back to indexYou Are Your Own Existence Proof (Karl Friston) | AI Podcast Clips with Lex Fridman
Chapters
0:0
0:10 What Is the Free Energy Principle
0:51 The Free Energy Principle
2:29 The Existential Imperative
4:30 The Free-Energy Principle
11:42 The Difference between Living and Nonliving Things
28:5 Philosophical Notion of Vagueness
00:00:03.240 |
So you've also formulated a fascinating principle, 00:00:31.960 |
So I think it's interesting to acknowledge that. 00:00:42.060 |
On the one hand, but I should also acknowledge 00:00:47.880 |
it inherits an awful lot from machine learning as well. 00:01:17.240 |
the probability of existing as the evidence that you exist. 00:01:21.720 |
And if you can write down that problem of existence 00:01:32.200 |
to understand and characterize the ensemble dynamics 00:01:37.200 |
that must be in play in the service of that inference. 00:01:44.280 |
you can always interpret anything that exists 00:01:47.200 |
in virtue of being separate from the environment 00:01:53.760 |
as trying to minimize variational free energy. 00:01:58.760 |
And if you're from the machine learning community, 00:02:01.600 |
you will know that as a negative evidence lower bound 00:02:06.440 |
which is the same as saying you're trying to maximize 00:02:13.880 |
are trying to maximize the compliment of that, 00:02:22.520 |
So that's basically the free energy principle. 00:02:26.160 |
- But to even take a sort of a small step backwards, 00:02:33.120 |
There's a lot of beautiful poetic words here, 00:02:44.080 |
of trying to describe if you're looking at a blob, 00:03:03.200 |
That's just a fascinating sort of philosophically 00:03:21.440 |
So maybe can you talk about that optimization view of it? 00:03:26.240 |
So what's trying to be minimized and maximized? 00:03:29.520 |
A system that's alive, what is it trying to minimize? 00:03:40.000 |
But you've assumed that the thing exists in a state 00:03:52.720 |
what licenses you to say that something exists? 00:04:04.040 |
of things that exist, then they have certain properties. 00:04:25.320 |
So it's good you introduced the word optimization. 00:04:28.160 |
So what the free energy principle in its sort of 00:04:38.080 |
and simplest says is that if something exists, 00:05:04.560 |
as the evidence lower bound in machine learning 00:05:07.360 |
or Bayesian model evidence in Bayesian statistics, 00:05:14.360 |
list of ways of understanding this key quantity, 00:05:19.360 |
which is a bound on surprisal, self-information, 00:05:45.280 |
to be a physicist who was trying to understand 00:05:48.400 |
the fundaments of non-equilibrium steady state. 00:06:04.640 |
As a sort of more specific kind of case study. 00:06:11.720 |
but at its simplest, a single-celled organism 00:06:32.680 |
how on earth can you even elaborate questions 00:06:37.680 |
about the existence of a single drop of oil, for example? 00:06:56.640 |
which is the solvent in which it is immersed, 00:07:03.440 |
Why doesn't the oil just dissolve into solvent? 00:07:14.600 |
and the external states in which it's immersed, 00:07:18.280 |
if you're a physicist, say it would be the heat path. 00:07:27.520 |
an ensemble of atoms or molecules immersed in the heat path. 00:07:32.400 |
But the question is, how did the heat path get there 00:07:40.480 |
I mean, it's such a fascinating idea of a drop of oil 00:07:54.640 |
and also the idea of like, where does the thing, 00:07:58.120 |
where does the drop of oil end and where does it begin? 00:08:03.120 |
- Right, so I mean, you're asking deep questions, 00:08:12.440 |
- But what you can do, so this is the deflationary part of it. 00:08:20.920 |
I answer from the point of view of a psychologist 00:08:22.760 |
and we talk about predictive processing and predictive coding 00:08:27.800 |
but you haven't asked me from that perspective, 00:08:30.080 |
I'm answering from the point of view of a physicist. 00:08:37.200 |
but if it exists, what properties must it display? 00:08:40.640 |
So that's the deflationary part of the free energy principle. 00:08:43.080 |
The free energy principle does not supply an answer 00:08:53.880 |
