The Creative Uses of Uncertainty

Karl Friston is a British neuroscientist and professor at the Institute of Neurology, University College, London. In March 2021 amid the COVID pandemic, we spoke to Karl Friston by Zoom from his home in London. It was one of the most marvelous professional experiences of my life in the satisfactory resolution of pre-game anxiety and conversational delight in his brilliant mind, and humorous and generous spirit.

We discussed the role of play in creativity; the advantages and limitations of what brain imaging "sees;" whether the particular problem and benefit of human language casts more light or greater opacity on the hard problem of what is consciousness?
We also hear Karl Friston pose a great definition of art, and share with us his view on the near-future of artificial intelligence and our fears it will take over the world.

Producer and Host: Ellen Berkovitch
Co-Interviewer this episode: Iliyan Ivanov
Guest: Karl Friston
Post-Production and Theme Music: Dennis Javier Jasso
Music Samples: Uncertainty by The Fray

STEAMPlant Team: Ellen Berkovitch, Mary Jo Vath, Iliyan Ivanov, Agnes Mocsy

(In)coherence podcast is funded by a STEAMPlant grant from Pratt Institute. 

Transcript

Ellen Berkovitch: When we started our podcast (In)coherence, funded by a Steamplant grant at Pratt Institute, one of our goals as an interdisciplinary team was to frame meaningful questions for the humanities and the arts to ask of neuroscience. I’m Ellen Berkovitch, the host and producer. And at the outset it definitely seemed, on deeper investigation, that the disciplines of the humanities and neuroscience didn’t always speak the same language.

For Example:

[Fx: Montage of  questioner at a neuroscience conference asking, In the IIT theory, is there a plausible estimate for what the value of Phi is, for people and for other systems?

And a movie clip of Mrs. Dalloway 
"Mrs. Dalloway said she would buy the flowers herself."

[Theme Music]

EB: One of the big shockers to me about neuroscience at first illustrates my own naïveté — and that was, this dichotomy between the material brain and the immaterial mind.

Material is heavy, material is physical.

Immateriality .. by contrast, doesn’t weigh. It’s the world of spirit. In language, the word immaterial can even connote inconsequentiality or even irrelevance.

Yet what’s called the hard problem of consciousness as it puzzles philosophers also puzzles brain scientists.

Karl Friston—It’s hard because it’s all in words.. you can’t write down the equations. As a physicist.. that would certainly be one of my misgivings about engaging with philosophers. 

EB—What brain scientists can see in the new forms of imaging that have been developed only over the last 30 years for looking deeply into the brain—represent truly scientific revolutions in imaging technology. But for all that there remains an ongoing human problem —how to pose better questions.

Today we have the great pleasure of speaking with Karl Friston of University College, London.

KF: It’s a great pleasure to talk to you. I don’t know what we’re going to be talking about but it’s sure to be interesting.

EB: Karl Friston is the inventor of no fewer than three forms of brain mapping, and also of something called the free energy principle.

What little part of the free energy principle that  I understand well enough to paraphrase, is that — the free energy principle seeks to describe the behaviors that living things or living systems must evince or show. In other words, what behaviors are those that give evidence of being alive?  

A seemingly simple question with an incredible depth —so deep in fact that this guest is able not only to hold it but hold the edges and transits of other’s fleeting thoughts on his thoughts—a truly amazing gift to have had while making this podcast.

EB: Inviting KF to our show was the doing of colleague Iliyan Ivanov. He is a neuroscientist, psychiatrist, visual artist and musician here in NY. He considers Karl Friston one of his mentors and it was through Iliyan’s connection that Karl agreed to talk with us. He joins us today from his home in London etc. We met up by Zoom one early March afternoon. It was one of the true honors of my life We present you now with our talk with him. 

Here’s what happened.

EB: So one of the key things that we’ve been exploring in this project is of course what is known and what isn’t known about consciousness — and what is known, what is known, through what can be seen of the brain. And I wonder if we can start off by asking you what you see as the most advantageous thing about these new forms — 30 years old and less — of brain analysis for humankind?

KF: Well, since the inception of brain mapping — known as sort of imaging neuroscience — a lot of unanswered questions about functional anatomy, sometimes known as functional architectures, have been answered in part.

