott Barry Kaufman,
39:16
SPEAKER_01: an old friend of mine,
39:16
SPEAKER_01: where we talked about,
39:17
SPEAKER_01: like,
39:18
SPEAKER_01: somebody wants to be the skee-ball champion of the world.
39:20
SPEAKER_01: I mean, the Olympics are going on right now, right?
39:22
SPEAKER_01: So we can talk about,
39:23
SPEAKER_01: like,
39:23
SPEAKER_01: why does somebody decide that they want to be,
39:25
SPEAKER_01: like,
39:26
SPEAKER_01: the world champion pole vaulter?
39:28
SPEAKER_01: What, you know, what really good is that?
39:31
SPEAKER_01: It's certainly a long way from, you know, a rat figuring out where the best resources are.
39:37
SPEAKER_01: And of course,
39:38
SPEAKER_01: if you can tie that back into the basic evolved goals like drive for status and
39:45
SPEAKER_01: status leads to resources,
39:47
SPEAKER_01: there are obvious ways in which even our most
39:50
SPEAKER_01: to peculiar and specific out there goals as human beings are constrained or tied to
39:55
SPEAKER_01: these original goals.
39:57
SPEAKER_01: But the fact remains that we live in this incredibly free environment.
40:03
SPEAKER_01: I mean,
40:03
SPEAKER_01: some of us live in more free environments than others because of socio-political
40:07
SPEAKER_01: constraints and their type of constraints.
40:10
SPEAKER_01: But, you know, we have to be able to navigate all of this freedom.
40:15
SPEAKER_01: And so that's where you get this interplay of the traits that are related to
40:22
SPEAKER_01: stabilizing you and allowing you to continue to pursue your goals.
40:25
SPEAKER_01: Other traits that are focused on essentially allowing you to expand your horizons,
40:31
SPEAKER_01: expand your capabilities,
40:33
SPEAKER_01: expand your knowledge,
40:35
SPEAKER_03: and expand your goals.
40:42
SPEAKER_03: just to jump in on that,
40:43
SPEAKER_00: Collins,
40:43
SPEAKER_00: I'm really,
40:44
SPEAKER_00: I was actually going to bring up that paper that John wrote,
40:46
SPEAKER_00: uh,
40:46
SPEAKER_00: with Gary.
40:47
SPEAKER_00: Um, I do, is it not an activist big five theory?
40:50
SPEAKER_00: I think,
40:51
SPEAKER_00: um,
40:51
SPEAKER_01: yeah,
40:51
SPEAKER_01: so they're taking,
40:53
SPEAKER_01: uh,
40:53
SPEAKER_01: you know,
40:53
SPEAKER_01: they're taking cybernetics theory and playing more of an activist in on it,
40:57
SPEAKER_01: but you know,
40:58
SPEAKER_01: uh,
41:01
SPEAKER_01: And they kind of set it up as a contrast.
41:02
SPEAKER_01: I don't see it as a contrast.
41:04
SPEAKER_01: I think that my cybernetic approach is already deeply inactivist, right?
41:10
SPEAKER_01: Because it focuses on the... There is this physical system that has these particular needs, right?
41:18
SPEAKER_01: And so everything is built around how the system pursues its needs.
41:24
SPEAKER_01: You can talk about it from a third-person perspective, and maybe that's something more like cybernetics.
41:28
SPEAKER_01: You can talk about it more from a first-person perspective.
41:30
SPEAKER_01: phenomenological perspective,
41:31
SPEAKER_01: and maybe that's something that's more like inactivism,
41:35
SPEAKER_01: but I really see them as fundamentally compatible.
41:39
SPEAKER_01: And,
41:39
SPEAKER_01: you know,
41:40
SPEAKER_01: I always keep in mind that personality psychology is fundamentally phenomenological
41:45
SPEAKER_01: in its nature mode of measurement,
41:46
SPEAKER_01: which is that we ask people to describe their experience,
41:49
SPEAKER_01: right?
41:49
SPEAKER_01: It may be
41:50
SPEAKER_01: Maybe that you're just reading it,
41:52
SPEAKER_01: you know,
41:52
SPEAKER_01: you see,
41:52
SPEAKER_01: like,
41:53
SPEAKER_01: I am an extroverted person or something like that,
41:55
SPEAKER_01: and you rate it on a scale of one to five.
41:57
SPEAKER_01: Nonetheless, that is at some basic level of phenomenology, right?
42:01
SPEAKER_01: You're saying,
42:02
SPEAKER_01: this is what my experience of myself is like,
42:05
SPEAKER_01: this is what I believe I am like as an eating person.
42:10
SPEAKER_00: Yes.
42:12
SPEAKER_00: Yeah.
42:12
SPEAKER_00: My understanding of that as well with John and Gary's paper is that they tied into
42:17
SPEAKER_00: relevance realization that these kind of stability versus plasticity or personality traits,
42:22
SPEAKER_00: um,
42:24
SPEAKER_00: I don't know if bias is the right word,
42:25
SPEAKER_00: but frame a person's relevance realization,
42:28
SPEAKER_00: like what they're going to identify as relevant in different circumstances.
42:31
SPEAKER_00: That ties into like distributed cognition between like different,
42:35
SPEAKER_00: they even go as far as populations,
42:36
SPEAKER_00: I think,
42:37
SPEAKER_00: that are skewed one particular way versus another.
42:40
SPEAKER_00: And then that that's an argument for, you know, distributed cognition and problem solving.
42:45
SPEAKER_00: They're like one person's personality,
42:46
SPEAKER_00: they see it one way,
42:47
SPEAKER_00: another person's personality sees it differently.
42:49
SPEAKER_00: And then through the communication, distributed cognition, you're able to
42:53
SPEAKER_00: have more flexible general adaptation sure yeah i mean and i definitely see
42:58
SPEAKER_01: stability and plasticity as these fundamentally complementary uh modes or uh or
43:05
SPEAKER_02: drives or knees it's still cybernetics though yeah i mean like extended homeostasis
43:11
SPEAKER_02: into like sociocultural homeostasis we're still talking homeostasis and cybernetics
43:16
SPEAKER_02: like
43:18
SPEAKER_01: it's something that i still haven't settled uh for myself um the degree to which
43:25
SPEAKER_01: you can have distributed cybernetic systems as opposed to just cooperations among
43:31
SPEAKER_01: um you know individual cybernetic systems but um you know i don't know this might
43:36
SPEAKER_01: be more detailed than we want to get into in terms of relevance realization but at
43:39
SPEAKER_01: some level i think that uh
43:43
SPEAKER_01: John's idea about the solution to relevance realization as residing in these
43:47
SPEAKER_01: opponent processes doesn't get him far enough.
43:50
SPEAKER_01: It actually doesn't get him to solve the frame problem,
43:52
SPEAKER_01: which is what relevance realization is trying to do,
43:54
SPEAKER_01: basically.
43:55
SPEAKER_01: sets up the conditions by which the frames problem can be solved by a system or an
44:00
SPEAKER_01: organism that lives in a complex and changing environment.
44:04
SPEAKER_01: But if you think about,
44:05
SPEAKER_01: the frame problem was actually supposed to answer the question of how,
44:08
SPEAKER_01: in a specific moment,
44:11
SPEAKER_01: a system,
44:12
SPEAKER_01: either a robot or an organism,
44:16
SPEAKER_01: makes an adequate decision to pursue its goals given the essentially infinite sea
44:20
SPEAKER_01: of information that it is surrounded by.
44:24
SPEAKER_01: and the only way that you can actually solve those kinds of problems is to have a
44:28
SPEAKER_01: an adequately detailed world model right you have to have a world model that is
44:34
SPEAKER_01: adequate both in terms of like correlational expectations and predictions and
44:39
SPEAKER_01: causal expectations and predictions um you know like the famous example then it's
44:44
SPEAKER_01: famous example of the robot that needs to uh
44:48
SPEAKER_01: You know,
44:48
SPEAKER_01: it needs a new battery and it needs to go and get the battery that's on a wagon and
44:53
SPEAKER_01: roam somewhere.
44:53
SPEAKER_01: But there's a bomb on the wagon.
44:55
SPEAKER_01: Right.
44:55
SPEAKER_01: And it doesn't realize that the person doesn't realize that the bomb is a bomb.
45:00
SPEAKER_01: And so it blows up and then it gets hung up and trying to figure out,
45:05
SPEAKER_01: like,
45:05
SPEAKER_01: what are all the things that are relevant to this situation?
