# Full Text: Transcript of: Karl Friston, 1st Applied Active Inference Symposium, Active Inference Lab, June 21, 2021

> Extracted from `v2 - Karl Friston, Applied Active Inference Symposium, ActInfLab, June 21, 2021.pdf`

---

## Page 1

Prof. Karl Friston, Applied Active Inference
Symposium, June 21, 2021
Abstract:
On June 21st, 2021, Active Inference Lab (activeinference.org/) hosted its first Applied Active
Inference Symposium, featuring Professor Karl Friston. The Symposium was structured in three
sections, corresponding to the Organizational Units of the Active Inference Lab: Education,
Communication, and Tools. This publication reflects an edited and enriched transcript of the
proceedings of the Symposium.
Authors (Participants)
Karl Friston 1, David Standish Douglass 2,3, Maria Luiza Iennaco de Vasconcelos 2,4, Stephen
Sillett 2,5, Lorena Sganzerla 6, Dean Tickles 2, Ivan Metelkin 2,7, Alex Vyatkin 2,7, Daniel Friedman 2,8
Author Affiliations
1. Wellcome Centre for Human Neuroimaging, Institute of Neurology, University College London, London WC1N 3AR, UK
2. Active Inference Lab, activeinference.org
3. American Society for Cybernetics
4. University of São Paulo, São Paulo, Brazil. Department of Philosophy
5. Canterbury Christ Church University, Salomons Institute for Applied Psychology
6. Berlin School of Mind and Brain, Humboldt-Universität zu Berlin
7. Systems Management School (Moscow)
8. University of California, Davis, USA, Department of Entomology & Nematology
v2.0 (March 2022)
1

## Page 2

Video Sections & Links:
●
Part 1: Education (.edu), page 3, youtube.com/watch?v=INRaCBikpso
●
Part 2: Communication (.comms), page 27, youtube.com/watch?v=X2GwqUVLlcs
●
Part 3: Tools (.tools), page 47, youtube.com/watch?v=hW9IiOujS1E
2

## Page 3

Session 1 (.edu), Introduction
00:01 Friedman:
Hello, and welcome to the Active Inference Lab, to our first-ever Applied Active
Inference Symposium. Today it's June 21st, 2021; and we're very honored to be here
with Professor Karl Friston, and many of our lab participants.
00:20 Just as a way of quick introduction, the Active Inference Lab is a non-profit
organization that is a participatory open science laboratory. We're working to curate and
develop applications related to the Active Inference framework -- something that,
hopefully, we'll be going into a lot more in detail today. And this is a screenshot of our
website [bottom of page 2].
00:45 As far as the overview of this symposium, there are three organizational units in
the lab: .edu [education], .comms [communication], and .tools. And each of these units
are going to facilitate a 45 minute or so session, and we'll have a short break in between
sessions.
Lab organizational units: .edu, .comms, .tools:
3

## Page 4

In our weekly meetings over the past weeks, for each organizational unit, we've been
developing questions and getting excited about things that we wanted to talk to you
about, as far as a few overarching themes that were kind of spoken to, really through the
whole journey of our Lab, but also across organizational units.
Theme 1. Applying Active Inference Across Systems
01:32 The first theme is [1] Applying Active Inference across systems (again something
that will come up probably in all sections);
Theme 2. Research Debt
01:40 The idea of [2] Research Debt, the idea that we don't want to be developing
research frameworks that have a huge burden on those who are learning and applying;
and that, especially early in the formalization of frameworks, it's extremely valuable to
increase the accessibility, so that we don't end up with major headaches and
incompatibilities later on;
Theme 3. Collective Intelligence
02:06 [3] Collective intelligence and the ways in which that is manifest across different
systems;
Theme 4. Transdisciplinary Teams, Projects, Communities
02:12 [4] Transdisciplinary teams, projects, and communities, which are kind of like
nested levels of organization (but transdisciplinarity is something that is necessary for
the type of work that we're all interested in);
Theme 5. Challenges and Opportunities for Research
02:27 And also just modern [5] Challenges and Opportunities for Research and all that
that means related to online and everything else;
And, of course, [6] Anything else that you have tumbling around, and wanted to bring to
the table, thematically.
So there we are, with our sort of lab overview and introduction.
4

## Page 5

Session 1 (.edu), Education
02:54 Let's go to our first organizational unit, .edu.
02:59 The goal of .edu is to scaffold and create a participatory and dynamic Active
Inference Body of Knowledge, which we'll talk more about in a second.
Ontology Term-Development Timeline
03:10 Our progress and actions this year have been to release a terms list, v1, which
benefited greatly from your feedback. And also we're now updating the terms list to
version 2, which now includes five complete language translations and many references
and citations for the terms.
5

## Page 6

Wong, Wilson; Wei Liu; & Mohammed Bennamoun Ontology learning from text: A look back and into the future (2012)
03:34 The way that we're approaching the development of the terms is by using
approaches that place ontology, and progressively more formalized versions of
ontologies, as kind of the backbone of an educational Body of Knowledge.
So we started on the left side here, with a terms list in the first quarter of 2021. The
Ontology Working Group is like a train that's pushing to the right, as they're learning
ontology by doing, and developing progressively stronger and stronger ways of relating
the terms and the concepts that are essential for understanding Active Inference. And
this will help us develop principled educational material that's also able to be translated
rapidly.
6

## Page 7

04:24 Alex, do you want to give a quick thought on where knowledge engineering
comes into play?
04:33 Vyatkin:
Yes, thanks. At this slide, we are showing this work with ontology with a system
engineering approach, which we are also using in the Lab; and considering possible
deliverables of working on educational materials and creating them. We should have at
some point of time textbooks and educational courses, and actually maybe this Lab is
started from the idea that a textbook for Active Inference should be created. Also, we see
some connections that can be applied to organizational management for creating
translations and to make it multi-language from the beginning. And also we should look
for some domain-specific use cases that we can understand in terms of that ontology that
we are going to create.
05:35 Friedman:
Thanks, Alex. So on to the questions section. We're going to start off pretty broad here in
the .edu:
.edu Question: 'How do we determine the core terms and ideas for Active Inference?'
05:45 How do we go about determining the core ideas and terms for Active Inference?
This will be the format of the question slides, Karl, so feel free to jump in.
7

## Page 8

06:09 Friston:
Right! I guess it will be structured around the key ideas, and essentially ingredients that
underwrite the Free Energy Principle, and how that translates into Active Inference. So,
without thinking about it too deeply, my mind just goes to what are the things, what are
the basic ingredients you need to explain to somebody, what Active Inference is, and
why it works.
And it normally starts off with the notion of a generative model, and then from that,
you spin off all the appropriate mathematical ideas and constructs and descriptions that
would attend that. I mean it may be best to reflect the question back to you.
Role of a Formal Ontology
07:10 This is a really neat idea -- having an ontology! And it's certainly my experience
that people are entertained by, sometimes the poetic use of phrases and descriptions,
such as epistemic affordance, when trying to grapple with, "What are the fundamental
ideas behind Active Inference?" Some of them are fundamental and some of them are
not. So, it's certainly an interesting idea to try and tie down the ontology.
But let me ask you: This ontology just means what it says, in the sense that you're trying
to define the essential concepts and how they relate to each other? Is that the basic idea?
07:57 Friedman:
Yep. Going back to this slide here, [shows Active Inference Working Ontology] we want
to have a continuum from a list of terms, potentially, that could be developed into
coherent and, again, principled course material and competencies; but also develop a
logic. And we're developing within the SUMO ontology development framework, which
defines not just relational edges, but an actual logic.
And so we hope to be able to ask, "Is this a complete Active Inference model? Have we
really checked off all the boxes?" -- and used those kinds of logical tools that are
accessible to the well-developed ontological frameworks.
08:41 Friston:
Okay. Well that's very compelling and very clear.
It strikes me then that it would be useful to link that operational ontology to the
underlying maths.
8

## Page 9

Much of the conceptual steps, both in understanding and implementing Active Inference,
(usually in terms of simulating your interesting behavior, or using it as an observation
model to explain some empirical data from a study) -- much of it can be developed in
terms of a series of moves that usually (or, in fact, almost universally) inherit from, are
framed in terms of, either information theory or linear algebra or differential equations;
and you can just build the story from that.
So, if you're looking for that degree of formal and useful detail, then one principle you
might refer to is basically: "Where does one equality assertion or description or variable
or object -- where does it come from in terms of inheriting from the more basic
formalism?"
So, what I'm thinking of here is: "Where does Active Inference start? And how do you
get to the calculus and the Bayesian mechanics that you'd associate with Active
Inference?" And my guess is: given the structure or the way that you have approached
the ontology, you've probably actually done that already or are in the process of doing
that.
Are you going to go through some examples that would sort of highlight the strategy and
the problems, which are usually more illuminating than the solutions that you've
encountered so far?
9