So in the 1980s people would understand the way the brain works in terms of functional specialization. There were bits of the brain that were specialized for this and that so let’s take visual  brain as a nice example.

EB: The subjects of initial studies were patients having some sort of a brain disease or a brain injury.

KF: What deficits do these people show? And can we impute or infer the functional role of that particular part of the brain? Leading to a picture of functional specialization where particular parts of the brain were specialized for this or that: Color. Motion. Visual form. And high-level abstractions.  Segregated in particular parts the brain— hence functional segregation.

However, this was just a hypothesis. You know. Nobody could actually say when you perceived the redness of a nice red apple your entire brain wasn’t getting very excited.

EB:  Remember all these technologies are just about 30 years old and it took about a decade for imaging techniques to respond to real-time stimuli being experienced by living subjects.

KF: Until the early 90s when we were able to look at the brain in action. To measure neuronal activity, the activity of brain cells or neurons when confronted with a stimulus that would evoke a sensation of redness, for ex., relative to a black-and-white stimulus that didn’t.

And then you were able to compare patterns of activity under those two different conditions and then see for the first time, yes indeed, a very small part of the visual cortex was engaged by the perceptual act of seeing color.

So that provided if you like an endorsement and a verification of a long-held hypothesis, namely that the brain did in fact organize itself around this notion of functional specialization.

And then the next game which has really been in play since the 1990s has taken us to the next level.

EB: Using Voxel-based morphometry or VBM—one of Friston’s inventions— researchers looked rather famously into the brains of London taxi drivers.

[FX-London taxi driver

Your brain gets conditioned to work this way..]

EB: London taxi drivers have to spend a long time learning what’s called the knowledge— the name of the test they have to pass to prove they have mastered how to navigate the maze of London’s central streets.

KF: One I think possibly unkind characterization of brain mapping was as a sort of new wave or neo-phrenology..You remember in the 19th-century people palpating the skull and trying to infer your competencies and your deficits in terms of lumps and bumps in the bony encasing of your brain, and in a sense, brain imaging was just revisiting that false ideology.

And to a certain extent that was true in the sense that if you want to explore a new country the first thing you do is establish landmarks— the mountains, the rivers, the cities.

You have to identify those spatially segregated features that define the geography and the landscape but the key question really is how does this country work? So now we're talking about how do messages pass around the brain? And how does information or the activity of one part of the brain influence the processing in other parts of the brain?

EB- KF is describing is a hierarchy. At the bottom of the hierarchy are raw sensory data.

KF: Raw visual signals or auditory signals or indeed intercepted signals from body—literally gut feelings.

EB: for those who’ve long suspected we have a form of brain in our gut.

KF: And then the intermediate levels of the hierarchy then see only the lower levels ,and they exchange reciprocally messages so they are both in play top-down and bottom-up messages.

Coordinated distributed processing that is contextualized by high levels in the hierarchy — also referred to as deep.

I use the word deep deliberately because when people talk in AI or Machine Learning about deep learning they mean learning under model of the world that has a hierarchical depth. And of course they’re taking inspiration from your brain and my brain..

EB- Here now is Iliyan Ivanov. Iliyan again introduced us to KF. He works in NY as a neuroscientist and child psychiatrist. 

Iliyan Ivanov: To flip Ellen's question on its head:  In your view what might be the most important limitations of this method to advance our knowledge about neuroscience especially around the complex phenomena of consciousness.

KF: Right. So the limitations practically are simply in terms of the resolution, physically in terms of the spatial resolution, the temporal resolution. So,  imagine that you are a cosmologist trying to understand the way the heavenly bodies couple to each another and interact, and you were equipped with a telescope that is not very good. It has very limited spatial acuity or you can only take an image every day or so.

Then you may be very much compromised in terms of trying to make sense of those data.

With fMRI (you) have a spatial res the order of about millimeters to centimeters. Probably more problematically their temporal resolution is lower-bounded — by the fact you’re not looking at neuronal activity; you’re looking at the hemodynamic or the blood flow responses. the blushes, literally the blushes of the brain,

So when you’re using things like fMRI, you are very severely limited in terms of the temporal resolution.