45:08
SPEAKER_01: right and and so it still blows up uh you know because it's designed to say well it
45:13
SPEAKER_01: can't just ignore the other things that are in the room it has to actually think
45:16
SPEAKER_01: about what else is there too and then the question is well how does a system
45:19
SPEAKER_01: adequately think about what else is there too and not just get overwhelmed by all
45:24
SPEAKER_01: the uh potentially extraneous details that are there and um you know and
45:32
SPEAKER_01: Merely saying that,
45:33
SPEAKER_01: well,
45:33
SPEAKER_01: the robot has to be able to explore and be stable doesn't actually...
45:36
SPEAKER_01: I mean,
45:37
SPEAKER_01: those are preconditions,
45:38
SPEAKER_01: presumably,
45:38
SPEAKER_01: to having it solve the frame problem,
45:42
SPEAKER_01: but they don't get it all the way there.
45:45
SPEAKER_01: It doesn't get all the way there until it's actually built a model of the world
45:48
SPEAKER_01: that is sufficiently detailed to be able to quickly recognize that a bomb is a bomb
45:52
SPEAKER_01: and that there are problems with bringing it along with you when you're getting the
45:56
SPEAKER_01: things that you want.
45:57
SPEAKER_01: And,
45:58
SPEAKER_01: you know,
45:58
SPEAKER_01: so the robot can't solve the frame problem,
46:01
SPEAKER_01: not because it lacks,
46:02
SPEAKER_01: you know,
46:03
SPEAKER_01: stability and plasticity or some sort of opponent processes,
46:06
SPEAKER_01: but because it hasn't learned enough yet.
46:08
SPEAKER_01: Just in the same way that a three-year-old would also get itself blown up, right?
46:11
SPEAKER_01: Because it wouldn't understand that a bomb was a bomb.
46:15
SPEAKER_01: Or,
46:15
SPEAKER_01: you know,
46:16
SPEAKER_01: in the way of three-year-olds,
46:17
SPEAKER_01: it might get distracted and,
46:19
SPEAKER_01: you know,
46:19
SPEAKER_01: start paying attention to
46:21
SPEAKER_01: some toy that's sitting in the corner of the room or something.
46:24
SPEAKER_01: So the capacity to model the world and solve the frame problem,
46:29
SPEAKER_01: like first of all,
46:30
SPEAKER_01: the frame problem is always relevant to what one's goals are.
46:33
SPEAKER_01: The goal sets the frame fundamentally.
46:36
SPEAKER_01: And so,
46:38
SPEAKER_01: you know,
46:39
SPEAKER_01: you have to basically have the capacity to build a detailed enough world model,
46:43
SPEAKER_01: and that's a set of characteristic adaptations that you have to adapt to your world
46:46
SPEAKER_01: in such a way that you can understand it adequately,
46:49
SPEAKER_01: that you can do the right thing in relation to your goals.
46:52
SPEAKER_01: So merely having these kinds of sets of opponent processes that ultimately boil
46:57
SPEAKER_01: down to disability and plasticity is merely like the precondition for...
47:05
SPEAKER_01: uh, solving your friend problem.
47:06
SPEAKER_03: It's not actually the solution itself.
47:10
SPEAKER_03: What, um, a couple of comments on that question from earlier.
47:16
SPEAKER_03: So like,
47:18
SPEAKER_02: so it seems like like with like in activism,
47:20
SPEAKER_02: you know,
47:21
SPEAKER_02: um,
47:22
SPEAKER_02: and so like oftentimes like,
47:24
SPEAKER_02: um,
47:24
SPEAKER_02: you know,
47:25
SPEAKER_02: like,
47:25
SPEAKER_02: like,
47:25
SPEAKER_02: or for E for EA cognition,
47:27
SPEAKER_02: like,
47:27
SPEAKER_02: you know,
47:28
SPEAKER_02: minds are,
47:29
SPEAKER_02: um,
47:30
SPEAKER_02: uh,
47:30
SPEAKER_02: you know,
47:31
SPEAKER_02: embodied,
47:31
SPEAKER_02: uh,
47:32
SPEAKER_02: extended,
47:33
SPEAKER_02: um,
47:35
SPEAKER_02: you know, interact with the environment.
47:38
SPEAKER_02: It's like,
47:40
SPEAKER_02: they're helpful,
47:41
SPEAKER_02: but really it's like,
47:43
SPEAKER_02: and also like,
47:43
SPEAKER_02: so they may be helpful in addition for like part of solving frame problems would be
47:48
SPEAKER_02: like both like distributed cognition,
47:54
SPEAKER_02: you know,
47:54
SPEAKER_02: ways in which like,
47:55
SPEAKER_02: you know,
47:57
SPEAKER_02: we're um computing between us and like some of like our working memory is like
48:02
SPEAKER_02: offloaded into us and we're taking cues from each other and like it's helpful but
48:07
SPEAKER_02: without like each individual being able to like toll its own weight as a cyber
48:10
SPEAKER_02: attack system and world modeling um you don't go as far and so so i feel like the
48:14
SPEAKER_02: inactivism cuts both ways and like it does show like additional i guess it both
48:20
SPEAKER_02: shows ways you can augment your um
48:24
SPEAKER_02: world modeling capacities or your intelligent capacities on top of the world
48:27
SPEAKER_02: modeling through extended cognition it brings your attention back to embodiment um
48:33
SPEAKER_02: it's uh
48:35
SPEAKER_02: Also, it's a good prior that basically, like Rodney Brooks said, the world is its own best model.
48:41
SPEAKER_02: And there's a sense in which it's like,
48:44
SPEAKER_02: to the extent of cryptoprocessing,
48:45
SPEAKER_02: you can minimize your prediction error as efficiently as possible,
48:51
SPEAKER_02: including things like unconscious reflex arcs,
48:53
SPEAKER_02: even things like the way your environment is structured,
48:55
SPEAKER_02: you do so.
48:56
SPEAKER_02: It's like the most effective way to minimize your cybernetic entropy.
49:00
SPEAKER_02: You try to find that it tends to be like pushing the control
49:04
SPEAKER_02: you know, as down to a low level across hierarchies as possible, even off into the world.
49:11
SPEAKER_02: And so in general,
49:11
SPEAKER_02: like,
49:11
SPEAKER_02: you know,
49:12
SPEAKER_02: do not,
49:13
SPEAKER_02: it's kind of like,
49:15
SPEAKER_02: do not,
49:16
SPEAKER_02: almost like Occam's razor,
49:17
SPEAKER_02: like,
49:18
SPEAKER_02: do not invoke representations and models unless you're forced to do so.
49:21
SPEAKER_02: Yet we are indeed forced to do so.
49:23
SPEAKER_02: And so that's part of the cutting both ways is I worry about this with radical and activism.
49:28
SPEAKER_02: And some of like,
49:29
SPEAKER_02: there's a point to the Cartesian,
49:31
SPEAKER_02: like anti-Cartesian rhetoric,
49:32
SPEAKER_02: and they're like,
49:33
SPEAKER_02: you know,
49:34
SPEAKER_02: a passive observer,
49:35
SPEAKER_02: that that doesn't work.
49:37
SPEAKER_02: But the idea that,
49:37
SPEAKER_02: for instance,
49:38
SPEAKER_02: like centralized locuses of control centered on the body based on our particular
49:44
SPEAKER_02: perspectives on the world,
49:46
SPEAKER_02: where this is a modeling system of self and world,
49:49
SPEAKER_02: and where there are things like explicit representations,
49:52
SPEAKER_02: these actually might be essential for understanding autonomous systems like us.
49:56
SPEAKER_02: And so if you're in activism, it ends up making you, there's so much to be added.
50:01
SPEAKER_02: But if you go radical in activism, I feel all is lost.
50:04
SPEAKER_02: And that's, so that's, I think there's like, so there's a sense in which like, you know, Sidi Fatih,
50:11
SPEAKER_02: A lot of it was like,
50:12
SPEAKER_02: it is already an activist,
50:13
SPEAKER_02: I'd say,
50:14
SPEAKER_02: and that like,
50:15
SPEAKER_02: you know,
50:15
SPEAKER_02: it's about like controlling a cybernetic system and embodying a cybernetic system
50:19
SPEAKER_02: and incorporating the free energy principles like,
50:20
SPEAKER_02: you know,
50:20
SPEAKER_02: and there's like evolutionary priors and there's like the way you're getting these
50:25
SPEAKER_02: causal loops and these bottlenecks,
50:28
SPEAKER_02: these,
50:28
SPEAKER_02: I forget the exact term you use,
50:30
SPEAKER_02: but the constraints,
50:31
SPEAKER_02: the source of constraints,
50:32
SPEAKER_02: they end up like unfolding both within the system and even though including
50:36
SPEAKER_02: dynamics without the system.
50:38
SPEAKER_02: So I think there's a valuable thing in re-emphasizing the inactivism,
50:41
SPEAKER_02: but I worry that if we go radical inactivists,
50:46
SPEAKER_02: we're losing something essential for
50:49
SPEAKER_02: actual agents.