## Page 10

10:49 Friedman:
Sure! I'll switch here to this screenshot of the current state of what it looks like.
And we're starting just in tabular form by compiling up to five references and citable
definitions.
First just looking for exact cases where a term is used. And then we'll go from how the
term has been used, towards synthetic definitions that capture different senses of the
term.
And then along with the concise narrative of the field, and also ontology experts who are
here with us, we're going to then be working to make the actual logical underpinnings,
elucidated in terms of specifiable code, rather than just concise English definitions.
And then from that sort of generator of the formal relationships we'll be able to descend
into mathematical formalisms, or other natural human languages.
11:48 Friston:
Yeah.
11:49 Friedman:
We'll keep you posted on this project though, for sure.
Let's go to this next question, and imagine that we had that set of terms in development
(it's going to be a work in progress our whole lives):
.edu Question: 'How Do We Go from Core Terms and Ideas to Interactive and
Enlivening Education?'
12:04 "How would we go from core terms and ideas, to an interactive and enlivening
education that speaks to people from many different backgrounds?"
Aim at non-specialists.
12:18 Friston:
So, I'm going to answer this question from the point of view of my experience as a
supervisor, which is probably a little bit of a narrow remit from your more general
ambition. I imagine that this is related to this notion of -- (was it "research debt?") -- but
this notion that you don't want to put too much pressure on people, when becoming
acquainted with the utility and application of either the code or the ideas.
10

## Page 11

Build toy models
12:51 In my experience, in an academic setting, just having toy simulations is usually
the best way to give people a feel for what this approach does and how it can be used.
So, it's enormously potent in terms of demystifying and also illustrating the functionality
at hand, or that can be accessed. Having a sort of a working, or at least a toy, model
provides a proof of principle, and that can strip away the magic as well.
And, I think your ambition to try and make this accessible to people who are not
necessarily fluent in the underlying information theory or dynamical systems, is very
laudable and perfectly feasible. So, again, in my experience, some of the most creative
applications of Active Inference can be by people who don't really necessarily wonder
too much, "What's underneath the hood?"
Get the generative model right!
14:11 It all comes back again to the design of the generative model. So, if you get the
generative model right, and it's apt to describe the thing that you want to understand or
to simulate, then usually everything else follows suit.
And I mean that in the sense that you can just take off-the-shelf software, which I
presume that your ultimate ambition is to make available, and make it work in the
service of saying, well... "What would this agent (or this synthetic creature or person) do
in exchange with her environment, if this was the generative model and this was the
generative process?"
So, a lot of this really, I imagine -- in terms of answering your question -- "how do we
go from core terms to interactive and enlivening education?" -- is just establishing a
language, a lexicon, that allows you to talk through somebody in constructing their own
simulations that speak to the issues, that engage them either academically, or beyond
academia.
So, clearly then, the core terms play the role of literally a language, in terms of
communication, which brings us back again to the importance of the ontology, and
having the terms linked in a formal way to the mathematical expressions and also
procedures and processes. So, I guess that a precondition to use the core terms in an
interactive and enlivening, educative sense will rest upon getting that ontology right.
In my experience, you know, the best way to get the ontology right, in the sense of it
being enabling, is just to talk about the terms, until there's some consensus and
11

## Page 12

everybody understands them, both in terms of their teleology, but also in terms of where
they come from -- from the point of view of the code and ultimately the maths that
underwrites all this.
Is that the sort of answer you're looking for here, or thinking along with? Are you
thinking along the same lines?
16:46 Friedman:
It sounds great. There's so many dimensions there!
Just to provide a summary, or just jump in at one place:
.edu Question: 'What is Active Inference? - - what does Active Inference do?
16:56 What is Active Inference, and what does Active Inference do?
17:03 Friston:
Right! That's perfect! - Because, I was just thinking: it would be really useful just to go
down the terms that you had in the previous slide highlighted in green, because all the
heavy lifting here is really just shouting about, what are the core aspects and claims, or
the core things that you're trying to communicate with any one of those terms.
So, for me, Active Inference would be a description of a process that can be seen as
something that arises from the Free Energy Principle. So you can either tell that story
from the point of view of a physicist, and say that Active Inference is a teleological
description of processes that systems that self-organize must possess; or you could tell
the story, or define Active Inference, from the point of view of neurobiology and
ethology, from a point of view of, say, predictive processing, and describes what it
entails.
And I've used the word "Bayesian mechanics" before, because from the point of view of
the physics definition, it would be a teleological description of a Bayesian mechanics
that necessarily arises, you know, (with certain assumptions) from any self-organizing
system.
One key thing about Active Inference, which I think would be important to put in the
definition in the ontology (I'm not sure if it's already there…): -- It's beyond predictive
processing. It's beyond sentience. And it emphasizes, or reflects, the pragmatic turn at
the beginning of this century ( epitomized by the 4Es - embodied, embedded,
extended, and the like), to make it clear that sentience is active, and that you are talking
12

## Page 13

about the circular causality of engagement of any particle, person, or plant, with
whatever is out there.
So, that would be certainly one thing to emphasize in terms of what Active Inference
means.
The "inference" is interesting, in the sense that it does imply a process, and a process
with purpose, which is to infer; which is why I keep using the word, "a teleological
description" -- of something that's actually underneath the hood from the point of view
of physics.
Getting the name right
20:11 One final point here is: There's an easy confusion, I think, between, first of all,
active inference and passive inference. So, that's certainly something which probably
needs resolving, certainly in the philosophical literature. So, I often come across
philosophers who say, "Well, there's passive inference, or perceptual inference (which is
just basically inferring states of affairs in the world on the basis of some sensory
evidence). And then there's the 'extra' bit, which is the active bit which is: 'Now you're in
charge of gathering that sensory evidence upon which you are now going to prosecute
your perceptual inference.'"
That's an interesting dichotomy, which I'm not sure is a correct dichotomy. If it's not
right (I'm not sure that it is not right, in the sense that it is a useful distinction) -- but
certainly is not what Active Inference was originally termed to mean.
You know, by conjoining "active" and "inference," there were a number of motivations.
First it was a generalization of David MacKay's active learning, but probably more
importantly, it was a nod to the notion of active sensing, and active perception -- that
perception is in and of itself, an active process, a constructive process -- that you have to
put policies, plans, and action into the game. So that, I think, would be one important
aspect of Active Inference to define, and I don't know that it has been defined so far. So,
you know, perhaps it's your job to define that.
Inference about the consequences of action - in the future!
22:06 The other thing which is important, I think, in terms of emphasizing what Active
Inference entails actually comes from that enactive perspective, which is inference
about the consequences of action.
13

## Page 14

And that has an important but really simple concomitant: that the consequences of action
are in the future. And that means you now have to think: if you're thinking about Active
Inference in terms of teleology or as a normative theory of behavior -- of sentient
behavior. And you have to now think about -- I should qualify: When I say normative, I
mean it can be operationally defined as an optimization process that, in turn, requires
you to define the objective function or functional. And that's important practically,
because if you're now thinking about sentient behavior, Active Inference, and its
influence about things that haven't yet happened, because you haven't yet acted, then
you're necessarily talking about objective functions or functionals that are about states of
affairs in the future. And that is an important move, and something that Active Inference
embraces, which goes beyond predictive coding. Much of the literature in the 1990s,
and subsequent, much of the literature that inspired that sort of enactive perception or
active sensing take -- situated cognition take on sentience originated in things like
predictive coding. But predictive coding is not what is meant by Active Inference: you
can do predictive coding just by , if you're a statistician, just minimizing Variational
Free Energy. That's only half the game, once you move into the world of Active
Inference.
From a teleological perspective, you have to do that, you have to form beliefs about
hidden states of affairs in the world , using the perceptual side of perceptual inference -
but that is only in the service of rolling out into the future, and deciding what the best
thing is to do next. And that running out into the future and deciding clearly calls for an
objective function.
So in Active Inference, that would be the expected free energy, which may or may not
be unfortunately named - but that's what it is. And therefore, Active Inference sort of
implies that you are committed to optimizing an expected free energy; and implicitly it's
all about choosing the next thing to do.
So, for me those would be two cardinal things that should be embraced by a definition
of Active Inference that, you know, transcend other normative approaches. So, for
example, you know, reinforcement learning in behavioral psychology would be all
about what the good things are to do, and you commit to a loss function, or a value
function of states , if that was the kind of behavior that you're trying to describe.
14

## Page 15

If, on the other hand, you were all about the psychophysics of perception, or just
building basal digital terrorist optimal recognition systems, where you weren't in charge
of gathering those data, then your objective functions would be very, very different.
But what Active Inference says, well you can't carve up the two problem domains,
because they're just both sides of the same coin. And thereby you're now facing the
problem of defining an objective function that is fit for purpose, that does both the belief
updating about latent states or hidden states generating the data, and also the best way
to solicit or cause those data or outcomes under some prior preferences or some
goal-directed constraints.
Is that a good long-winded answer?
26:45 Friedman:
Thank you for the comprehensive answer! It leads directly to our next questions, which
are:
.edu Question: 'What is the Free Energy Principle?'
26:51 What is the Free Energy Principle? And especially,
.edu Question: 'What is the relationship between Active Inference and the Free Energy
Principle?'
26:54 What is the relationship between Active Inference and the Free Energy Principle?
27:02 Friston:
Right! Well, that's, I think, a slightly easier question to answer!
The Free Energy Principle is just a variational principle of least action. Why is it
special? -- or not formally identical to all the other variational principles that we use?
If you look under the hood, right from quantum, through statistical and stochastic, to
classical mechanics -- Well, the only thing that differentiates, really, the variational
principle of least action that is the Free Energy Principle, is that you're paying careful
attention to the separation of states to which you apply that principle - the separation of
states into the states of an agent, or a particle, or a part of a person -- and the outside
states.
So technically, , if you were in statistical thermodynamics, for example, you'd normally
assume that separation, in terms of some idealized gas that was contained within the
15