That's the sort of the thing that's on offer. 00:08:57.720 |
And it so happens that these properties it must display 00:09:01.400 |
are actually intriguing and have this inferential gloss, 00:09:09.840 |
that inherits on the fact that the very preservation 00:09:14.120 |
of the boundary between the oil drop and the not oil drop 00:09:18.880 |
requires an optimization of a particular function 00:09:29.240 |
which is why I started with existential imperatives. 00:09:51.920 |
in computational chemistry with self-assembly. 00:10:08.160 |
from the states or the soup in which they are immersed. 00:10:13.680 |
So from the point of view of computational chemistry, 00:10:21.640 |
to minimize its free energy, its thermodynamic free energy. 00:10:27.720 |
that thermodynamic free energy is just the negative elbow. 00:10:34.420 |
So the very emergence of existence of structure or form 00:10:38.720 |
that can be distinguished from the environment 00:10:43.480 |
necessitates the existence of an objective function 00:10:56.520 |
- And so just to clarify, I'm trying to wrap my head around. 00:11:01.120 |
So the free energy principle says that if something exists, 00:11:13.640 |
we can't just go into a soup and there's no mechanism. 00:11:17.280 |
A free energy principle doesn't give us a mechanism 00:11:30.580 |
to think about, like study a particular system 00:11:42.160 |
to your previous question about what's the difference 00:11:54.920 |
but you kind of drew a line between living and existing. 00:12:19.840 |
from the environment in which they are immersed, 00:12:27.280 |
Taking this sort of model evidence interpretation 00:12:34.120 |
that self-evidencing, another nice little twist 00:12:36.960 |
of phrase here is that you are your own existence proof, 00:12:41.480 |
statistically speaking, which I don't think I said that. 00:12:54.040 |
- I'm gonna have to think about that for a few days. 00:13:11.760 |
First of all, you have to define what it means to exist, 00:13:18.040 |
you have to define what probabilistic properties 00:13:32.000 |
Again, it's not what's connected or what's correlated 00:13:35.700 |
or what depends upon what, it's what's not correlated 00:13:52.920 |
And in this instance, basically being able to identify 00:14:05.160 |
well, there are actually four kinds of states 00:14:08.760 |
in any given universe that contains anything. 00:14:32.600 |
stuff inside the surface, and the surface itself, 00:14:39.720 |
about trying to explore what it means to exist 00:14:53.560 |
- Anyway, so what you were just talking about, 00:14:57.320 |
- So the surface or these blanket states that are the, 00:15:08.240 |
these independences and what different states, 00:15:27.000 |
that you would find in non-equilibrium physics 00:15:29.640 |
or steady state or thermodynamics or hydrodynamics, 00:15:39.240 |
And what it looks like is if all the normal gradient flows 00:15:44.160 |
that you would associate with any non-equilibrium system 00:15:52.200 |
part of the Markov blanket and the internal states 00:15:55.080 |
seem to be hill climbing or doing a gradient descent 00:16:09.160 |
You can write down the existence of this oil drop 00:16:12.040 |
in terms of flows, dynamics, equations of motion, 00:16:20.200 |
we call them active states and the internal states 00:16:27.880 |
trying to look as if they're minimizing the same function, 00:16:31.520 |
which is a log probability of occupying these states. 00:16:41.720 |
So what we're talking about are internal states, 00:17:12.600 |
Well, it means the active states, the internal states 00:17:15.320 |
are now jointly not influenced by external states. 00:17:40.560 |
even a little oil drop with autonomous states 00:17:47.400 |
their variational free energy or their negative elbow, 00:17:53.880 |
And that would be an interesting intellectual exercise. 00:18:08.120 |
Now we make the next move, but what about living things? 00:18:17.600 |
and a little tadpole or a little larva or a plankton? 00:18:22.600 |
- The picture was just painted of an oil drop. 00:18:43.760 |
and acting capabilities and it maintains something. 00:18:53.700 |