50:51
SPEAKER_01: I definitely want to hear a little bit more about your take on the sort of
50:55
SPEAKER_01: Cartesian problem,
50:56
SPEAKER_01: but I think it would be really helpful first if we sort of take a step back
50:59
SPEAKER_01: historically and think about this debate between,
51:04
SPEAKER_01: you know,
51:04
SPEAKER_01: inactivism and some earlier form of,
51:07
SPEAKER_01: let's call it computationalism.
51:10
SPEAKER_01: Because, you know, in some ways I think, you know, Varela won, right?
51:14
SPEAKER_01: The inactivists won.
51:15
SPEAKER_01: But I think that
51:18
SPEAKER_01: in some sense, they're still fighting the same battle, even though they've already won.
51:23
SPEAKER_01: And that,
51:24
SPEAKER_01: I think,
51:25
SPEAKER_01: has to do with the fact that there are all these different meanings floating around
51:28
SPEAKER_01: of the term representation.
51:30
SPEAKER_01: So when you think about classic representationalism and what the kind of an
51:35
SPEAKER_01: activist perspective is pushing back on,
51:39
SPEAKER_01: it essentially treats
51:42
SPEAKER_01: human brains as if they were symbolic computers,
51:46
SPEAKER_01: that they are carrying out computational operations on symbols.
51:50
SPEAKER_01: And so if what you mean by a representation is that something has to be translated
51:56
SPEAKER_01: into a symbol string and then the symbol string gets manipulated,
51:59
SPEAKER_01: then that,
52:01
SPEAKER_01: I think,
52:01
SPEAKER_01: is...
52:03
SPEAKER_01: Fair enough.
52:05
SPEAKER_01: I think that's out the window, right?
52:06
SPEAKER_01: That's not how brains work.
52:08
SPEAKER_01: They don't do computations on syllable strings, except for the fact that they do sometimes, right?
52:14
SPEAKER_01: But that's not the foundation of cognition.
52:18
SPEAKER_01: The foundation of cognition is embodied and...
52:27
SPEAKER_01: I don't know even what the right contrast is.
52:29
SPEAKER_01: Well, I mean, to use some of Deacon's terms, it's indexical and iconic, right?
52:32
SPEAKER_01: It's about detecting correlations among things and similarities and differences among things.
52:38
SPEAKER_01: It does not have to translate everything immediately into a symbol string and then
52:42
SPEAKER_01: manipulate symbol string the way that traditional serial computers did.
52:47
SPEAKER_01: Okay.
52:48
SPEAKER_01: But yeah,
52:49
SPEAKER_01: so if that's what you mean by representation,
52:51
SPEAKER_01: fine,
52:52
SPEAKER_01: you know,
52:52
SPEAKER_01: representation is inadequate,
52:53
SPEAKER_01: but I think most of us now use representation to mean something different and we
52:57
SPEAKER_01: use it to mean something more like the capacity of a cybernetic system to basically
53:07
SPEAKER_01: build regularities within the system itself
53:11
SPEAKER_01: that in some way model or reflect or represent,
53:16
SPEAKER_01: if you will,
53:17
SPEAKER_01: some aspect either of the external world or of their own functioning that then
53:22
SPEAKER_01: allows them to control their outcomes more effectively,
53:27
SPEAKER_01: right?
53:27
SPEAKER_01: You mentioned the good regulator theorem, Adam.
53:29
SPEAKER_01: So that's a classic cybernetic theory, right?
53:31
SPEAKER_01: Yeah.
53:32
SPEAKER_01: to optimally regulate something, you need a model of that thing.
53:36
SPEAKER_01: And what is a model of the thing?
53:37
SPEAKER_01: A model of the thing is not necessarily like a set of symbolic linguistic or
53:42
SPEAKER_01: mathematical representations of the thing.
53:45
SPEAKER_01: Rather,
53:45
SPEAKER_01: it is some kind of set of physical variables that mimic or mirror a subset of the
53:54
SPEAKER_01: dynamics of the thing that is being modeled.
53:56
SPEAKER_01: And, you know, preferably...
53:58
SPEAKER_01: aspects of the dynamics that are most relevant to whatever goal the system is trying to achieve.
54:03
SPEAKER_01: And so that is
54:08
SPEAKER_01: Okay,
54:08
SPEAKER_01: so I hope that helps us to get past this kind of old debate that sometimes I see
54:14
SPEAKER_01: people like John still fighting,
54:16
SPEAKER_01: even though I don't know if they need to,
54:18
SPEAKER_01: right?
54:19
SPEAKER_01: Because you're fighting against concepts that are attached to the vocabulary that
54:24
SPEAKER_01: is no longer being used to represent those same concepts.
54:29
SPEAKER_01: But, okay, so with that kind of in mind, we're freer to talk about representation.
54:34
SPEAKER_01: Because I think you're talking something more about the idea that there is some
54:37
SPEAKER_01: kind of centralized self-process.
54:41
SPEAKER_02: So I do go that far into something that you might even call a kind of neo-Cartesianism.
54:50
SPEAKER_02: um and you know there's like cartesian there's like um aspects of cartesian
54:54
SPEAKER_02: thinking i think you know we rightly reject but i think um uh the corrective might
54:59
SPEAKER_02: have been excessive in a problematic way that hid the plots even about things like
55:04
SPEAKER_02: the central nature of embodiment as a fundamental context context of mind and data
55:08
SPEAKER_02: structure and the way in which the mind is like
55:11
SPEAKER_02: in a sensual way scaffolded off of the body, and this being essentially essential for cybernet control.
55:17
SPEAKER_02: But in terms of the issue of representation,
55:20
SPEAKER_02: one thing that could be people who are interested in this issue,
55:25
SPEAKER_02: I organized a special issue of Frontiers in Neuro-Robotics
55:30
SPEAKER_02: recently with Andy Clark and Ines Hipolito.
55:33
SPEAKER_02: And it was kind of like a benevolent adversarial collaboration with Ines being a
55:39
SPEAKER_02: little bit more radical in her inactivism,
55:41
SPEAKER_02: me being a neo-Cartesian.
55:45
SPEAKER_02: And basically, in this intro, the papers kind of take different sides of this debate and explore them.
55:51
SPEAKER_02: in the context of biologically inspired artificial intelligence.
55:56
SPEAKER_02: And in the intro,
55:56
SPEAKER_02: we kind of like described,
55:57
SPEAKER_02: like there's different ways you can think of representation ranging from a kind of like,
56:02
SPEAKER_02: you know,
56:04
SPEAKER_02: any kind of predictive information,
56:06
SPEAKER_02: like,
56:06
SPEAKER_02: you know,
56:07
SPEAKER_02: an attractor that can wiggle another attractor when they couple to,
56:10
SPEAKER_02: you know,
56:11
SPEAKER_02: something like at the limit,
56:13
SPEAKER_02: something that approaches like the,
56:15
SPEAKER_02: um what they use in like good old-fashioned cognitive science in terms of like you
56:19
SPEAKER_02: know these types of symbol systems and ways that either we can do this via like
56:25
SPEAKER_02: virtual machines in terms of like this is a thing that the systems all could do but
56:29
SPEAKER_02: possibly even ways that some of the neural dynamics might in a more direct way um
56:34
SPEAKER_02: be mappable in those terms like so for instance like a deep mind they'd have some
56:39
SPEAKER_02: like architectures they call like a
56:41
SPEAKER_02: It's like a neural Turing machine,
56:43
SPEAKER_02: where they're actually describing,
56:44
SPEAKER_02: for instance,
56:46
SPEAKER_02: a somewhat brain-inspired cognitive system that has connections elements,
56:51
SPEAKER_02: but then has a kind of sequential operation where it functions mechanistically like
56:55
SPEAKER_02: a Turing machine.
56:56
SPEAKER_02: And another kind of example of this,
56:59
SPEAKER_02: you could think of maybe even the hippocampal system and certain choke points,
57:04
SPEAKER_02: where you have basically these equilibrium points that you can move between in a
57:10
SPEAKER_02: kind of discrete fashion,
57:11
SPEAKER_02: that these can enable something like a kind of...
57:15
SPEAKER_02: symbolic the conceptual chunking and order of operations and putting things
57:20
SPEAKER_02: together into hierarchies where it's you're starting to approach something which
57:26
SPEAKER_02: we've called neuro symbolic if you were to realize it as an architecture so like I
57:32
SPEAKER_02: think the inactivists are like they're crucially right like it's like
57:36
SPEAKER_02: and like what Colin was saying,
57:37
SPEAKER_02: it's like the foundation is like in a way,
57:40
SPEAKER_02: um,
57:41
SPEAKER_02: either non-representational or a notion of representation that's more like thin and,
57:47
SPEAKER_02: and more applicable.