## Page 16

container or heat reservoir, or a heat bath -- without really worrying about where the heat
bath or the heat reservoir came from.
But the Free Energy Principle says: Well no, you can't really do that. You've really got to
attend very carefully to what licenses a separation of different kinds of states, so you can
assign to the inside of something -- and the outside of something -- and the states that
mediate the exchange between the inside and the outside. And then you get into the
Markov blanket and Markov boundary literature.
So, just to summarize: A Free Energy Principle is just a principle of least action, by
which I mean, that there is a description of dynamics in terms of the most likely paths
any system will take. That is the special provenance of a partitioning, or a separation,
of the states of some universe into the states that are owned by an agent (or a particle),
and those that are not, and the states that mediate the exchange between them. So that
would be the Free Energy Principle.
Active Inference, as I say, is a sort of teleological spin-off from the Free Energy
Principle, in the same sense that you have now at hand a principle of least action. It
allows you to identify, simulate, define the most likely paths, trajectories, or narratives
that a system will pursue under certain conditions. And those conditions are just that
there is an attracting set of states which that system will converge to, or will look as if
it's attracted to.
So, sorry! – What I was working towards, was the notion of an attracting set, as a
metaphor for equipping that physics with a teleology; and that teleology is nicely
illustrated by the notion of attraction. So when mathematicians talk about attractors --
in the particular case, in the Free Energy Principle, these are these sort of pullback
attractors or the kind of attractors that you get in random dynamical systems.
There's a proper and natural tendency to think that these particular states of the
attracting set literally attract in the sense of gravitational attraction, or any other kind of
attraction: "They pull -- they pull states towards them." So that, to me, would be a
teleological interpretation which, I think, is much closer to Active Inference. -- That
you're saying, "That "influence is a process that has a purpose; and the underlying Free
Energy Principle allows you to say the way it looks, as if self-organizing systems show
these certain properties; they're attracted to certain states; they're attracted to certain
16

## Page 17

paths; and we can describe those in terms of the teleological ontology." And that would
be Active Inference.
One practical difference between Active Inference and the Free Energy Principle, is that
the Free Energy Principle is just a principle. It's neither right or wrong; it's just like
Beren Millidge has noted: it's sort of like [Amalie Emmy] Noether's Theorem or
[William Rowan] Hamilton's principle of least action. But as soon as you start to say,
"Well, I think that this principle applies to this population or person or particle," that
certainly commits you, or requires you, to define the attracting set of states - a pullback
attractor (in another jargon, the equivalent would be a generative model); and as soon as
you commit to a generative model to explain the teleology of this system, or this agent,
or this person, then you've moved into [or "from"? -Ed.] the world of non-falsifiable
principles into falsifiable hypotheses, because you could have chosen the wrong
generative model, and thereby there will be evidence for choosing this generative model
or that generative model.
So, the relationship between Active Inference and the Free Energy Principle is
operationally quite simple: Active Inference is the application of the Free Energy
Principle to a particular system. But in that application you're bringing a lot of teleology
to the table, and more specifically you're having to commit to a particular generative
model. And as soon as you do that, that becomes your theory or your hypothesis about
what is an apt description for this system. So, a number of interesting distinctions, in
terms of the relationship between Active Inference and Free Energy Principle that I
imagine your ontology is already addressed, or it's certainly addressing.
33:43 Friedman:
Well, we'll get there! Thank you for that excellent answer!
For the next question: Lorena, please read it out.
.edu Question: 'How and Where Does the Idea of Information Play a Role in FEP /
Active Inference?'
33:52 Sganzerla:
Still in the spirit of broad questions and broad terms, and that, I think, comes in line with
(what) came before. So, how and where does the idea of information play a role in the
Free Energy Principle, and how does it relate with Active Inference -- in the sense of,
17

## Page 18

what is something to keep in mind when thinking about information dynamics in Active
Inference?
34:20 Friston:
Right! Well, these are great questions! I'm getting the hang of this now. You just want
me to talk! I've presented a question to another talking! -- Which I'm very happy to do.
Are you sure you want me to do that? Or should this be a conversation? Perhaps it'll turn
into a conversation at some point. Anyway.
So, information. It plays a dual role, in the sense that information theoretic formulations
underpin most of the derivations behind that principle of least action. And it can be no
other way -- in the sense that all mechanics from physics, is really articulated in terms of
probability densities or distributions.
And as soon as you have a mechanics, or a calculus, or probability distributions, you're
effectively in the world of information theory. And you see that at many different levels.
One nice example of this is that the central quantity that we often use to score the
likelihood of being in a particular state -- if you're a statistician, that would be the
marginal likelihood; if you were fluent with an FEP ontology, it would be surprisal (or
more simply surprise)-- and that is just basically the self-information. If you're a
physicist, you look at this as a potential -- it's a negative log probability.
So you start, really, when thinking about the physics, with this central concept of
self-information, which, I repeat, can be read as a potential function, or a surprisal
function, or surprise. And it is the thing that the Variational Free Energy is a bound
approximation to. So at that level -- and then every other move you make
mathematically, in terms of the expected self-information being the entropy -- and why
that is important as a characterization of various probability distributions in the setting of
self-organization -- would testify to the fact that information theory is absolutely central
to all the maths that underlies the physics of the sentience that emerges from having a
distinction between the states of the system and states that are not in the system, namely
the Markov blanket.
Having said that, I think "information" to most people's minds usually means more.
Certainly in the folk psychology context, it's really information about something. And
the FEP Active Inference has, I think, something quite special to bring to the table here,
that goes beyond information theoretic treatments that you get in communication and
signal processing and rate distortion theorems. All of that kind of information is just
18

## Page 19

your extensions of information theory that inherit from self-information or the
implausibility of a particular event or message -- or, in more abstract domains, such as
sentience and consciousness, you would go to something like integrated information
theory. But that is all about this "Shannon-esque" kind of information.
The opposite, the other kind of information, which is information about something -- So,
what I wanted to try and put on the table, is the very fact that you've got this Markov
blanket, or separation of states on the inside and states on the outside, means that now
you can equip the states on the inside with the role of encoding posterior conditional
Bayesian beliefs about states on the outside. And that introduces, technically, a different
kind of information geometry, a different kind of information theory -- where,
crucially now, you can read the internal dynamics as containing or having information
about what's going on, on the outside. And this is a really important move, equipping
your neuronal dynamics or variational message passing or belief propagation in a
computer, with an information geometry, that now allows you to read off the state of the
computer, or the state of the neural activity, in terms of what it is encoding, or the
information it contains about the outside.
And so that sort of dual aspect information geometry has been celebrated to a minor
extent in the philosophy literature by Wanja Wiese, asking the question, "Is this really
the maths of sentience?" -- where you now have information about things. And in a
sense, that really is the heart of the Free Energy Principle -- or Active Inference, anyway
-- in the sense that it equips that information geometry. I mean, technically, what you are
saying is that any particular internal state of a computer, or a person, or a brain, now can
be read as encoding a Bayesian or a posterior belief about other states, namely, hidden or
latent causes outside the Markov blanket. And that defines, technically, something called
a statistical manifold. And as soon as there is a statistical manifold, there's an
information geometry. And any movement on that manifold necessarily implies a
change in your Bayesian beliefs, namely Bayesian belief updating. Which means now
there's an interpretation of neuronal dynamics, movements on a statistical manifold on
the inside, in terms of belief updating. So the notion of Active Inference, as the process
of belief updating, really , rests upon this fundamental notion that there's information
about stuff that is encoded or parameterized, by the internal machinations, and the
mechanics, and the dynamics of the inside.
19

## Page 20

If you're trying to educate people in terms of how they should understand information, I
think it'd be important to differentiate between:
INFORMATION OF THE FIRST KIND
41:27 the mathematical notions of Shannon information, or self-information –
INFORMATION OF THE SECOND KIND
41:42 And the information implicit in an information geometry; namely, the information
about something. This second kind of information is implicit in an information geometry
and the sentience that is afforded by Active Inference -- when now you're understanding
neuronal dynamics, or message passing in the computer on some Forney factor graph.
Because in this instance, each of those messages, or those neuronal dynamics, can now
be read as belief updating -- namely changing your mind about other things -- so that the
stuff on the inside has information about stuff on the outside.
Associated terms
“Information of the first kind”
Information theory
Shannon information
Self information
Surprisal / Surprise
Log evidence
Log marginal likelihood
Mutual information
“Information of the second kind”
Information geometry
Bayesian beliefs
Posterior beliefs
Conditional beliefs
Information length
Information geometry
Table 1. Two kinds of information, based upon the transcript.
42:26 Friedman:
Thanks for this important answer.
And we're going to How can the integrity of the Active Inference process theory be
maintained when blankets, blanket states, and generative models are being interpreted in
20

## Page 21

novel ways? What do you think of discussions around Markov, or Pearl, Friston
blankets, etc.?
43:12 Friston:
All right, that's an excellent question. I have quite a technical answer! So if it's getting
too technical, tell me now -- [laughs] I'll try and get back to what you were really trying
to unearth!
This is not a fast moving field; but [has] certainly been a delicate and important area of
discussion over the past few years.
So, in the original introduction of Markov blankets, there was an explicit nod to Pearl's
construction of Markov blankets, and how Markov blankets are used practically in terms
of simplifying message passing in computer science. However, that may have been
something of an oversimplification.
Because from the point of view of the Free Energy Principle, the kind of causality that
the Free Energy Principle deals with is not the kind of causality that people, particularly
people like Pearl, but also people dealing with things like Granger Causality deal with.
From the point of view of the Free Energy Principle: that starts with a stochastic
differential equation, or a random dynamical system written as a random dynamical
equation -- and OU [Ornstein-Uhlenbeck] processes being simple examples -- in physics
these would be Langevin-like equations.
Common to all of these starting points is time, and evolution, and dynamics. Now, there
is nothing in Pearl's formulations (well, certainly there's nothing in Pearl's book), on
causality, that deals with time. And I know that, because -- before the days of PDFs and
being able to go and search particular words -- I had to go through [laughs] and find out
-- there's one paragraph that mentions dynamics! If you were in statistics, computer
science, you know -- this will be the world of dynamic Bayesian nets. This is their
"take" on something which is actually much more universal, which is basically the
universe as a Markovian dynamical process.
So, just stepping back, the challenge now, is to articulate independences -- that
underwrite Markov blankets in the sense of Pearl, in terms of dynamics. So you've now
got to link two quite distinct fields, which is basically the fields of dynamics and
Langevin processes, and things that have paths of least action -- to the world of statistics
21