The converse if you move to things like EEG and MEG that stand for basically the measurements of the electrical and magnetic activity, you’ve got the opposite problem.

There you’ve got an exquisite temporal resolution — you can see fluctuations in neuronal activity down to level of a few milliseconds or least tens to hundreds of millisceconds.

But your spatial resolution  --you don’t know where the signal is coming from.

EB: So here’s a key nexus between the tool and the questions that scientists are using it to attempt to get closer to.

KF-: And that brings us to the other limitation which is not so much the resolution of the data but the resolution of our notions of how the brain works and how neuronal activity underwrites qualitative experience or possibly consciousness, or whether that is even the right kind of question to pose.

These are all just great tools..they’re only useful in generating data and dats is only useful when placed in service of testing hypotheses so now we’re taking about quality of conceptions, hypos and models of how the brain actually works.

To my mind that’s almost a  bigger limitation than the practical limitations imposed by the resolutions of brain imaging devices.

EB: I was reading something about your own history, including that your own origin story of the free energy principle dates to looking at woodlice in childhood.

That led me to wonder whether you subscribe to Carl Jung’s idea that play is fundamentally an essential of creative idea generation.

KF:  Yeah. I mean not only do I subscribe to that, that’s one of the holy grails of some of our theoretical work to  simulate that, to actually get in silico, curious behavior — building little curious creatures in a computer..   and showing that they evince the kinds of properties that you would hope to see in your own children, in your real-life curious creatures.. (laughs)

So that notion of curiosity and creativity is absolutely essential, I think, to the right kinds of theories that explain the way that our brains work. One could even argue that the resolution of uncertainty which underpins curious behavior and creativity—

What would happen if I did that?

What would this look like if I arranged it in that particular pattern?

What sorts of experiences would that evince?

These are all questions that our actions are designed to answer by soliciting the right kind of sensory evidence or information that enables us to answer those questions for ourselves..

Then it is curiosity it is creativity, it can’t be anything else really.

So I certainly subscribe to that yeah.

II: In your mind is that an underlying principle for the scientific type of creativity ?

When Scientists process a lot of data etc. etc. and you reasonably have to build up a hypothesis, versus artists who would say I’m greatly in the moment.. Intuition.. some inbuilt talent/ ability comes out and then you create artistic part of it.

Do you see these as slightly different or is that an underlying principle that is more universal?

KF: I think you’ve just answered your own question very eloquently.

They’re clearly causatively different in the domains of application but there’s a cross-cutting principal theme here, which is resolution of uncertainty through creative and curious behavior. So that principle is enshrined mathematically in the principle of optimal Bayes-ian design.

That there is one particular way of acting to generate new data that will resolve the greatest amount of uncertainty. Mathematically that’s the one that has the most information gain.

[Fx-song The Fray Uncertainty

Uncertainty is killing me and I’m certainly not asleep..]

KF: So you can actually design an experiment as a scientist in a Bayes-optimal fashion simply by knowing what I know,  or knowing I have say a portfolio of hypotheses to explain this phenomena.  And I also know my unknowns.

But if I had those data ,then my unknowns and my uncertainty would shrink but would shrink to different amounts. So you can actually go through and work out what the best next experimental move would be. So that's where I think the truly creative part comes in, new ways of looking at things that may or may not be evidence, that may or may not be endorsed by the next round of data you go and solicit from the world.

EB: It’s so interesting though because as a writer when you say hypotheses I think to be communicable a hypothesis must be stated.. right? articulated in language.

I wonder if you feel language is going to evolve to meet the rapidly advancing requirements of science to try to explain the hard problem.

KF- Yes I think it will but that answer is almost trivial in relation to a key theme you’ve both implicitly brought to the table--comprehension, communication, language—

You’ve moved now into a set of questions that pertain not to just model building or hypothesis building that will explain my world. It’s a question of explaining a world inhabited or constituted by creatures like me — and the notion that I have a sense of me and you are you —

That is a really challenging, and probably one of the most, the most exciting and challenging aspects of applying these basic principles of curiosity, uncertainty reduction.

How would that play out in a world, in an eco-niche where 99 % of all your sensory evidence is actually being generated by something very much like yourself?