57:49
SPEAKER_02: But, um,
57:51
SPEAKER_02: uh, and more broadly applicable.
57:53
SPEAKER_02: And you might not even think like, is that a representation?
57:54
SPEAKER_02: I'm not so sure.
57:56
SPEAKER_02: Or,
57:57
SPEAKER_02: but,
57:57
SPEAKER_02: but I would argue that basically,
57:59
SPEAKER_02: uh,
57:59
SPEAKER_02: in terms of the brains like ours as an overall architecture,
58:03
SPEAKER_02: um,
58:04
SPEAKER_02: that for kind of like CL against,
58:07
SPEAKER_02: you know,
58:07
SPEAKER_02: maybe we're,
58:08
SPEAKER_02: we're more like,
58:08
SPEAKER_02: you know,
58:09
SPEAKER_02: talking,
58:10
SPEAKER_02: uh,
58:10
SPEAKER_02: more purely an activist, and maybe you want to squint a bit for, say, representation.
58:15
SPEAKER_02: Like, ah, I would still, like, I might want to use that language.
58:18
SPEAKER_02: I'm happy with it.
58:19
SPEAKER_02: And Talon is a cyberneticist.
58:21
SPEAKER_01: Yeah,
58:22
SPEAKER_01: what makes it a representation is precisely the kind of function that it plays in
58:25
SPEAKER_01: the cybernetic system.
58:27
SPEAKER_01: Like, that's the way that I think about it.
58:29
SPEAKER_02: Yeah, so I agree.
58:31
SPEAKER_02: And I prefer that use of like language because of like the ability for you to like
58:37
SPEAKER_02: see things like life-mind continuity,
58:40
SPEAKER_02: to understand like,
58:41
SPEAKER_02: you know,
58:43
SPEAKER_02: the same principles showing up across scales and to make comparison,
58:46
SPEAKER_02: you know.
58:47
SPEAKER_01: And in terms of vocabulary,
58:48
SPEAKER_01: I think it's easier for people to adopt the idea that there are different kinds of
58:52
SPEAKER_01: representations than to work with the idea that there are no representations at all
58:58
SPEAKER_01: or that they're only in a very specialized context.
59:03
SPEAKER_02: I think there is an important corrective to be had.
59:06
SPEAKER_02: It's like all you're doing is focusing on the mind and it's high levels of all
59:09
SPEAKER_02: character and you're just neglecting it.
59:12
SPEAKER_02: That was a problem and it still is a problem.
59:14
SPEAKER_01: That's why I say Morello won, right?
59:15
SPEAKER_01: Because I think that that whole movement did largely address that problem.
59:21
SPEAKER_01: And once that's why robotics is now having a lot of success where for a decade it's struggled miserably.
59:28
SPEAKER_02: I tend to go, you know,
59:34
SPEAKER_02: I wouldn't say for sure that this is the way it works.
59:36
SPEAKER_02: But for instance,
59:37
SPEAKER_02: in some of my attempts at doing dynamical systems handling of systems neuroscience
59:45
SPEAKER_02: in a computational lens,
59:47
SPEAKER_02: I've interpreted,
59:48
SPEAKER_02: for instance,
59:51
SPEAKER_02: cortical predictive processing as a kind of auto-encoding.
59:55
SPEAKER_02: And the idea is like,
59:56
SPEAKER_02: you're basically doing this dimensionality reduction,
59:58
SPEAKER_02: finding patterns of patterns of patterns and of increasing or reducing
1:00:02
SPEAKER_02: dimensionality with a,
1:00:03
SPEAKER_02: with a bottlenecking.
1:00:05
SPEAKER_02: Um, and, uh,
1:00:08
SPEAKER_02: information biohack where you're basically,
1:00:10
SPEAKER_02: by virtue of this,
1:00:12
SPEAKER_02: able to identify the,
1:00:15
SPEAKER_02: like,
1:00:15
SPEAKER_02: large,
1:00:16
SPEAKER_02: the salient features that could potentially be mapped onto meaningful things in the world.
1:00:21
SPEAKER_02: And it's been actually shown that,
1:00:22
SPEAKER_02: like,
1:00:22
SPEAKER_02: autoencoding actually,
1:00:23
SPEAKER_02: it's equivalent to principal component analysis in a way.
1:00:27
SPEAKER_02: So,
1:00:28
SPEAKER_02: but the idea is,
1:00:28
SPEAKER_02: like,
1:00:29
SPEAKER_02: if you get to basically the dynamics within,
1:00:32
SPEAKER_02: like,
1:00:32
SPEAKER_02: deeper areas of cortical hierarchies,
1:00:34
SPEAKER_02: and if you look at their manifold activity,
1:00:37
SPEAKER_02: in terms of the population level activity,
1:00:41
SPEAKER_02: there's a way in which this might be viewed as maybe something closer to
1:00:45
SPEAKER_02: representation proper.
1:00:46
SPEAKER_02: and rather like a more like a thick notion a little bit closer to both like um
1:00:54
SPEAKER_02: cognitive science and i would argue like a goal to actually target science in some
1:00:57
SPEAKER_02: ways and and maybe even i would suggest closer to experience in terms of reduce the
1:01:04
SPEAKER_01: just because you reduce the dimensionality doesn't mean that you have to make it a
1:01:07
SPEAKER_01: symbolic representation right doesn't mean that no doesn't mean that i think
1:01:12
SPEAKER_01: One thing that I think is interesting to think about is the possibility that rather
1:01:15
SPEAKER_01: than thinking about it in terms of like principal components analysis,
1:01:17
SPEAKER_01: factor analysis,
1:01:18
SPEAKER_01: we think of this like partially squares analysis,
1:01:20
SPEAKER_01: right?
1:01:20
SPEAKER_01: Where you're only refactorizing part of the matrix so that you maintain the
1:01:26
SPEAKER_01: essentially control of the data structure by
1:01:31
SPEAKER_01: goals, fundamentally.
1:01:33
SPEAKER_01: And I just thought of this connection because I was thinking about,
1:01:36
SPEAKER_01: you know,
1:01:36
SPEAKER_01: what Neil McNaughton would say about the way in which,
1:01:40
SPEAKER_01: you know,
1:01:41
SPEAKER_01: the reason that you get these things,
1:01:42
SPEAKER_01: the data gets compressed and compressed,
1:01:44
SPEAKER_01: basically,
1:01:44
SPEAKER_01: right?
1:01:44
SPEAKER_01: The dimensionality gets reduced until it feeds into the hippocampus,
1:01:48
SPEAKER_01: because essentially the hippocampus is the top of the cortical hierarchy.
1:01:52
SPEAKER_01: And
1:01:55
SPEAKER_01: What you're doing there,
1:01:56
SPEAKER_01: basically,
1:01:56
SPEAKER_01: according to McNaughton and,
1:01:59
SPEAKER_01: you know,
1:02:00
SPEAKER_01: originally Gray and McNaughton and Jeffrey Gray's whole life,
1:02:04
SPEAKER_01: is that you are the primary function of the hippocampus.
1:02:08
SPEAKER_01: Yes,
1:02:09
SPEAKER_01: it performs all this integration that allows you to create,
1:02:11
SPEAKER_01: you know,
1:02:11
SPEAKER_01: like maps and space and to store episodic memories.
1:02:16
SPEAKER_01: in a form where they can be retrieved with all of their elements being pulled back together.
1:02:19
SPEAKER_01: But the reason you need all that integration basically is that you have to be able
1:02:23
SPEAKER_01: to detect when different high-level goals are in conflict with each other.
1:02:28
SPEAKER_01: So what the hippocampus is doing is basically checking to say,
1:02:33
SPEAKER_01: if I do this,
1:02:33
SPEAKER_01: is it going to conflict with this other thing?
1:02:35
SPEAKER_01: But this other thing can be very abstract, right?
1:02:37
SPEAKER_01: And so you actually need to have everything compressed down to this high-level abstract dimensionality.
1:02:42
SPEAKER_01: But it also has to be compressed in such a way that what anchors the dimensions are
1:02:47
SPEAKER_01: the goals that the sensory input is actually relevant to.
1:02:52
SPEAKER_02: That is so crucial because the maintenance of the data structure to be able to buy goals.
1:02:58
SPEAKER_02: And now I feel like we've actually achieved something like agency.
1:03:02
SPEAKER_02: And without that, you don't necessarily need it to be via some sort of anatomical choke point.
1:03:12
SPEAKER_02: It happens to be a useful thing to do to varying degrees for letting the whole
1:03:17
SPEAKER_02: system turn on a dime more flexibly and have large-level things...