## Page 22

and Pearl-esque independences and causality cast as interventions that have observable
consequences.
The problem in doing that linking, is that you have to really abandon the notion of
causality in the world of Granger Causality and Pearl, because causality is baked into,
and is inherent in, writing down any differential equation (be it stochastic or random or
deterministic), in the sense that states cause motion.
So the causality in this context would be a more control- theoretic causality. So, that
means that you can't then use the causality concept later on. But it does mean that you've
now got to derive from a dynamical Markovian calculus the necessary conditions that
would lead to the conditional independences that are necessary to define Markov
boundaries.
Just to slip in here: The Markov blanket is composed of minimal blankets: namely
boundaries, in the sense of Pearl. And on most recent analyses, it looks as if the blanket
is actually two Markov boundaries, in the sense of Pearl.
But to get to the sense of Pearl, you've got to think very carefully about, "What are the
constraints that lead to the conditional independences?" -- where those constraints are
specified in terms of equations of motion and things like the amplitude of random
fluctuations.
So, once you've seen that that is the link that needs to be made, that actually simplifies
the thing! It simplifies things, in the sense there's no real latitude for interpretation. So,
I'm going back to the part of your question: "Blankets and generative models are being
interpreted in novel ways." And I don't think there's any latitude of... any novel
interpretation other than the… ... Sorry! If "in novel ways" you mean "the best way," or
"The correct way -- and we just haven't got that yet!" -- then I would concur entirely
with that! [Laughs.]
If you think that there is some latitude, there's some library of insightful reinterpretations
and redefinitions, all of which have equal veracity, then I would suggest that's not the
case. There's only one way, and there's only one Markov blanket, or there's only one
particular partition that can be articulated in terms of Markov blankets. And the only
"novelness" there is really in tying down very precisely and defensibly how you get
from a Langevin formulation to a Markov blanket.
22

## Page 23

At the moment, the "novel" way of doing that looks as if it's that the conditional
independences arise from sparse dynamical coupling, or causal coupling. So, if you
read the causality as the influence that a state has on the motion of itself or any other
state, in this sort of minimal Langevin-like description of the universe, then it is the
sparsity of influence, or the sparsity of coupling, that leads to conditional
independences. And if the system has a sufficiently rich sparsity of conditional
independences and implicitly coupling, then it will have a particular partition; and if it
has that particular partition, then the Free Energy Principle holds.
So, I think the discussions around Markov, Pearl, Friston, and blankets are essential --
they're fascinating. And the conclusions of those discussions (that I think I'm gonna
probably have to refer back to the underlying maths) - and that maths is all about
connecting Langevin formulations of physics to the kind of calculus that Pearl has
established, in a more statistical sense.
51:04 Friedman:
Thank you for the educational answer! This brings us almost to the end of the .edu
section.
So, I will pass to the final question to be read by Dean, who had several excellent points
and questions.
So, Dean, feel free to ask however you would like.
.edu Question: 'What Is the Difference Between “Subject Matter Expert” and “Prediction
Matter Expert”?'
51:25 Dean:
Good morning!. The question is: what's the difference between a subject matter expert
and a prediction matter expert -- and how does this relate to your mode of interaction?
51:40 Friston:
You're going to have to unpack what "subject" and "prediction matter experts" means,
for me.
51:48 Dean:
For me, you become a "subject matter expert" by gaining a certain amount of
concentration in a particular field or area. And you become a "prediction matter expert"
when you are able to think more distributively, more dispersively. And so, when I read
23

## Page 24

some of the things and listen to some of the stuff that I've heard you talk about – You
brought these two worlds together.
And so, I'm interested in hearing what you think - in terms of introducing some of the
ideas and principles that you brought into a world where, traditionally, we focused on
concentrating -- whether it's materializing something from an engineering perspective,
or deciding what's in and what's out. You've brought in another aspect to look at. And
I'm curious what you think of that.
52:49 Friston:
Okay, that's a fascinating distinction! I'm not sure it's terribly important what I think
about it, because clearly you're the expert on this.
But it certainly would be fascinating to consider the conditions under which you were
able to simulate the emergence of a subject matter versus a prediction matter expert in
silico, for example. This is a proof of principle that these are both effectively Bayes
optimal ways of responding to a particular environment.
And my guess is that you would be able to do that relatively easily, by appealing to the
ideas that you find in applying some of Active Inference notions to structure learning
and development, where the basic idea is:
The 'Prediction Matter Expert' Worldview
53:43 IF you've got a very volatile environment, by which I mean that there's lots of
uncertainty in the contingencies; or possibly there are lots of random fluctuations, that
are irreducible (in terms of your ability to predict the outcome of the trajectory of latent
states of the world in which you are becoming an expert) -- THEN, (formally: in terms of
the precision of various likelihood mappings or probability transition matrices in
discrete state-space generative models) -- when you parameterize your beliefs about that
irreducible uncertainty and volatility, then agents that believe (or have inferred) that they
are in a very volatile, changeable, capricious world usually become better at the
prediction side of things, in the sense that they rely less upon deep past experience, and
assign more precision or more potency to the more recent evidence.
So, they have a different style of evidence accumulation, and also, they have the right
level of uncertainty about what will happen next. So, it looks as if in their predictive
engagement and epistemic foraging in that world, it looks as if they are better at
24

## Page 25

predicting changes -- because they're not committed to a particular explanation or
understanding of how their world works.
The 'Subject Matter Expert' Worldview
55:43 On the other hand, if you create a world which is incredibly predictable and
learnable, then, over time, the natural pressure to minimize free energy translates into a
pressure to minimize complexity, namely a way of modeling your world and your
exchange with it, in the simplest way possible, in accordance with Ockham's principle.
And what that leads to is somebody - it sounds as if - it's somebody who becomes a
subject matter expert. But the subject matter is their lived world, that has now become so
predictable that they do not entertain all possible other outcomes, because they have
precise beliefs about the way that things will unfold, and they can make very wise, very
parsimonious -- or using parsimonious degrees of freedom, they can make moves and
become very expert in the way that this particular non-volatile, predictable (i.e precise)
world works.
Making Our Worlds More Predictable
56:56 And the link with aging here is that: If you allow for the fact that we create our
own environments, and your many levels of Active Inference will permit -- or is a way
of framing -- our eco-niche construction. The story people tend to tell is that: as you get
older, you basically make your world more predictable and you become a subject matter
expert in your own lived world. So, I like the example: I no longer go bungee jumping
nor go to discos, because my world is very, very predictable. And I'm, you know, very
much am expert -- because my world is basically my conservatory, my study and my
bedroom. So, I'm a complete subject expert on that! (Laughs.) You take me out to a
disco, and I will not be able to predict what's going to happen next. Because I'm old.
'Prediction Experts' in the Making
57:47 Whereas, adolescents, and children, and certainly newborn infants -- or newborn
artifacts discovering their world, which is full of uncertainty -- they are not yet subject
experts. And the epistemic pressures or motivation for them to learn about, "What
happens if I do that?" and "What can I control? - What can't I control?" - that will make
them very quickly into prediction experts, until they become sufficiently fluent that they
can now engineer their world to make it non-volatile. And then they presumably will
become subject matter experts.
25

## Page 26

Simulating Cognitive Styles
58:29 So, yeah. I'm sure that would be fairly simple to simulate, using all the toy Active
Inference schemes that we currently use. And it would be really interesting, if these two
different kinds of synthetic agents did develop some cognitive styles and confidence in
what they were doing that looked exactly like the distinction you're talking about! I'm
not sure it would work, but if it does that would be an illuminating proof of principle.
59:02 Friedman:
Thanks for this answer -- and for this session from the Lab and .edu.
Moderator's Recap: 'Maturing Cognitive Styles'
59:06 That last answer really spoke to the importance, also, of intergenerational
learning.
At this point, we will take a break, and we will return for .comms.
26

## Page 27

Session 2 (.comms), Communications
00:01 Hello, welcome back! This is the second session of the ActInfLab symposium on
June 21st 2021.
.comms Organization Unit Overview
00:10 We're here in the .comms, or communication, organizational unit. The goal of
.comms is to organize the Lab's internal projects and activities, as well as to carry out all
forms of communication with external entities. So it's like our connective tissue and our
neuroectoderm, in a way.
What has been done so far, is that we've had about 75 Livestreams, variously on
presentations and participatory discussions, since July 28th, 2020.
We're taking an Active Inference approach to communication, and learning by doing.
Our very first Livestream was on the paper "Narrative as Active Inference," and shortly
after "A World unto itself: Human Communication as Active Inference." That's
something that we like thinking about here, and that we want to explore. Also, some of
our Lab members framed online communication and team collaboration in terms of
Active Inference in the 2020 paper "Active Inference & Behavior Engineering for
27