That it could die or it can't, I mean, yeah, mortality. 00:19:06.720 |
like an essential element to being able to act 00:19:08.880 |
in the environment, but the oil drop is doing that. 00:19:15.760 |
but does it in and of itself move autonomously? 00:19:26.840 |
I was, I didn't find a passive little oil drop 00:19:36.600 |
- What I'm trying to say is you're absolutely right. 00:19:45.400 |
then you've got, I think, something that's living. 00:19:58.640 |
but they're not influenced by the external states, 00:20:03.200 |
So there are two types of oil drops, if you like. 00:20:06.480 |
There are oil drops where the internal states 00:20:08.840 |
are so random that they average themselves away. 00:20:27.200 |
There's lots of intrinsic autonomous activity going on, 00:20:31.880 |
because it doesn't have the deep in the millennial sense, 00:20:34.120 |
the hierarchical structure that the brain does, 00:20:37.040 |
there is no overall mode or pattern or organization 00:20:50.160 |
but on mass, at the scale of the actual surface of the sun, 00:20:54.320 |
the average position of that surface cannot in itself move 00:20:59.000 |
because the internal dynamics are more like a hot gas. 00:21:04.520 |
Whereas your internal dynamics are much more structured 00:21:15.800 |
your autonomic nervous system and its effectors, 00:21:24.320 |
if you haven't thought of it like this before, 00:21:26.480 |
I think it's nice to just realize there is no other way 00:21:31.000 |
that you can change the universe other than simply moving. 00:21:35.320 |
Whether that movement is articulating with my voice box 00:21:44.800 |
there's only one way you can change the universe, 00:21:48.920 |
- And the fact that you do so non-randomly makes you alive. 00:22:00.840 |
it would be realized in terms of essentially swimming, 00:22:06.560 |
a morphogenesis that is dynamic and possibly adaptive. 00:22:19.080 |
its active states are actually changing the external states. 00:22:30.080 |
that depends upon this deeply structured autonomous behavior 00:22:45.680 |
impressed upon their surface or the blanket states, 00:22:49.880 |
but they are actively resampling those data by moving. 00:22:54.880 |
They're moving towards say chemical gradients and chemotaxis. 00:22:58.600 |
So they've gone beyond just being good little models 00:23:07.160 |
For example, an oil droplet could in a panpsychic sense 00:23:25.960 |
with the ability to go out and test that hypothesis 00:23:30.120 |
So it can actually push its surface over there, over there, 00:23:34.760 |
or then you start to move to much more lifelike form. 00:23:38.760 |
Now this is all fun, theoretically interesting, 00:24:01.840 |
this sort of the central importance of movement, 00:24:05.600 |
I think has yet to really hit machine learning. 00:24:10.040 |
It certainly has now diffused itself throughout robotics. 00:24:15.040 |
And perhaps you could say certain problems in active vision 00:24:32.000 |
with the movement problem and the active sampling of data, 00:24:35.240 |
it's just said, we don't need to worry about it, 00:24:36.760 |
we can see all the data 'cause we've got big data. 00:24:55.560 |
to nearly all the data that we need to be exposed to, 00:25:08.160 |
Let's go and move and ingest food, for example, 00:25:33.760 |
do you find within this framework a possibility 00:25:52.160 |
How can we start to think about consciousness 00:25:56.840 |
- Well, yeah, I think it's possible to think about it, 00:26:00.000 |
- Get it, you heard it, it's another question. 00:26:08.720 |
I think you'd have to speak to a qualified philosopher 00:26:21.840 |
to try and tie down the maths and the calculus 00:26:30.040 |
either in terms of sort of a minimal consciousness, 00:26:38.360 |
And what I'm talking about is the ability effectively 00:26:47.320 |
So you could argue that a virus does have a form of agency 00:26:53.360 |
in virtue of the way that it selectively finds hosts 00:27:05.240 |
to think about planning and moving in a purposeful way 00:27:22.160 |
It talks to its friends en route during its foraging. 00:27:43.960 |
I mean, it would be beautiful if we can find a line 00:27:47.520 |