1:03:23
SPEAKER_02: able to be able to enslave the dynamics of classes that actually have like things
1:03:27
SPEAKER_02: like small world structure have like hubs where things are concentrated convergence
1:03:32
SPEAKER_02: divergence sounds like the master might call it and then you have this capacity and
1:03:36
SPEAKER_02: so the nature of this like and as you're doing this though you know you're a long
1:03:40
SPEAKER_02: way from the initial inputs and so you know dennis like you know there is no double
1:03:44
SPEAKER_02: transduction in the brain i don't think that's right i think there's multiple
1:03:47
SPEAKER_02: transductions and re-representations
1:03:49
SPEAKER_02: And it takes on different forms with different properties.
1:03:52
SPEAKER_02: And in some of them,
1:03:53
SPEAKER_02: it lets you basically have things like a chunking of experience in ways where it
1:03:58
SPEAKER_02: can be mapped onto things like coherent actions.
1:04:01
SPEAKER_02: And so I worry that... I feel like Dennett has a very mixed legacy.
1:04:05
SPEAKER_02: And in Activist, there's a valuable corrective, but it's excessive.
1:04:09
SPEAKER_02: And then we end up basically...
1:04:12
SPEAKER_02: In our attempt to bring together the manifest and scientific images and inactively
1:04:17
SPEAKER_02: find our place in the world,
1:04:18
SPEAKER_02: we end up losing it by basically annihilating the very thing you need for autonomous,
1:04:25
SPEAKER_02: sovereign selfhood.
1:04:30
SPEAKER_01: Right.
1:04:32
SPEAKER_01: Now we get to the crusade against Buddhism.
1:04:37
SPEAKER_01: I'm glad we finally got there.
1:04:38
SPEAKER_01: The annihilation of the self.
1:04:41
SPEAKER_01: Yeah, well, you know, the self is an illusion, but it's a necessary illusion.
1:04:44
SPEAKER_01: You need to build the model of the agent in order to control the agent effectively.
1:04:50
SPEAKER_01: That's a good regulator, isn't it?
1:04:53
SPEAKER_00: And I was hearing there,
1:04:54
SPEAKER_00: Colin,
1:04:54
SPEAKER_00: what you were saying about the hippocampus conscientiousness,
1:04:58
SPEAKER_00: like the maintenance of the goal hierarchy.
1:05:01
SPEAKER_00: We talked about that last time, Adam, the connection between conscientiousness and agency.
1:05:05
SPEAKER_00: Is that,
1:05:08
SPEAKER_00: like,
1:05:09
SPEAKER_00: do you think,
1:05:09
SPEAKER_00: because you've said that's the top of the cortical layer,
1:05:11
SPEAKER_00: is that...
1:05:13
SPEAKER_00: Yeah,
1:05:13
SPEAKER_01: well,
1:05:14
SPEAKER_01: I mean,
1:05:14
SPEAKER_01: because it's older than the cortex that massively expanded,
1:05:18
SPEAKER_01: but one way to think about that is that it massively expanded in us
1:05:22
SPEAKER_01: is the capacity to map out sensory information,
1:05:26
SPEAKER_01: that's the posterior of the cortex,
1:05:29
SPEAKER_01: and then the capacity to design complex actions,
1:05:31
SPEAKER_01: that's the anterior of the cortex.
1:05:33
SPEAKER_01: But in terms of the control hierarchy,
1:05:35
SPEAKER_01: so all that stuff kind of ballooned out,
1:05:38
SPEAKER_01: but it still has to feed into this,
1:05:39
SPEAKER_01: which you might think of as,
1:05:41
SPEAKER_01: in some sense,
1:05:42
SPEAKER_01: a central decision.
1:05:43
SPEAKER_01: That's oversimplified because decisions are
1:05:46
SPEAKER_01: product distributed throughout the system.
1:05:49
SPEAKER_01: But you need to have somewhere where everything comes together so you can check for
1:05:54
SPEAKER_01: conflicts within the system in terms of these high-level goals.
1:05:58
SPEAKER_01: And that seems to be what the hippocampus does.
1:06:00
SPEAKER_01: And so, I mean, obviously, in some sense, that's conceptually relevant to conscientiousness.
1:06:05
SPEAKER_01: But the thing is, it's also relevant to anxiety.
1:06:08
SPEAKER_01: Because what do you do when you detect a conflict?
1:06:10
SPEAKER_01: Well,
1:06:10
SPEAKER_01: you need to get worried that something bad might happen if you don't pay attention
1:06:13
SPEAKER_01: to some of these details that might interfere with your ability to
1:06:16
SPEAKER_01: survive the limit right or just to you know you want to you know piss off certain
1:06:23
SPEAKER_01: people or whatever these counter you know these goals that might be counter to your
1:06:27
SPEAKER_01: current goal to be able to detect those conflicts and the more basic
1:06:32
SPEAKER_01: response than sort of conscientious planning and inhibition of impulses or whatever
1:06:37
SPEAKER_01: is anxious anxiety,
1:06:39
SPEAKER_01: right?
1:06:40
SPEAKER_01: And inhibition and wait, let's stop and check this out and what if things go wrong?
1:06:44
SPEAKER_01: Let's pay more attention to potential consequences.
1:06:48
SPEAKER_01: But also even Ray linked it to consciousness.
1:06:51
SPEAKER_01: He said that what essentially the hippocampus is doing is selecting the contents of consciousness.
1:06:58
SPEAKER_01: And what it is,
1:07:00
SPEAKER_01: the way that it does that basically is that you pay attention to things that are
1:07:03
SPEAKER_01: not yet predicted because you need to bring more processing power to bear on them
1:07:08
SPEAKER_01: in order to make them predictable.
1:07:11
SPEAKER_01: You know, and it's,
1:07:14
SPEAKER_01: Yeah,
1:07:15
SPEAKER_01: and telling out this idea from the accurate inference framework to understand
1:07:19
SPEAKER_01: really the way in which goals are just another kind of prediction,
1:07:22
SPEAKER_01: you know,
1:07:22
SPEAKER_01: I was having trouble kind of piecing that all together.
1:07:24
SPEAKER_01: But once you see that it's all just predictions,
1:07:27
SPEAKER_01: then you get the sense that the hippocampus is basically just trying to resolve
1:07:32
SPEAKER_01: conflicts between these particularly high-level predictions.
1:07:36
SPEAKER_01: And the highest-level predictions are, you know, organized in terms of our goals.
1:07:43
SPEAKER_01: We have a lot of freedom to see the world in ways that are not immediately goal-relevant,
1:07:46
SPEAKER_01: because that's how you store information that could be useful in the future,
1:07:51
SPEAKER_01: even if it's not right at the moment.
1:07:54
SPEAKER_01: But some fundamental things are organized around at least a very high-level goal.
1:08:03
SPEAKER_01: I think we probably talked about this the last time we talked about a different,
1:08:08
SPEAKER_01: particularly large-scale distributed network in the brain that we think is most
1:08:11
SPEAKER_01: relevant to conscientists.
1:08:13
SPEAKER_01: So I guess my point is just that,
1:08:14
SPEAKER_01: yes,
1:08:15
SPEAKER_01: that kind of output of the capacity of the hippocampus to detect the result of a
1:08:20
SPEAKER_01: conflict is clearly related to conscientiousness,
1:08:23
SPEAKER_01: but it's maybe more related to this larger stability factor,
1:08:26
SPEAKER_01: where what conscientiousness and neuroticism and even agreeableness share.
1:08:29
SPEAKER_01: Because one of your goals is to respect other people's goals,
1:08:34
SPEAKER_01: and then their goals also become high-level goals that you have to check for
1:08:38
SPEAKER_01: conflict with your own goals.
1:08:43
SPEAKER_02: We didn't necessarily have to do it this way of like there being structures that
1:08:51
SPEAKER_02: are so central,
1:08:52
SPEAKER_02: even like you can point to that.
1:08:54
SPEAKER_02: It could have been done otherwise,
1:08:55
SPEAKER_02: but it just so happened for things like maybe for anatomical optimization and like evolvability,
1:09:02
SPEAKER_02: things like minimizing connection links to actually make a brain fit in your head
1:09:06
SPEAKER_02: and be energy efficient.
1:09:08
SPEAKER_02: We utilized small worldness.
1:09:09
SPEAKER_02: And one of the ways to do this was creating modularity.