## Page 28

Teams." Our aim here is to make Active Inference accessible, well known, and well
understood.
So let's get right to the questions, Karl.
.comms Question: 'How can we “show not tell” the idea of Active Inference? (E.g.
through embodied experiences, experiments)'
01:30 The first question is, "How can we show -- not tell -- the idea of Active Inference,
for example through embodied experience, experiments, or what other mechanisms?"
How can we best communicate in a way that makes it resonate?
01:47 Friston:
Well... Pursuing that very interesting notion -- of using the principles of Active Inference
to optimize the role of the .comms team. At its simplest, Active Inference means that the
imperatives for all behavior (and it's likely that most behavior is of an epistemic sort) is
to resolve uncertainty. So if you want to engage people and be a service to the people
you want to engage -- which may be internal members of your own team -- and then
you've got to know, what they don't know. Because, that will define the epistemic
affordances that will get them engaged with you, and you with them, that will incur the
best kind of belief updating.
Embodied Illustrations
02:53 So, practically, what that might mean, is that(, it may well be that... you um…-->
one has to identify didactic or informative illustrations that are tailored, or specialized, to
the person or people that you're talking about.
You know, I have never thought about using embodied experiences before! But that's a
brilliant idea! You're just illustrating to people who want to know "How my body and my
mind works?" -- to illustrate to them the mechanics of it working, using the language of
Active Inference. And it can be very, very powerful!
Sensorimotor Attenuation
03:47 One example of this would be --
Using saccadic suppression as an illustration of the potency of getting your beliefs
about the predictability of sensory evidence right.
If you're a physiologist,you would... you would know this as sensory attenuation.
28

## Page 29

If you were in machine learning, you're working with transformers, this would be, I
think, attentional selection, basically deploying the gain, or switching on the right
channels, in order to select those data that are going to resolve the most uncertainty or
maximize the information gain, subsequent on this sort of covert action, often sold as a
sort of "mental action" in the philosophy literature.
So that mental action is really endemic, and a vital part of our sensory engagement with
the world -- and beautifully illustrated by saccadic suppression. This speaks to the notion
of attenuating the sensory consequences of your own action; so that any evidence that
you're not actually acting is precluded from your belief updating.
Parkinsonian Paralysis
05:14 A clinical example of this would be Parkinson's disease, for example: If I'm
sitting still; and I wasn't able to ignore all the messages from my muscles that tell me,
"at the moment I am sitting still;" then I am never going to be able to realize an a priori
intention that I'm going to initiate a movement -- Because as soon as I initiate a
movement, I have… I put in place a plausible hypothesis that I'm lifting, that I'm going
to stand up. Immediately, all the evidence at hand suggests that I am not standing up, so
I'm going to revise my belief: "No, I'm not standing up; I'm not in the process of
standing up.", And it becomes impossible to move. So that would be an example of what
would happen if you didn't have this capacity to attend to, or to select or apply the
principles of optimal Bayesian design, in terms of selecting those data for your own
belief updating.
More on Saccadic Suppression
06:15 But a really pragmatic and easy example of that is saccadic suppression, when we
do the simplest of movements - epistemic foraging, which is moving our eyes, making
saccadic eye movements. Because, when we do that, we actually induce masses of visual
information on the photoreceptors in the retina, sometimes referred to as retinal slips.
So when I look from the left to the right, there's a flood of information that I have
caused. And yet it is not useful information, because it doesn't tell me anything I didn't
already know. So what the brain does,... it suppresses that information by transiently
suspending the precision -- or the Kalman gain, if you are taking a Kalman filter-like
perspective on predictive coding as one kind of variational filtering).
And that's really easy to demonstrate to an audience. Just get them to either fixate on an
essential stimulus, and then pay attention to something that's moving around. Or the
29

## Page 30

converse: they fixate on the thing that's moving around, and then ask them questions
about the central stimulus. And with the right timing it's a very potent illustration of,
beyond just gathering information, but actively selecting and triaging that information in
accord with the principles of your optimal Bayesian design.
Exemplary Experiments
07:46 So I haven't thought about going beyond that; but I'm sure there are lots of lovely
examples of embodied experiences that really do illustrate Active Inference in action, as
it were.
Ideomotor Theories; Hypnotism
08:04 I'm just reminded, because Active Inference could be read as if you like a 21st
century version of ideomotor theories, which were very popular in the 19th century.
And, of course, that was demonstrated through embodied experiences in a very alluring
way, through hypnotism and the like! So I can imagine somebody doing a sort of 21st
century version of hypnotism and all those wonderful Victorian illusions about the way
you use your sense organs or deploy them actively. But now in the service of just
illustrating some basic phenomena that underwrite Active Inference.
Visual Illusions
08:52 In terms of experiments (or the classic ones that immediately come to mind) that
are really engaging, are visual illusions.
On one reading, all visual illusions are just ways of getting out your perceptual priors in
the context of Bayesian inference. If you can conjure a particular pattern of sensory
information that you know was caused in one way, and yet you think your subject or
your audience has sufficiently precise prior beliefs that it could only be caused in
another way, which is not the way you caused it; and then you let them experience that;
and then you reveal how you actually generated those data - then that's a very powerful
way of demonstrating the innate priors, the sort of formal priors, in terms of the
connectome or the the sparse coupling on a factor graph. And that, again, is part of the
lived experience. So I think visual illusions would be -- and there are loads of beautiful
illusions out there -- and all that one would have to do, is to harness their beauty and
allure; and use them as a vehicle to give people insight into the their own, usually
sub-personal priors about the way that the world is constructed.
30

## Page 31

Simulating Illusions
10:17 And then, in my world, what you generally try and do, is to actually put this in
silico by just, you know, creating little in silico creatures. And because you've now got
Active Inference with this information geometry, you now have the opportunity, not just
to simulate what these creatures do, but what they perceive. Because now you've got the
quantitative estimates of their posterior beliefs. So you can actually show a subject who's
just experienced a visual illusion, that this is perfectly Bayes optimal. And, indeed, when
you write down this variational message-passing scheme in this synthetic subject, this
synthetic person also experiences exactly the same illusions; and this is Bayes optimal,
for this kind of work.
So you can leverage Active Inference activities in that sense.
Focus on .edu's Educational Work
11:16 And deliberately referring back to the .edu discussions - What Active Inference
brings to the table:
Because it's got information about stuff out there, in the numerics -- you can go a lot
further, than you can if you were doing, say, deep learning or machine learning. Because
you've got this information geometry at hand, the state of your variational autoencoder
actually means something in relation to a belief about what generated those data.
You can create lovely little movies, you know, showing what this simulation of you was
actually experiencing.
So, I'm interested: Are there any other ways that you've thought about in terms of
showing people?
Moderator's Recap, 'Developing a Curriculum'
12:07 Friedman:
Thanks for the answer! It's like "look left, look right, now you're an Active Inference
agent!" And, as far as potential avenues for embodiment:
Somatosensory Simulations
12:14 Some of the work with Ryan Smith and others, bringing people into the
somatosensory dimensions and their own priors and expectations about their body and
about motion could be very powerful, as well as auditory modalities. And, indeed Active
Inference is a framework by which we can think about how our perceptions are related
31

## Page 32

to our inference and our action. So, in various domains, I think they'll be excellent
experiments.
.comms Question: 'How Can Active Inference Engage in Better Dialogue with Adjacent
Areas? (E.g. Machine Learning, Systems Engineering, Psychiatry/Neuroscience, What
Other Fields?)'
12:44 And it brings us to our next question. You actually addressed several areas in your
answer. You addressed machine learning, as well as neuroscience, as well as just
everyday lived experience.
How can Active Inference engage in better dialogue with adjacent areas? For example:
Machine Learning, Systems Engineering, psychiatry, and neuroscience, as well as any
other fields that you think are relevant, too?
Active Integration: An Integrative Framework
13:12 Friston:
The obvious answer here is in either academic or commercial collaboration, and what
would license that.
The simplest answer is that the Free Energy Principle and its (if you like) teleological
correlative, Active Inference, and is not there, and was never intended, to replace extant
theories. It was there , to endorse them, and to reveal the interrelationships between
them. So anything that's worked and survived into the 21st century has some veracity
and a proven utility.
And therefore it's just a question of reformulating or changing the words, so that people
can see immediately how their particular formulation relates to somebody else's
formulation, where both formulations are special cases of the most generic and simplest
explanation -- which would from my point of view would be the the Free Energy
Principle, and Active Inference in the case of sentience.
Cross-Disciplinary Interviewing
14:26 So, I think -- As an an integrative framework, you're very well positioned to say:
"Look, can we understand the way that you think about this; and can we now articulate
this, either using simulations or mathematical analysis?
"Can we understand what you've been doing -- in this integrative framework? And if we
can, can we show how it relates to another discipline's formulation of this problem?"
32