beyond which we can say a being is conscious. 00:28:08.360 |
So you're saying it would be wonderful to draw a line. 00:28:29.080 |
with at what point does a pile of sand become a pile? 00:28:33.040 |
Is it one grain, two grains, three grains, or four grains? 00:28:57.440 |
I think it's systems that have the ability to plan. 00:29:13.240 |
literally the thermodynamics and gradient flows 00:29:30.360 |
That self-evidencing must be evidence for a model 00:29:47.120 |
it must include a model of the future consequences 00:29:49.800 |
of your active states or your action, just planning. 00:29:52.760 |
So we're now in the game of planning as inference. 00:30:01.200 |
because again, it's the consequences of moving. 00:30:04.440 |
It's the consequences of selecting those data 00:30:10.200 |
And that tells you immediately that even to be a contender 00:30:22.720 |
Then you've got to have movement in the game. 00:30:25.240 |
And furthermore, you've got to have a generative model 00:30:28.520 |
of the sort you might find in say a variational autoencoder 00:30:31.880 |
that is thinking about the future conditioned 00:30:37.880 |
Now that brings a number of things to the table, 00:30:41.320 |
well, those who've got all the right ingredients 00:30:46.680 |
of different courses of action into the future 00:31:05.800 |
as if it's selecting amongst different alternative ways 00:31:08.800 |
forward as I actively swim here or swim there 00:31:23.320 |
I don't think it gets you quite as far as self-aware though. 00:31:51.840 |
I think most people that I know would probably say 00:31:56.680 |
that a goldfish, a pet fish was not self-aware. 00:32:02.960 |
They would probably argue about their favorite cat, 00:32:12.320 |
very well connect to some level of complexity with planning. 00:32:44.640 |
with the free energy principle might be difficult. 00:32:49.280 |
I don't think that, well, perhaps we should just, 00:32:57.840 |
you'd see the truisms that you've just exposed for us. 00:33:07.360 |
Well, what's the free energy principle good for? 00:33:17.320 |
It can be regarded, it's gonna sound very arrogant, 00:33:20.040 |
but it is of the sort of theory of natural selection 00:33:38.040 |
It tells you nothing about the actual phenotype 00:33:40.720 |
and it wouldn't allow you to build something. 00:34:03.000 |
has that same deflationary tautology under the hood. 00:34:18.440 |
And you just keep on going round and round and round. 00:34:31.680 |
like differential evolution or genetic algorithms 00:34:53.280 |
a probabilistic description of causes and consequences, 00:34:57.720 |
causes out there, consequences in the sensorium, 00:35:00.560 |
on the sensory parts of the Markov Planckian, 00:35:07.800 |
and then cause it to autonomously self-evidence. 00:35:11.920 |
So you should be able to write down oil droplets. 00:35:16.080 |
where you have supplied the objective function 00:35:30.120 |
when you can write down your required evidence 00:35:40.800 |
the probability of these sensory data or this data, 00:35:48.960 |
that the ELBO, or the Variational Free Energy, 00:35:54.240 |
That means that you can actually write down the model, 00:35:57.720 |
and the kind of thing that you want to engineer, 00:36:00.640 |
the kind of AGI, or Artificial General Intelligence, 00:36:12.720 |
but you would engineer a robot and a computer 00:36:15.760 |
to perform a gradient descent on that objective function. 00:36:32.160 |
I've just told you, all you need to do is write down 00:36:39.880 |
in the form of a probabilistic generative model, 00:36:42.480 |
probability distribution over the causes and consequences 00:36:45.840 |
of the world in which this thing is immersed. 00:36:50.680 |
And then you just engineer a computer and a robot 00:36:54.040 |
to perform a gradient descent on that objective function. 00:37:04.040 |
So it's the form and the structure of that generative model, 00:37:08.160 |
which basically defines the artifact that you will create, 00:37:11.640 |
or indeed, the kind of artifact that has self-awareness. 00:37:17.960 |
very much like natural selection doesn't tell you 00:37:23.000 |