1:09:12
SPEAKER_02: of the architecture and one of these varying degrees of modularity a lot of the
1:09:16
SPEAKER_02: module is flexible and remixable and i forget what anderson says you know but
1:09:21
SPEAKER_02: there's some of them which is like um a little bit more like refined in its
1:09:25
SPEAKER_02: operation and its modes like for instance the campus it's like hunt puts in the
1:09:29
SPEAKER_02: back of the brain distinction you see this reflected along the posterior to
1:09:33
SPEAKER_02: anterior axis of the campus in terms of like the nature of like what's being
1:09:36
SPEAKER_02: tracked anterior
1:09:38
SPEAKER_02: going more seeing more information related to counterfactual and future goal states
1:09:43
SPEAKER_02: and posterior more like the presence like estimate of what's going on perceptually
1:09:48
SPEAKER_02: and so like you'll see like there's this holistic front of the brain versus the
1:09:52
SPEAKER_02: back of the brain
1:09:53
SPEAKER_02: I forget which side of the switch.
1:09:54
SPEAKER_02: Distinction being mirrored within this convergence-divergence zone.
1:10:00
SPEAKER_02: And where this convergence-divergence zone is actually chunking things together
1:10:04
SPEAKER_02: sequentially in these different equilibrium points,
1:10:08
SPEAKER_02: you're moving to the overall cybernetic system as it moves through time and space.
1:10:16
SPEAKER_02: Basically, like...
1:10:19
SPEAKER_02: what is happening here and what does it mean?
1:10:21
SPEAKER_02: What is happening at this next point?
1:10:23
SPEAKER_02: What does it mean?
1:10:24
SPEAKER_02: And so this is different.
1:10:25
SPEAKER_02: So like,
1:10:26
SPEAKER_02: you know,
1:10:28
SPEAKER_02: it's,
1:10:29
SPEAKER_02: we didn't,
1:10:30
SPEAKER_02: you didn't necessarily have to do this way in theory,
1:10:32
SPEAKER_02: but practically that's what evolution seemed to have done.
1:10:34
SPEAKER_02: Right.
1:10:35
SPEAKER_01: It's not just evolutionary agency, right?
1:10:37
SPEAKER_01: Evolution is stuck with working with what was there already.
1:10:40
SPEAKER_01: So if you've got a simple organism,
1:10:42
SPEAKER_01: you're building more complex organism and then more complex organism than that.
1:10:46
SPEAKER_01: The only reason there's only the option of saying,
1:10:48
SPEAKER_01: well,
1:10:48
SPEAKER_01: let's just tear the whole thing down and do it a different way,
1:10:51
SPEAKER_01: right?
1:10:51
SPEAKER_01: You're always building on what was already there and tweaking what was already
1:10:55
SPEAKER_01: there and including things together.
1:10:57
SPEAKER_01: And so, you know, you get specificity in parts.
1:11:02
SPEAKER_02: And one thing that's potentially,
1:11:04
SPEAKER_02: like,
1:11:04
SPEAKER_02: interesting data kind of pointing to this,
1:11:06
SPEAKER_02: like,
1:11:06
SPEAKER_02: hippocampal centrality for,
1:11:07
SPEAKER_02: like,
1:11:08
SPEAKER_02: being able to,
1:11:09
SPEAKER_02: you know,
1:11:09
SPEAKER_02: at the very least,
1:11:10
SPEAKER_02: like,
1:11:11
SPEAKER_02: parameterize consciousness in terms of,
1:11:13
SPEAKER_02: like,
1:11:13
SPEAKER_02: part of binding.
1:11:13
SPEAKER_02: And
1:11:17
SPEAKER_02: And often as part of the generation processes,
1:11:20
SPEAKER_02: like temporal lobe epilepsy is associated with some of the most intense states of
1:11:25
SPEAKER_02: experience people ever have.
1:11:26
SPEAKER_02: Or also things like, if you look at DMT experiences, similar, like the theta rhythm seems to be elevated.
1:11:34
SPEAKER_02: And so that's, I think, points in that direction.
1:11:37
SPEAKER_02: But also just...
1:11:38
SPEAKER_02: The anatomy points in that direction.
1:11:39
SPEAKER_02: If you think of it as like the hippocampus as a convergence,
1:11:42
SPEAKER_02: divergence,
1:11:42
SPEAKER_02: and its high centrality,
1:11:44
SPEAKER_02: or as this person on the water calls it,
1:11:46
SPEAKER_02: like a semantic pointer architecture in computational terms.
1:11:50
SPEAKER_02: And so in the hippocampal system,
1:11:52
SPEAKER_02: I've actually like,
1:11:54
SPEAKER_02: and part of it is it's somewhat related to this neocartesianism,
1:11:57
SPEAKER_02: which I may not be able to get into,
1:11:59
SPEAKER_02: but I was collaborating with these roboticists,
1:12:02
SPEAKER_02: active inference roboticists,
1:12:03
SPEAKER_02: and for them solving the problem of robot autonomy,
1:12:07
SPEAKER_02: um they both um like take some of their main inspiration is from the functioning of
1:12:14
SPEAKER_02: the hippocampal system and its ability to um basically give you this high level
1:12:20
SPEAKER_02: knowledge of what the overall system is doing and constraining that and using it to
1:12:25
SPEAKER_02: govern it as it's selecting actions and using it for the
1:12:33
SPEAKER_02: into the future and navigating through the world.
1:12:35
SPEAKER_01: Somehow there needs to be integration, right?
1:12:39
SPEAKER_01: And I think that's kind of part of your point about let's not throw out the navy
1:12:45
SPEAKER_01: with the last word on Cartesianism.
1:12:47
SPEAKER_02: It doesn't all need to come together in the pineal gland.
1:12:50
SPEAKER_02: But it does need to come together.
1:12:52
SPEAKER_02: And one way for it to come together is like having potentially hub structures with high centrality.
1:12:59
SPEAKER_02: And then when you have these hub structures to be able to refine their cybernetic
1:13:03
SPEAKER_02: functioning in particular ways,
1:13:04
SPEAKER_02: some of which might even include sequential orders of operation that are closer to
1:13:12
SPEAKER_02: symbols and their character.
1:13:15
SPEAKER_02: Not necessarily interpreted that way, but that's a good way to do it.
1:13:20
SPEAKER_02: That seems to be the way we did it.
1:13:22
SPEAKER_02: And there was an active effort in AI.
1:13:26
SPEAKER_02: And I'd say some of the most exciting things were trying to do things like system to AI.
1:13:31
SPEAKER_02: Systems that are smart in the way we are.
1:13:32
SPEAKER_02: Systems that have actual agency.
1:13:34
SPEAKER_02: They are taking some of their main inspiration from
1:13:36
SPEAKER_02: from the hippocampal system.
1:13:38
SPEAKER_02: This is part of what gave Dennis the confidence to found DeepMind.
1:13:42
SPEAKER_02: And so you have work on things like the Thomann-Eichenbaum machine and various other architectures.
1:13:47
SPEAKER_02: And it's actually the cutting edge of AI in terms of getting something like...
1:13:52
SPEAKER_02: systems that can do like hierarchical models of the world and be intelligent with
1:13:58
SPEAKER_02: respect to planning and be able to be,
1:14:01
SPEAKER_02: um,
1:14:02
SPEAKER_02: uh,
1:14:02
SPEAKER_02: proficient at this conceptual level,
1:14:05
SPEAKER_02: um,
1:14:05
SPEAKER_02: and,
1:14:05
SPEAKER_02: and to be sensible in a way.
1:14:08
SPEAKER_03: And so, um, it's, uh, yeah.
1:14:14
SPEAKER_03: So.
1:14:14
SPEAKER_03: Oh,
1:14:22
SPEAKER_03: Sorry, I'm still muted.
1:14:26
SPEAKER_00: Yeah, to clarify, so you're not saying that this is like an executive control center?
1:14:34
SPEAKER_01: Executive control is a really big term, right?
1:14:38
SPEAKER_00: Yeah.
1:14:39
SPEAKER_01: So, you know, a cybernetic system is, in some sense, an executive control center.
1:14:44
SPEAKER_01: That's what it's for.
1:14:44
SPEAKER_00: But it's,
1:14:45
SPEAKER_00: like,
1:14:45
SPEAKER_01: distributed through...
1:14:47
SPEAKER_01: Yeah,
1:14:47
SPEAKER_00: but so,
1:14:48
SPEAKER_01: like,
1:14:49
SPEAKER_01: no,
1:14:49
SPEAKER_01: you wouldn't want to think of it as...
1:14:50
SPEAKER_01: It's not a homunculus,
1:14:51
SPEAKER_01: right?
1:14:51
SPEAKER_01: On the hippocampus, it's not a homunculus.
1:14:53
SPEAKER_01: That's kind of what I'm thinking with the neocartesianism, because I'm curious about that.
1:14:57
SPEAKER_02: I do have homunculi.
1:14:58
SPEAKER_02: They're not at the campus.
1:15:00
SPEAKER_01: Let's not even touch that one today.