## Page 33

And sometimes you can get synergistic or added value from doing that. There must be
loads of examples! You've written "Machine learning, Systems Engineering, Psychiatry,
Neuroscience…" here.
Dialog with Machine-Learning
15:06 So, Machine Learning, for example: "How would Active Inference help Machine
Learning?" At the moment, there seems to be two answers floating around. We've
already discussed a couple of these issues in depth.
Machine Learning commits to (usually) a normative approach to good behavior, that can
be quantified by a loss or a value function. But we've just said, "Well, if we now want
these machines to learn to act, then we have to go beyond state-action value functions,
and consider the belief- based calculus that is Active Inference, which is all about the
reduction of uncertainty."
So now you are in a position to say: "Well, look! If you consider your objective
functions as a part of a more generic "objective function," think what you might be able
to get from this!" And, of course, what you might get from this is a deep learning
scheme, that actually can now go and solicit the right kind of data to optimize its own
learning.
And, you know, people in Bayesian RL [Reinforcement Learning] might argue: "Well!
That's what we're doing" - with a series of bright ideas and heuristics! - "to try and
augment classical value functions!"
But you can say, "Well, okay - You’ve clearly put a lot of work into that! But there is
actually a simple objective function already out there, that is provably appropriate to
describe systems that self-organize and maintain themselves, that actually has what you
want! Why don't you try this?" - for example. So that would be one example.
You have to tread carefully because, you know, a lot of people have dedicated their lives
to solving these problems! And they're very reluctant to change their rhetoric, or see
their contributions as a special case.
But in many instances -- certainly from my perspective, mathematically -- they are
special cases. And sometimes, if you catch the entrepreneurs, the innovators, the creative
academics, at the right stage in their career, before they have committed to a particular
33

## Page 34

church, or ideology, or calculus, or group, or company -- and you can actually point
them in the right direction -- they become extremely creative!
Dialog with Systems Engineering
17:43 And I'm not so sure about Systems Engineering. But, certainly, I always celebrate
the expected free energy as with just taking away various bits and pieces, various
sources of uncertainty, as reducing to KL control. And then what I say is, what KL
[Kullback-Leibler] control is [is] what grown-up engineers use in a control-theoretic
setting! That would be another example.
Dialog on Tolerating 'External Uncertainty'
18:14 You could also (I don't know this, because it's not my field) -- but certainly in
terms of introducing, say, a fault tolerance in control theoretic approaches in
engineering, where the fault tolerance required uncertainty about the operation of some
external part, you could, again, motivate a more complete objective function, that takes
you beyond KL control and introduces the information gain into the mix.
Because to get from the complete objective function to KL control - you have to ignore
uncertainty about the latent states that are in the mapping from latent states of the plant
you're controlling, to the sensors or the observables. So you're moving from a partially
observed Markov decision process (for example), to an observable one; and then
expected free energy becomes KL control - or risk-sensitive control, in economics.
So you could say, "Well, look, why don’t you just augment your KL control, and then
put this extra term in? And now what you've got, is a kind of anticipatory fault tolerance,
in the sense that if there's uncertainty about latent causes, that's automatically resolved,
in the way that you go and switch on various sensors or switch off various sensors. As
you know, there's a principled way of doing that!"
I think to have any influence, you need to be able to show or provide proof of principle,
that this more integrative, more universal, normative approach to problems can offer
speed ups, or increased efficiency, or do what the people actually in that field want it to
do.
So (for example) you've got to be able to show that Active Inference can outperform
"vanilla deep learning" by an order of magnitude. Which is easy to do, because of course
most benchmarks in machine learning are actually inference problems! So if you just
recast it, it's actually quite a trivial thing to do, just by saying: "Well, actually, what
34

## Page 35

you've been dealing with is an inference problem!" -- which looks a lot like one-shot
learning, from the point of view of somebody in machine learning.
But I think there will be some pressure to get people's attention, to make yourself
attractive. They will, first of all, find you interesting; and also have the potential that
they can place an epistemic trust in you. You've got to give them a clue and a cue, as to
why they should engage with you. And very often there's a two-way, or two-road,
exchange.
One simple example of that, which I see emerging in the field, is the use of deep
learning to amortize certain mappings, when they can be amortized in Active Inference
schemes to evaluate the expected free energy, for example, or doing very deep tree
searches.
So that's the kind of innovation you've seen coming out of 20-year-olds at the moment,
who haven't yet decided whether they're going to do deep learning or Active Inference -
because they want to do both, and do it very, very effectively. So that's a nice example,
from my perspective, on the sort of integrative role that could be played - or you could
play.
Moderator's Recap, 'Active Inference Dialogues'
22:15 Friedman:
Thanks! I really heard this "Yes! - And!" maxim from communication and
improvisation. It's like, "Yes, there's been a disciplinary way of approaching it. And
we're going to be working together to come back to first principles, or to make it more
efficient. So that's really powerful!
.comms Question: 'How Does Active Inference Help Us Rethink the Nature of (Online)
Communication?'
22:32 How does Active Inference help us rethink the nature of online communication,
where so much of our communication nowadays does occur?
22:44 Friston:
That's a big question, isn't it! , And, certainly in the context of social media, politics,
fake news, and the like, you could take that question in lots of different directions, which
I won't do because that's not my field of expertise. [Laughs.]
35

## Page 36

'What did you Mean?'
23:01 But just off the cuff, in terms of first principles --
"What is communication?" It's the ability for me to infer what you meant. It's the
hermeneutics problem. If it's a hermeneutics problem, that's most efficiently resolved in
terms of dyadic or multi-system interactions -- when we come back to first principles -
which is the generative model, when we share the same narrative or same generative
model.
'Who Is Talking To Who?'
23:31 , "How does Active Inference help us rethink the nature of online
communication?" Just from a first-principles point of view, it would be the importance
of establishing, "Who is talking to who?" -- and (if you want to optimize the efficiency
of that exchange - literally from the point of view of this principle of least action) - the
speed with which you can resolve uncertainty and minimize your uncertainty or
surprisal.
'Are You Like Me?'
24:03 And then it's ensuring that like-minded communicators are actually
communicating, because it's only them that will understand each other. So everybody has
to speak the same language, they have to commit to a shared narrative, and a shared
generative model. And then by things like rate distortion theorem, or rewriting that in
terms of Active Inference, -- the joint free energy minimization between two
interlocutors -- that's the most efficient sort of shared path of least action.
"How does that help engineer, or intervene on, things?" I'm not so sure. Certainly, just in
reference to communications with people like Maxwell [Ramstead] and other
colleagues, there is this interesting notion that, "If the real problem of communication is
not really the messages that you send, but the inferring whether to send the messages to
this person or not," that itself now becomes conditional upon inferring, "That's a
member of my in-group!," or "That's a creature, or a person, like me!" And then the
question is: "How does self-organization (say, in terms of social media exchange) - how
is that underwritten by an inference about the kind of people who I am listening to, or
who I am talking to? And what are the basic principles of that?"
And again, in accord with the minimization of complexity in our generative models -- It
may be a useful hypothesis to say that "There's an inevitable coarse graining of the way
36

## Page 37

that we conceive of the people that we generate information for (say, on social media),
and reciprocally; and the kinds of people that I will be able to solicit by listening to this
Twitter feed, or that Wikipedia page, or this news channel."
So understanding how people carve up -- the degree of similarity to them -- may be very
useful in just getting an idea of the dynamics of message passing amongst communities
that will be defined by, on average, how each member of that ensemble (or individual)
coarse grains and has a generative model of the kinds of people in the communication
grain.
'Why Always a 50-50 Split?'
26:33 And just to finish this -- which is something I've heard - and I found it really
interesting notion (that, again, would be great, if one could simulate this and understand
the maths behind it) -- is that:
The only evolutionarily stable (from the point of view of the Free Energy Principle) --
the only one that will be selected by a process of Bayesian model selection, the only
partitioning into in-groups and out-groups is a 50/50 in-group/out-group -- in the sense
that anything that departs from that sort of dynamically unstable (but evolutionarily
stable) partition, means that the smaller group, the out-group -- the odd man out -- will
necessarily ultimately be absorbed into the larger group. So the only stable partitioning
is 50-50. Which makes a lot of sense, when you look at Trump versus Biden, when you
look at Brexit versus not- Brexit -- wherever you look, all the important allegiances, in
terms of our political, ideological, and possibly even theological communication, seems
to be split right down the middle. And perhaps it can be no other way.
So it'd be very, very interesting to simulate that, and see if that is a truism that inherits
from all of these marginal likelihood or free-energy minimizing processes, implemented
at multiple hierarchical levels.
You know: Communication is just message passing. And message passing is just the
way you articulate belief updating. And belief updating just is the process of inference;
which just is the paths of least action according to the FEP.
Moderator's Recap: 'Rethinking Online Communication'
28:25 Friedman:
The 50-50 politics, it's maximally confusing! - something we all experience. And a few
key points there about the nature of online communication, is that, at the core, it is
37

## Page 38

dyadic, even when you're broadcasting to many. It's actually about that connection and
the hermeneutic relationship of unpacking meaning.
And then also you brought up the importance of context, and identity, and who's talking
to who, and our inferences about that - which is essential to rhetoric, and something that
often gets left off when people take big-data approaches to online discourse.
.comms Question: 'How Does Active Inference Help Us Think about Science
Communication and Participation?'
29:00 The next question is:
How does Active Inference help us think about science communication and
participation? Specifically, as we move into broader citizen-science initiatives, and as
scientists are in the loop - something you've been recently involved in as well, with
society and with decision making. So:
As science and the nature of science is changing - Who is doing it, and how [do] they
communicate it; how does Active Inference help us navigate that?
29:30 Friston:
Right! Well, I'm sure you've thought about this much more deeply than I have. It's just
drawing upon my experience you know, in terms of science communication during the
coronavirus epidemic.
Yeah - I think you're absolutely right. As with the previous questions, I think you can
take the principles of Active Inference, and just think about, "What does that mean for
optimal communication and belief updating, and shared belief updating, and shared
narratives?" - or not - and use that as a point of reference for the way that you articulate
your own science. And you've asked all the challenging and exactly-right questions,
about how you communicate, how you engage other scientists, or other partners, within
or beyond academia. And I think the same principles apply exactly to the public.
And just to reinforce your beautiful observation that "all communication is dyadic" from
the point of view of the person communicating: So it's this kind of person (as a unitary
object) I am talking to, or this population, or this mentality, on this discipline. So it is, I
think, fundamentally dyadic, from the point of view, the person generating the messages,
that may or may not incur belief updating in the recipients. And these kinds of
principles, I'm sure, would be useful in terms of science communication.
38