So you have to drill down on the actual phenotype, 00:37:32.360 |
that tells me immediately the kinds of generative models 00:37:36.720 |
I would have to write down in order to have self-awareness? 00:37:39.480 |
- What you said to me was, I have to have a model 00:37:44.200 |
that is effectively fit for purpose for this kind of world 00:37:49.680 |
And if I now make the observation that this kind of world 00:37:53.160 |
is effectively largely populated by other things like me, 00:38:12.320 |
then it becomes, again, mandated to have a sense of self. 00:38:21.280 |
by things like me, basically a social world, a community, 00:38:25.520 |
then it becomes necessary now for me to infer 00:38:30.400 |
I wouldn't need that if I was on Mars by myself, 00:38:42.480 |
a hypothesis, ah, yes, it is me that is experiencing 00:38:50.680 |
induced by the fact that there are others in that world. 00:38:54.240 |
So I think that the special thing about self-aware artifacts 00:38:59.240 |
is that they have learned to, or they have acquired, 00:39:04.280 |
or at least are equipped with, possibly by evolution, 00:39:10.600 |
there are lots of copies of things like them around, 00:39:13.360 |
and therefore they have to work out it's you and not me. 00:39:20.560 |
I never thought of that, that the purpose of, 00:39:24.440 |
the really usefulness of consciousness or self-awareness 00:39:29.000 |
in the context of planning existing in the world 00:39:31.920 |
is so you can operate with other things like you. 00:39:34.360 |
And like you could, it doesn't have to necessarily be human. 00:39:39.480 |
- Absolutely, well, we imbue a lot of our attributes 00:39:52.220 |
that basically you're me, and it's just your turn to talk. 00:40:00.200 |
the highest, if you like, manifestation or realization 00:40:05.600 |
I mean, the human condition doesn't get any higher 00:40:08.560 |
than this talking about the philosophy of existence 00:40:13.900 |
But in that conversation, there is a beautiful art 00:40:17.740 |
of turn-taking and mutual inference, theory of mind. 00:40:31.800 |
That's the highest, the most sophisticated form 00:40:34.320 |
of generative model, where the generative model 00:40:47.260 |
'Cause without that, we'd both be talking over each other, 00:40:50.640 |
or we'd be singing together in a choir, you know? 00:40:54.320 |
That was just probably not, that's not a brilliant analogy 00:41:09.240 |
I'll re-listen to this conversation many times. 00:41:11.800 |
There's so much poetry in this, and mathematics. 00:41:17.640 |
Let me ask the silliest, or perhaps the biggest question 00:41:53.680 |
- I'm tempted to answer that again as a physicist. 00:41:57.720 |
Free energy I expect consequent upon my behavior. 00:42:05.680 |
what that comprises in terms of searching for information, 00:42:09.160 |
resolving uncertainty about the kind of thing that I am. 00:42:12.560 |
But I suspect that you want a slightly more personal 00:42:26.120 |
and harps back to what you were taught as a child, 00:42:31.440 |
that you have certain beliefs about the kind of creature 00:42:57.240 |
usually in the form of bedtime stories or fairy stories. 00:43:03.160 |
who's gonna transform and it's gonna be a prince. 00:43:13.680 |
And then your objective function is to fulfill-- 00:43:17.000 |
- Exactly, that narrative that has been encultured 00:43:24.440 |
in which you grew up and you create for yourself. 00:43:26.880 |
I mean, again, because of this active inference, 00:43:52.800 |
So the question now is for me being very selfish, 00:43:58.240 |
It basically was a mixture between Einstein and Sherlock Holmes. 00:44:11.280 |
enjoy the fantasy that you're a popular scientist 00:44:16.280 |
who's gonna make a difference in a slightly quirky way. 00:44:24.280 |
and he loved sort of things like Sir Arthur Eddington's 00:44:40.840 |
as I was growing up were all about these characters. 00:44:49.160 |
But it's a journey of exploration, I suppose, of sorts. 00:44:54.200 |
what I imagine a mild-mannered Sherlock Holmes/Albert Einstein 00:45:04.000 |
- And you did it elegantly and beautifully, Carl.