1:15:02
SPEAKER_01: But no,
1:15:03
SPEAKER_01: I would think of it as just like a particularly crucial control element for a
1:15:07
SPEAKER_01: complex system that needs to integrate a lot of information and navigate a lot of
1:15:15
SPEAKER_01: potentially conflicting different goals.
1:15:19
SPEAKER_00: Is it connected to narrative in any way?
1:15:21
SPEAKER_00: I know you said it's connected to episodic memory.
1:15:24
SPEAKER_01: Right, well, you can't form a narrative if you don't have an episodic memory, right?
1:15:28
SPEAKER_01: I mean,
1:15:29
SPEAKER_01: a narrative is sort of inherently episodic in some sense,
1:15:35
SPEAKER_01: so you've got to be able to pull it together.
1:15:37
SPEAKER_01: You know,
1:15:37
SPEAKER_01: and we've talked about symbols and the translation into symbols and languages,
1:15:43
SPEAKER_01: and I would give us some way toward answering that question,
1:15:46
SPEAKER_01: right?
1:15:46
SPEAKER_01: And I really like Terrence Deacon's perspective on that.
1:15:53
SPEAKER_02: I think I can put a quick bow on the homunculus.
1:15:56
SPEAKER_02: So basically,
1:15:57
SPEAKER_02: in this collaboration with roboticists,
1:16:00
SPEAKER_02: for handling this SLAM problem,
1:16:03
SPEAKER_02: or simultaneous localization and mapping,
1:16:05
SPEAKER_02: a system's moving through the world,
1:16:07
SPEAKER_02: trying to figure out...
1:16:10
SPEAKER_02: And along the way,
1:16:11
SPEAKER_02: it has to tackle,
1:16:12
SPEAKER_02: to varying degrees,
1:16:13
SPEAKER_02: the frame problem.
1:16:14
SPEAKER_02: You're building a map of the world.
1:16:17
SPEAKER_02: To do this, there's a cart-horse problem of...
1:16:22
SPEAKER_02: To help build a map, it helps to know where you are.
1:16:24
SPEAKER_02: It's hard to know where you are without the map.
1:16:26
SPEAKER_02: And so it seems like one of the things that the campus does in terms of its spatial
1:16:29
SPEAKER_02: character is it's like a spatial mapping system,
1:16:32
SPEAKER_02: in addition to having these functions that Colin's talking about.
1:16:35
SPEAKER_02: And they're part of unified cybernetic functioning.
1:16:40
SPEAKER_01: That's essentially secondary to the output resolution
1:16:44
SPEAKER_01: problem, right?
1:16:45
SPEAKER_01: The reason that it's important for mapping is that mapping involves a lot of
1:16:48
SPEAKER_01: conflict resolution,
1:16:49
SPEAKER_01: like was it over here or was it over there?
1:16:51
SPEAKER_01: Was it next to this tree or next to that tree, right?
1:16:54
SPEAKER_02: That would be more fundamental.
1:16:56
SPEAKER_02: More fundamental, you're doing action selection, apparently.
1:17:00
SPEAKER_02: But as part of these architectures,
1:17:05
SPEAKER_02: and so I kind of came at this from a systems neuroscience perspective,
1:17:08
SPEAKER_02: in some phenomenology,
1:17:10
SPEAKER_02: some of the main things I emphasize is basically visuospatial awareness and
1:17:14
SPEAKER_02: somatospatial awareness as being central to consciousness.
1:17:17
SPEAKER_02: And in this collaboration, it turns out that in the work on robotic autonomy,
1:17:22
SPEAKER_02: Basically,
1:17:22
SPEAKER_02: one of the main data structures they use to situate the robot in the world is a
1:17:26
SPEAKER_02: conjunction of views and poses.
1:17:28
SPEAKER_02: There's a visual data structure of, like, what is the robot seeing?
1:17:33
SPEAKER_02: And there's basically, how's my body?
1:17:35
SPEAKER_02: And these are basically almost the same data structure.
1:17:38
SPEAKER_02: They're closely conjuncted.
1:17:39
SPEAKER_02: So it's almost, as you will, a Cartesian theater structure.
1:17:43
SPEAKER_02: and a homunculus,
1:17:44
SPEAKER_02: and even a homunculus looking at the theater in terms of the intense mutual
1:17:48
SPEAKER_02: information between my view informs my pose and vice versa.
1:17:53
SPEAKER_02: And so I do not necessarily... This is not the hardcore of any theory, but...
1:18:00
SPEAKER_02: I think we should look for such data structures in the brain as being core to their
1:18:05
SPEAKER_02: coherent cybernetic functioning.
1:18:07
SPEAKER_02: And I would not be surprised to find even within hub structures,
1:18:10
SPEAKER_02: potentially even things like geometrically structured subnetworks,
1:18:15
SPEAKER_02: that actually the dynamics do have some
1:18:19
SPEAKER_02: isomorphic maybe diffeomorphic character with actually the geometry of the body so
1:18:24
SPEAKER_02: you know it's a testable hypothesis that like you could basically you know drop you
1:18:28
SPEAKER_02: know you on musketeer link to go to work on it you drop some neuro pixels in theory
1:18:32
SPEAKER_02: if like the manifold activity in cortical hubs has both like some like kind of
1:18:38
SPEAKER_02: cartesian character
1:18:39
SPEAKER_02: Or is it actually mapping out the body in the population activity?
1:18:44
SPEAKER_02: And if you're finding these as being basically mapping them onto experience.
1:18:49
SPEAKER_02: And I think you do have something like a homunculus and even a Cartesian theater.
1:18:55
SPEAKER_02: And so basically, I would argue that Dennett got it precisely backwards and led us all astray.
1:19:00
SPEAKER_02: In terms of what might be one of the central means by which a system like us can
1:19:05
SPEAKER_02: actually model its body,
1:19:06
SPEAKER_02: its relationship to the world,
1:19:08
SPEAKER_02: and allow this to be interfaced with value signals in a way that you can both
1:19:13
SPEAKER_02: inform and be informed by action perception cycles as they unfold and be a more
1:19:18
SPEAKER_02: effective and flexible cybernetic system and an autonomous system.
1:19:22
SPEAKER_02: and even a conscious system.
1:19:24
SPEAKER_02: And so the claim I've been making is that we need to,
1:19:27
SPEAKER_02: to varying degrees,
1:19:28
SPEAKER_02: you don't have to buy the,
1:19:29
SPEAKER_02: you can take it a la carte,
1:19:31
SPEAKER_02: like not be so averse to some Cartesian thinking,
1:19:36
SPEAKER_02: like reject just the passive observer who just sits there and figures it all out
1:19:39
SPEAKER_02: without doing anything.
1:19:40
SPEAKER_02: Reject the fact that there's like a single place where it all comes together and that's the soul.
1:19:45
SPEAKER_02: But the idea of anything that looks like Descartes, we should reject.
1:19:48
SPEAKER_02: Descartes was a titan.
1:19:51
SPEAKER_02: We are not his superiors in so many ways.
1:19:53
SPEAKER_02: I could not do what Descartes did in his time.
1:19:55
SPEAKER_02: That's not to say we take everything wholesale, but Descartes was an idiot.
1:20:00
SPEAKER_02: I don't think so.
1:20:01
SPEAKER_00: I love how much you're repping Descartes here, Adam.
1:20:03
SPEAKER_00: I've never heard anybody fight so hard for him, so it's good to hear.
1:20:06
SPEAKER_02: Yeah, for somebody to fight for him, you know.
1:20:10
SPEAKER_00: Maybe we should end there.
1:20:13
SPEAKER_00: Yeah,
1:20:13
SPEAKER_00: and do you want to,
1:20:14
SPEAKER_00: I mean,
1:20:14
SPEAKER_00: Colin,
1:20:14
SPEAKER_00: do you want to say,
1:20:15
SPEAKER_00: it's very obviously,
1:20:17
SPEAKER_00: just any thoughts on what Adam's presented in terms of,
1:20:21
SPEAKER_00: my understanding of that,
1:20:22
SPEAKER_00: in a sense,
1:20:22
SPEAKER_00: is that
1:20:24
SPEAKER_00: very rough, but that self-modelling leads to self-consciousness and affords us control, Adam?
1:20:30
SPEAKER_00: Is it something like,
1:20:32
SPEAKER_00: is that what you're saying,
1:20:33
SPEAKER_00: that the modelling of the model forms this kind of homunculus Cartesian theatre
1:20:38
SPEAKER_00: that allows us to
1:20:41
SPEAKER_02: I want a data structure that's tracking the geometry of the body,
1:20:45
SPEAKER_02: that's tracking basically my visual fields in terms of what I see in the world,
1:20:51
SPEAKER_02: and is bringing those things together.
1:20:53
SPEAKER_02: Because in my experience,
1:20:55
SPEAKER_02: my experience of the world is a world appears to me with a certain projected geometry,
1:21:00
SPEAKER_02: and there's a live body at its center.