## Page 39

So at that level I don't know that there's much that I would have to what you already
know - and will, and possibly already are, implementing.
Active Inference as a framework for thinking about data collection
31:15 There is another level, though, which is using Active Inference not as a model for
the way that we work and communicate and participate, but as more a statistical, or
observation model of data.
So, in a sense, you can use the principles of Active Inference really to make the most of
data pertinent to a particular domain. So again I'm thinking here of the dynamic causal
modeling of the epidemiological and behavioral data that has been generated by the
coronavirus epidemic.
You can certainly use the perceptual inference side of it, (if you like) the Bayesian
filtering side of it; but also in principle the the data mining, or the optimal Bayesian
design, to select which data are useful (or not), in a very practical way when assimilating
big data in the service of understanding the system at hand.
Modeling the Epidemic
32:26 So, if the system at hand is, "How does a spike propagate from one neuron to
another neuron in a neural network?" or "How does a virus propagate from one person to
another person in a population network?" - then you can certainly use the data to apply
Active Inference to build generative models of how you think that occurs.
And what immediately confronts you is, you've got to put in all of the things that
generate those data. So you can't miss out any factors that are important, be they
psychological, be they behavioral, be they viral, be they transport-related -- all of these
things have to go into your generative model to best explain the data.
So, when we do this in a practical way, we use the instance rates from PCR testing, and
Google mobility data, and Department of Transportation data - anything that speaks to
them and reduces uncertainty about all the factors necessary. They're entailed by your
generative model - in this instance, a discrete state space model.
And the Active Inference is not explicitly part of it, in the sense that we're not trying to
predict people's behavior. But it does serve an indirect guide through the principles of
Bayesian optimum design. And all that basically means is: "Do I invest computational
resources, and thereby incur computational and statistical complexity, by including or
39

## Page 40

attending to this kind of data, or not?" And then you can actually evaluate the
information gain by including that data, or that data.
So, for example, "Do you need Google retail estimates and workplace activity, or just
one?" If you include both, that means that the complexity increases; and literally you
have to wait another half hour before you get the results for your dashboard! [Laughs.]
Or - "Do you or do you not have a more parsimonious model?" -- in the same sense of
that saccadic suppression of retinal slip. And you've actually said, "No, I don't need that
-- I've got everything I need. I got the right kind of data, just by focusing on these data."
And then, once you've got that in mind, you can now go foraging for different kinds of
data -- different collaborators from different disciplines, who've got different
perspectives. But also, crucially, different data to try, that will inform and shrink your
uncertainty about the model parameters -- and also, very importantly, about the
structural form of the model: "Do I need this node?" -- "Is this interaction important or
not?" -- "Is this degree of nonlinearity justified by the data?" So all of these questions
affected your hypotheses about how this system is responding, or would respond if you
intervene on it -- all of those questions now become amenable to an evidence-based
analysis, because you've got a generative model underneath the hood.
So that would be a more practical application of the principles that underwrite Active
Inference, even though your computer program is not actually doing Active Inference.
But it's certainly been deployed using the principles of Active Inference
Moderator's Recap: 'Science Communication'
36:19 Friedman:
Awesome! And we heard that integrative approach: "Yes, we're going to include multiple
data sets, potentially of unconventional type; and we're going to have a principled way
of deciding how to include that data". And also as you brought up at the end, who to
include in the conversation. And there was one piece you said in there about the dyadic
nature of communication, where a speaker is always, (I think you said) "speaking to a
person, or to a group, or to a community."
.comms Question: 'How Can We Appropriately Interact With Shared and Nested
Generative Models Across Scales (Person, Team, Community)?'
36:47 And it relates to our next question which is: "How can we appropriately interact
with shared and nested generative models, potentially across, scales -- be it person, team,
40

## Page 41

or community? Do we think about these levels of analysis as Active Inference agents in
their own right? Or how do we, for example, speak to a community or speak to a level of
analysis that's broader than the personal?"
37:17 Friston:
Yeah! I think that's a great and challenging question! Clearly, there has been some
provisional work in academia, looking at sort of "Markov blankets of Markov blankets,
which is effectively, from a stats point of view, from a physics point of view at least,
what we're talking about here.
As a physicist, you'll be tackling this with things like the apparatus of the
renormalization group -- which tells you immediately something interesting: that the
existence of this nested structure, if it is a renormalization group, means that there are
certain functional forms that are conserved.
So what that means is, from your practical point of view, that there will be certain kinds
of behavior that are actually conserved at different scales. So what works in terms of
talking to your children should also work as a president talking to your community, or a
governor talking to your state, or a team leader talking to your assembled team -- Simply
because, in order for there to be a hierarchical nesting that supports that hierarchical
structure, that has to be this conservation, usually mathematically written down as the
functional form of the Lagrangian (or it could be the marginal likelihood, or the surprise
that we're talking about) -- that underwrites these most likely paths, or paths of least
action.
So that, paradoxically, makes the problem slightly simpler! Because, what you're saying
is: What works at one level, will work at all levels - all you've got to do is find the
coarse graining operator that takes you from one level to the next. So what that would
look like, I think, would be very, very application-domain specific. So I think that there
is a great challenge ahead, which is taking the single-particle FEP approach now into a
world where it matters, where the world is actually an ensemble of particles.
Political Physics?
39:45 And we've already discussed the importance of thinking about worlds where all
the particles are identical. Whereas, actually, half the particles vote for Trump and the
other half vote for Biden! And this is interesting to reflect upon! Pre-21st century
physics, that was so powerful in articulating this kind of dynamics, because it just dealt
41

## Page 42

with the simplifying assumption that my idealized gas was an ensemble of identical
particles. And then you can spin up from that equilibrium physics and everything that
led from Carnot cycles and engines, through to current technology. So it's a powerful
assumption, if you just make some simplifying assumptions. But we've already said,
"Well, perhaps that's not the best kind of assumption to make, when you're dealing with
political mechanics!" [Laughs.] You'd at least like one bi-partition in there!
And so that would require a revisiting of that kind of physics from our point of view or
your point of view, basically simulating Active Inference agents, or ensembles of Active
Inference agents, particles, but where now there's a heterogeneity in play; and then
asking the questions: "Well, what are, at the next scale, the free-energy minimizing, or
surprisal-minimizing, or potential-minimizing solutions at the next scale up?"
So we come back to our , "Why is it the case that people are all split 50-50?" -- Which
has an enormous impact on the interactions at the scale below.
So I think, to tease… (I'm just hand waving here, because I don't think there are any any
formal answers.)
And I think those formal answers will probably have to come out of agent-based, and
possibly stochastic agent-based, modeling initiatives. But with the "twist:" You're
making each agent itself an Active Inference agent. So while each individual member of
the ensemble is trying to minimize their free energy; also the ensemble, through
cooperation and the shared narrative, is minimizing the joint free energy. -- And what
that means, when you move from one scale to the next scale.
If you're in physics, I imagine that this is the problem of "beyond non-equilibrium steady
states;" because you're actually now dealing with the multi-scale aspect of
non-equilibria. So at best we have good models of turbulent flow and solenoidal
dynamics in laser physics, that take us beyond equilibrium physics where all the
particles are the same, into non-equilibrium physics. But I don't know that there's an
equivalent maths or metaphor, in physics that would really speak to the hierarchical
nesting. So I think this is a really open and important research area, that I can only
recommend is dealt with by numerical analyses basically predicated on underlying
principles.
42

## Page 43

Moderator's Recap: 'Agreeing to disagree'
43:06 Friedman:
Thanks for the answer there! And it made me wonder if "agree to disagree" is a narrative
that can be shared even when there is a 50-50 split.
.comms Question: 'How Can We Move People and Teams Into the Co-Transformative
Space?'
43:15 And It brings us nicely to the final question of .comms, which is: "How can we
move people and teams into a co-transformative space; or, as some of your recent work
discussed, an interactionist space?"
43:34 Friston:
Well, I'm sure you know the answer to some of these. I'm now realizing you already
know the answers, because your knowing smiles when I say something that you
recognize!
So actually answering that on the basis of what you just said -- I think that's another
really useful insight - that "agreeing to disagree" is a surprise-minimizing, Bayes-optimal
explanation for the exchange with others. But it does rest upon committing to the
hypothesis that "You are not like me; you are not like-minded; and that's okay." So I've
now classified you as somebody who's not like-minded.
And I've resolved the ambiguity among the hypotheses that "You are either like-minded
-- Or you're not like-minded." Normally, we resolve your first impressions within a few
seconds, based upon all these epistemic cues we offer each other to define the sort of
person that we are.
So we make that job as easy as possible for us, and signaling to make this "so we know
our place." And I use that phrase because, of course you know, there's a paper called
"Knowing your Place," that exploited a shared generative model that allowed you to be
in a particular position in some space, even if you and I share the same understanding of
political ideology. But I know my place (because I'm right-wing) and you [know] your
place (because you're left-wing), or vice versa -- and then we can quite happily exchange
but agreeing to disagree.
So I think that that's a wonderful perspective to have, and to endorse it, and that is a
Bayes-optional perspective from both sides of the disagreement. That's resolving
uncertainty, and in a bounded rational way.
43