You know. Asking the question where does language come from and is it an almost inevitable outcome of an ensemble of curious creatures all trying to predict each other?

and of course if we all share the same language and share the same narrative then a lot of the uncertainty can be resolved for free. because if I lived in a world that didn’t have other people

Then I wouldn’t need language. And interestingly I wouldn’t need a sense of self bc there would be no need to disambiguate between self and other.. so I would know these sensations are caused by me or not me..it’s only when you become part of my world that I now need a sense of self.

II: At times actually I thought of language as like contributing to the hard problem..is making it a harder problem in the sense that, what is really the need of developing the tool of language to describe the experience? That’s how at least sometimes I think of language and I think the way you put it is it’s actually evolutionarily meaningful because if you’re surrounded by other creatures presumably like you. Language would be one of the tools to diminish uncertainty.. and so from that perspective that would make sense.

KF: There are two ways I can read your question. First of all: Has language confounded or indeed generated the hard problem in the first place?

II: It’s a harder problem now.

KF: Well, it’s hard because it's all in words you can’t write down the equations which would be my -as a physicist that would be certainly one of my sort of misgivings about engaging with philosophers because they don’t have the right kind of calculus because it’s all in words. I think that would probably be a disingenuous position to take.

Then the other part of your question is more along lines of it is only things like us equipped with language that can entertain the hard problem.

We can celebrate that puzzlement simply because we can share that puzzlement with someone else in the form of linguistic exchange and in absence of that there would be no real puzzlement.

I wouldn’t sit around as a philosopher puzzling about the fact that I can see red and are you seeing the same red. But it could in a more positive way be said, Is the very existence of hard problem conditioned upon language? And I would guess yes. The reason why I’m guessing that is because of David Chalmers’ most recent move which is the meta-hard problem..

[FX: David Chalmers, The Meta Problem. YouTube link]

The meta problem it’s called the meta problem bc it’s a problem about a problem

The problem of explaining why we think there is a problem about consciousness..

{edit to this segment.} The nice thing is you can proceed with some initial neutrality on that question—on the question of is consciousness real? Or an illusion?

KF: The peculiar role of language in communication was not so much trying to answer how we can have sentient experience or sentience that we are aware of.

It is more why we are puzzled by it. What would a creature or a brain or any artifact have to be able to do to be puzzled by the fact it is experiencing something?

One obvious answer to that is It has to entertain the hypothesis or the counterfactual notion that it’s possible to be someone without that experience. Or I can be in an altered state of consciousness where I do not see that red.

Then — I can now start to ask the question, is the other having the same experience as I am? But in asking that question of course I am now nibbling at the hard problem.

But the very fact that I am asking that question belies the fact that now I have a generative model of you and of course, that’s a wonderful gift.

EB: Yeah. But phrasing the question is itself a hard problem. But I wanted to come back to what you said because as a physicist--and we have a physicist not literally in our midst but as one of our team members, and I’ve discussed with her what she says is a consensus among physicists that any quantum explanations for consciousness are very likely untrue.

KF: So I’m answering carefully because it’s very easy to ridicule sort of quantum level explanations for consciousness — and I’m mindful that somebody might be listening who is very much committed to a quantum view (laughs.)

But if you want my honest answer. Yeah, I decline to review papers that rest upon the notion there is a quantum level explanation for consciousness,  for a number of reasons.

The problem is that quantum physics is limited to the domain of the very very small.

So I think the maths of quantum mechanics is exactly the right kind of maths you would need to have a formal explanation for sentience and in that, consciousness at some elemental level. However, it is not sufficient because it  deals with the probabilistic dynamics at the wrong scale, you need to move up a scale.

EB: You know, we’ve been dealing with a lot of different themes as we’ve been approaching our shows and one of the ones that our painter colleague has been working on is the enduring creation myth of Frankenstein, Mary Shelley's book written when she 18 years old. And this has led us to have some conversations with people reminding of Stephen Hawking and other’s warnings that if AI isn’t developed with attention to — whether they’re Asimov's laws or a do no harm foundational approach— that fairly soon and in fairly short order it will be humans who will be rendered obsolete in a new world—Sentient or no. 