1:21:02
SPEAKER_02: And these things are in close relation.
1:21:04
SPEAKER_02: And maybe that's just like, yeah, screw my phenomenology.
1:21:08
SPEAKER_02: But I also suspect it's not a coincidence that those happen to be the two main data
1:21:11
SPEAKER_02: structures that roboticists are using to actually control their systems.
1:21:16
SPEAKER_02: And that actually,
1:21:19
SPEAKER_02: there are convergence-divergence zones with respect to cortical hierarchies over
1:21:22
SPEAKER_02: vision and somatic cessation,
1:21:25
SPEAKER_02: which actually are part of rich clubs in terms of being in close communication.
1:21:29
SPEAKER_02: Whether you want to think of how homuncular you want to go,
1:21:32
SPEAKER_02: whether the positive,
1:21:34
SPEAKER_02: maybe the submembers are even having geometric properties.
1:21:38
SPEAKER_02: You might want to think, I don't know.
1:21:39
SPEAKER_02: It could be, it could not.
1:21:40
SPEAKER_02: But the idea of looking for basically things to come together,
1:21:45
SPEAKER_02: and maybe even with anatomical centrality,
1:21:48
SPEAKER_02: I don't think we should reject that because of the anatomical centrality.
1:21:54
SPEAKER_03: I second that.
1:21:57
SPEAKER_01: Robotics is important partly because you have to put your money where your mouth is.
1:22:02
SPEAKER_01: You have some kind of controller.
1:22:04
SPEAKER_01: It actually has to exert control successfully.
1:22:10
SPEAKER_01: Adam,
1:22:10
SPEAKER_01: that makes me think about something that one of my excellent students,
1:22:17
SPEAKER_01: actually one of my former students who's just finished his PhD,
1:22:20
SPEAKER_01: Tyler Sassenberg,
1:22:22
SPEAKER_01: he's been working on a paper that is using these cortical gradients.
1:22:30
SPEAKER_01: If you essentially take
1:22:34
SPEAKER_01: you take all of the interconnections functionally among parts of the cortex,
1:22:39
SPEAKER_01: uh,
1:22:39
SPEAKER_01: and you reduce their dimensionality down to two,
1:22:43
SPEAKER_01: uh,
1:22:43
SPEAKER_01: you get these two major dimensions,
1:22:46
SPEAKER_01: one of which,
1:22:47
SPEAKER_01: uh,
1:22:47
SPEAKER_01: basically goes from the,
1:22:50
SPEAKER_01: uh,
1:22:52
SPEAKER_01: primary sensory and motor areas to the most complex high-order cognitive networks,
1:23:01
SPEAKER_01: like the salience default network and frontal parietal control network.
1:23:05
SPEAKER_01: So you've got one that's basically going from the simplest inputs and outputs up to
1:23:13
SPEAKER_01: the most removed from immediate sensation and action type of processing.
1:23:19
SPEAKER_01: And that one seems fairly straightforward in terms of why it's organized that way.
1:23:23
SPEAKER_01: But then the other one is you've been sort of puzzling over the best way to
1:23:27
SPEAKER_01: interpret it,
1:23:27
SPEAKER_01: because it basically goes from the visual system on one side to the sensory motor
1:23:36
SPEAKER_01: system on the other side,
1:23:38
SPEAKER_01: you know,
1:23:38
SPEAKER_01: the primary sensory motor strips,
1:23:40
SPEAKER_01: with those higher cognitive networks in the middle.
1:23:44
SPEAKER_01: And so my initial intuition about that was like, oh, that's like input and output, right?
1:23:48
SPEAKER_01: Those are the two ends of the cybernetic system.
1:23:50
SPEAKER_01: You've got the input and you've got the output.
1:23:52
SPEAKER_01: But I think the fact that the somatosensory strip is together with the motor strip
1:23:59
SPEAKER_01: on one side of that is not with the sensory side.
1:24:02
SPEAKER_01: You know,
1:24:03
SPEAKER_01: it's not like clearly sensory on one side and output on the other because
1:24:06
SPEAKER_01: somatomotor all goes together about sensation and action control.
1:24:12
SPEAKER_01: Tyler wasn't happy with that.
1:24:15
SPEAKER_01: interpretation and so what you were just saying was making me think that maybe what
1:24:19
SPEAKER_01: you're seeing there essentially is like a gradient from the model of the external
1:24:24
SPEAKER_01: world to the model of the viewer of the external world uh and so it's more like
1:24:30
SPEAKER_01: distal sensory representation versus proximal system representation in other words
1:24:36
SPEAKER_01: representation of what is being seen uh contrasted with representation of what is
1:24:41
SPEAKER_01: doing with seeing um
1:24:44
SPEAKER_01: And those are the two ends of a major gradient of particle organization.
1:24:49
SPEAKER_01: And one of the things I had to look for was where the auditory system was distributed in that.
1:24:55
SPEAKER_01: And as I thought it would be, this was really about distal versus proximal.
1:25:00
SPEAKER_01: It is closer to the visual end of the gradient, but a bit more distributed along the gradient.
1:25:10
SPEAKER_01: So yeah, anyway, that's what, that's what, I mean, now we've opened up all additional kind of worms.
1:25:16
SPEAKER_01: That's what that made you think about.
1:25:17
SPEAKER_00: I might have to come back and have another conversation.
1:25:19
SPEAKER_00: Anyway,
1:25:19
SPEAKER_00: this is,
1:25:20
SPEAKER_00: I probably rocked very little,
1:25:22
SPEAKER_00: but it's been,
1:25:24
SPEAKER_00: it's opened some very interesting doors and I really appreciate the work that
1:25:27
SPEAKER_00: you're doing.
1:25:29
SPEAKER_00: Cause as, as we'd mentioned these, there's a therapeutic, there's existential consequences.
1:25:33
SPEAKER_00: There's the artificial intelligence stuff, you know, these questions are all very live at the moment.
1:25:39
SPEAKER_03: Yeah.
1:25:40
SPEAKER_01: And let me just add the punchline to Tyler's anchor there,
1:25:43
SPEAKER_01: which is that people's levels of creative achievement is essentially being
1:25:49
SPEAKER_01: predicted by the degree to which their
1:25:54
SPEAKER_01: their higher cognitive networks are functionally distinct both from each other and
1:26:01
SPEAKER_01: from the sensory inputs,
1:26:03
SPEAKER_01: right?
1:26:03
SPEAKER_01: So it's like having greater functional distance between your higher cognitive
1:26:11
SPEAKER_01: processing and your sensory inputs is potentially representative of more freedom to
1:26:15
SPEAKER_01: play with your representation,
1:26:17
SPEAKER_01: and use that term,
1:26:19
SPEAKER_01: that then potentially leads to creativity.
1:26:26
SPEAKER_00: Yeah, try and extend it, maybe, see if we can get some more credibility.
1:26:31
SPEAKER_00: That,
1:26:31
SPEAKER_00: I think,
1:26:31
SPEAKER_02: connects with the psychedelic literature,
1:26:33
SPEAKER_02: potentially,
1:26:33
SPEAKER_02: in treating ways,
1:26:34
SPEAKER_02: too,
1:26:35
SPEAKER_02: in terms of,
1:26:35
SPEAKER_02: like,
1:26:35
SPEAKER_02: you know,
1:26:40
SPEAKER_02: some models of psychedelics,
1:26:40
SPEAKER_02: like,
1:26:41
SPEAKER_02: I've been proposing,
1:26:41
SPEAKER_02: like,
1:26:42
SPEAKER_02: being akin to,
1:26:42
SPEAKER_02: like,
1:26:42
SPEAKER_02: waking dreams,
1:26:44
SPEAKER_02: and,
1:26:44
SPEAKER_02: like,
1:26:44
SPEAKER_02: the ways in which,
1:26:44
SPEAKER_02: like,
1:26:45
SPEAKER_02: for instance,
1:26:45
SPEAKER_02: like,
1:26:45
SPEAKER_02: hallucinations can occur with sensory deprivation,
1:26:48
SPEAKER_02: or,
1:26:49
SPEAKER_02: like,
1:26:49
SPEAKER_02: with Charles Benet syndrome,
1:26:50
SPEAKER_02: like,
1:26:51
SPEAKER_02: through,
1:26:51
SPEAKER_02: like,
1:26:52
SPEAKER_02: loss of a sensory modality.
1:26:53
SPEAKER_02: The
1:26:58
SPEAKER_03: That's interesting.
1:26:59
SPEAKER_01: We're almost ready to submit that paper, so I'll send it to you when we are.
1:27:03
SPEAKER_03: Thank you both very much.
1:27:07
SPEAKER_03: Pleasure.