So is sentience sort of going to be a never fully understood element of a vanishing world?

KF-That’s a great question (laughing.)

Of course, if you subscribe to the notion that sentience inherits from communication and language between similar kinds of creatures — then there’s actually quite a pacific, a pacifying angle on the singularity or the catastrophes that could ensue from paper-clip minded AI.

In the sense that to exist in that kind of world, you are going to have to have a good model of those things that occupy and constitute that world. So this would be more akin to evolution as not survival of the fittest but the survival of the nicest and most cooperative by which I mean, those phenotypes that have the capacity to share, to empathize and to work together as set of con-specifics with a shared goal — and build their eco-niche, construct their eco-niche in the service of all of them collectively.

From that point of view. If that is an inevitable and emergent property of this imperative to remove uncertainty, there is no room I am afraid for paper-clip building AI. (laughs)

The only way that AI is going to get a grip in the world. The only way to get into the game is by becoming sufficiently like us to understand us so that we understand them., which requires them to speak the same language, have the same goals. I can imagine them coming in through pets.

[FX: Japanese robot funerals]

EB:  That’s a recording of Japanese robot funerals—robot funerals for robot dogs whose manufacturer stopped making parts that would extend their lives indefinitely.

KF: Perhaps there are aspects of it which do speak to this evoking the same kind of empathy and affection and dyadic-ness that you would have with your cat or your dog. So if you can get active inference machines, artifacts, that evoke the same responses as our pets do, then they may have a chance. They certainly not going to have a chance if they go off and build paper clip factories, that just doesn’t work.

EB: In our first show we had the delight of interviewing jazz musician, guitarist Pat Metheny. Iliyan posed to Pat a number of questions that had been raised at a 2018 neuroscience conference. Iliyan decided it would be interesting to ask Karl Friston one of those same questions.

II: So it reads--Suppose one day we’re able to fully explain the musical experience in neurobiological terms. And the question is, do you have insight into why it might be possible or impossible to create such a device and if it exists would it subtract from music’s nature as a form of art?

KF: I would say it is possible to do that. Would, in so doing, would that remove the appeal and allure of music?

In a sense, the appeal or the allure of music can be reduced to this thing that we’ve been talking about — which is the resolution of uncertainty. So there are people who think that music is the perfect medium that offers the opportunity to resolve uncertainty if I just keep listening for a little bit longer, very much like the punchline of a joke or the reveal at the end of series 2 of your favorite TV show. You don’t know what it is but you know it’s going to be there.

So this is a beautiful situation from the point of view of curious creatures--that is actually the opposite of reducing uncertainty. They’re actually seeking out stimuli, sensations, and possibly even narratives, that are deliberately designed to be uncertain. But only because they know they’re going to resolve uncertainty at the end. So you could actually write that down mathematically and say that was basically the appeal and the reason that we enjoy music.

Now does that make any piece of music less alluring? No. You just know  why it's alluring now. 

If you had a clever hearing aid that substituted for part of your auditory cortex I am sure you’d still appreciate the charm of music, provided that of course you were being acquainted with your neuronal hierarchy, your brain hierarchy that was actually having these long-term predictions about what could happen at the end of this riff or in bar 16 or at the end of this story.

Or is he going to play it like this or like that?

Again, a  blush—not the same kind of blush but it certainly is a blush.

Of course, you would know better than many interviewers that that would also in our brains be accompanied by a blush of dopamine that usually signals this resolution of uncertainty.
I know what's happening, I know exactly what I'm going to do now.

And we feel good about that. And we have a little neurotransmitter blush which brings us back for more.

EB: I can say the pleasure is truly all mine. I’m honored to have had this conversation, thank you so much.

KF: This is the denouement, it’s been wonderful talking to you.

(In)coherence Podcast:

Producer and Host: Ellen Berkovitch
Guest: Karl Friston
Co- Interviewer this episode: Iliyan Ivanov
Post-Production and Theme Music: Dennis Javier Jasso
Music sample: Uncertainty, The Fray
STEAMPlant Team: Ellen Berkovitch, Iliyan Ivanov, Mary Jo Vath, Agnes Mocsy

Funding Provided by a STEAMPlant grant from Pratt Institute