## Page 44

'Confirming' a 'Commitment'
45:31 So, applying that notion to co-transformative space –
It reminds me of the problem where certain patients in psychiatry have committed to a
particular inference about whether they belong out there, in that kind of environment, or
not; and they have decided that they do not belong out there, and that those people are
not like them; and they start to avoid. (So in a very simple-minded way, if there are any
psychiatrists…) in this example… -- But I think it's illustrative and useful in that respect.
So, take depression or agoraphobia, you know, which is a completely Bayes optimal
response - if I have committed to the hypothesis that out there is full of people who are
not like me; and potentially will upset, confuse, and render me uncertain; and possibly
even injure me in some way. So withdrawal into your house, or into that silo (if you're
working in teams), is a perfectly Bayes-optimal response that says that "You've got a
precise belief that this is where you belong, these are the people that you speak to, and
not those people!" And, that's usually perfectly functional. In psychiatry that would be a
neurotic defense.
But it can become pathological when you become housebound; or say, if you've got a
pathological hypothesis, like your body got dysmorphophobia, and you nearly die
because of a failure to eat properly.
[subSection "It Can Be Another Way"]
47:14 So when you say "co-transformative," I imagine what you mean is, you want to
transform two teams into one team, or at least enable them to work together. If that's
right (and you're nodding partially, so I assume that it is), you're facing the same
challenge that a psychiatrist faces in terms of enabling people to revise the precision of
their precise beliefs about who they can interact with and who they should interact with.
That's not an easy thing, but it's certainly doable. And it usually reduces to presenting
evidence to a group or a person, that it can be another way; so that they start to revise
their prior beliefs, or at least the precision of their prior beliefs, in a safe space where it's
okay to explore other hypotheses, enabling them to think about other ways of
interacting. So this would normally be the objective of psychotherapy, basically by
illustration -- very much in the same way you were talking about, illustrating or
educating by embodied experience.
44

## Page 45

Very much, psychotherapy is thought to work like this. You provide a psychologically
embodied experience where you can try out different styles and different hypotheses.
And in so doing, you paradoxically introduce the right kind of uncertainty, about
different styles of engagement and who you are and who you are talking to. And by
relaxing that precision, you give the patient, or the naughty team that's become too
siloed, the latitude to explore other ways of behaving. So I would imagine that most of
the tried and trusted procedures to get teams into a co-transformative space, use one or
more of those mechanisms.
What would Active Inference bring to the table? It would just bring the narrative that,
"Everybody you're trying to get to talk to each other can come to share;" so they can see,
through the process of becoming more collaborative, or exchanging ideas more fluently,
or working with the same lexicon, or mechanics, or code. Having the same narrative will
actually shape their prior space, and understand the mechanics of actually enlarging the
hypothesis space in terms of interaction styles.
So, that was an incredibly hand-waving answer! But it was, in part, informed by my
understanding of the question from the point of view of the psychiatrist who wants to
transform the way that a patient relates to her world.
Drugs for Treatment…
50:47 Oh - and drugs can help. And I mean that literally! -- those drugs that are
responsible, neurobiologically, for setting the precision. If you can temporarily suspend
the precision in order to reveal other latent a priori hypotheses in terms of the way that I
am, or the way that I interact, and the way that I behave, or the way that I perceive - that
can actually have long-lasting effects, on bringing those other hypotheses to the table in
the moment in subsequent interactions.
Drugs for the Terminal…
51:25 So perhaps the most compelling example of this, which is trending at the moment,
is the use of psilocybin-assisted therapy, particularly in terminal care. If you know
you're going to die of cancer in the next six months, there are certain hypotheses that are
brought to bear in terms of how I would expect to feel, and how I engage with the world,
and how I engage with my loved ones, as a dying person who is near death and the
ultimate loss. Those are not necessarily the best or most functional hypotheses or ways
of being!
45

## Page 46

There are other ways of dying gracefully, and gloriously. But to get at them, sometimes
having a managed challenge to your 5HT-2a receptors via things like psilocybin and
other related drugs just allows you to suspend for a moment your very precise beliefs
about "the kind of thing I am," and allows you to experience other ways of being and
perceiving, which can be very useful when it comes to just trying out other hypotheses,
in this is "your cancer journey."
... Drugs for Teams
52:34 … But, you know, one can also imagine similar scenarios, when you get locked in
to a particular way of interacting either within a team or between teams, in a larger
organization. So that would mean you have to go on a retreat, and take lots of magic
mushrooms.
Moderator's Recap: 'Into Co-Transformative Space'
52:52 Friedman:
Never thought I would hear from you, Professor Friston, but there we have it: "Drugs for
teams!"
Thank you for this excellent interval with .comms unit!
We're going to take another break, and we'll return for the final session - for .tools.
46

## Page 47

Session 3 (.tools)
.Tools Overview
00:01 Welcome back, everyone! This is our third session of the Applied Active
Inference Symposium with Professor Karl Friston, hosted by the Active Inference Lab.
It's June 21st 2021.
We're here representing the .tools organizational unit of the Lab - the third organizational
unit in the Lab.
.Tools Goals
00:22 The goals of .tools is -- To enable effective tool and instrument use for all Active
Inference Lab processes - so that's just using the digital tools affordances that we have
better. -- As well as exploring and designing affordances for our niche, modifying our
niche, resulting in effective action; as well as innovations in tool development.
As with the other groups, we've been meeting weekly in Tools, and having a lot of
awesome insights related to where Active Inference might come into play. And that's
what we're excited to talk to you about.
47

## Page 48

Core Insights from .tools
00:58 Some of the core insights from the work in this unit relates to learning by doing:
The recognition that modern systems are cyberphysical -- everything is really
intercalated with the digital. And also we found it really refreshing, kind of like a
two-stroke engine, to be sidestepping, or complementing, or augmenting some of these
philosophical discussions with technical clarifications.
And two ways in which we've seen that play out:
01:27 Here is a quote from you during a 2019 Dropbox blog post when you wrote that
"Technology is the natural extension of Active Inference beyond the single person;"
48

## Page 49

which, of course, brings technology - far from being something artificial - into the realm
of extended and embedded cognition in our niche.
01:49 And then, a slide from a very recent talk by Bert de Vries on "Beyond Deep
Learning: Natural AI Systems," speaking to several applications in hardware and
software of Active Inference - for example gesture recognition, robotic navigation, and
also audiometry for hearing aids.
49

## Page 50

02:12 And one effort that we're starting up now is a NetHack challenge. It's kind of a
video game played in text characters. And we're assembling a team with already
multiple interested participants to get an Active Inference agent on the playing field so
to speak, and have people maybe update their generative model when they see that it
doesn't have to be a three billion parameter neural network trained for six months on the
GPU -- but what if it's enough to just be curious and to want to succeed? Those are the
kinds of things that motivate us in .tools.
.tools Question: 'How Can We Use Active Inference to Structure the process of
innovation & tool development?'
02:50 We can off with asking: "How can we use Active Inference to structure the
process of innovation and tool development?" And, "How can Active Inference concepts
help us design for complex agents that are interacting in complex niches?" For example,
thinking about niche modification, extension of affordances, reduction of uncertainty, or
structuring of communications.
03:21 Friston:
So: The use of Active Inference to structure the process of innovation and tool
development --
50

## Page 51

NeverEnding Curiosity
03:34 That is, , in itself, an entertaining notion -- in the sense that, you are a realization
of Active Inference! And, you know, I'm mindful that the emphasis on curiosity as the
imperative that drives most of our behavior is exactly the imperative that, as a scientist,
drives me and most of the people I know - and in a sense, I would imagine, also drives
your initiative and your laboratory.
So, all the questions you are asking are really, [Laughs.] "How do I make the next move,
in order to resolve uncertainty about your particular model of how, say, Artificial
Intelligence (A. I.) or human communication is going to evolve?" So, in that light, I
think there are two levels to the answer.
The first one is just to celebrate and acknowledge that you are engaging in the scientific
process as formulated by Active Inference; that you are on a journey of trying to satisfy
curiosity that will be never ending. And that speaks to one of your themes in the
previous slide about "learning is doing." The only way you're going to resolve or sate
that curiosity is to go out there and see what happens. And that is exactly the right thing
to do.
A more practical-level answer, though, speaks to the tool development. Because one of
the fundaments of Active Inference is the appreciation that,
"If you just want to maximize the likelihood that your kind of world model or generative
model (that entails the way that you exchange with and interrogate and ping a world), is
the right world (that is articulated out there, in the sense of extended cognition for
example, in terms of the software tools, or the educational tools that you're making
available) -- Then all of this is still subject to the imperative to minimize complexity.
So, in maximizing the likelihood that these tools will be out there -- and in a sense you're
saying, "This model, this way of narrating the way that the world works -- you provide
an accurate description that is as simple as possible."
So, you cannot escape the complexity - I'm speaking like Jürgen Schmidhuber now –
which is a good thing (in this instance)! That means you've got to find the simplest tools.
It's interesting that you highlighted Bert de Vries's contribution. Because, again, just
practically thinking, "what's the game" here? The game here is to find the best
hypothesis, the best explanation for my lived world -- and my "me" could be Active
Inference Lab's -- and the "lived world" is everything that you have to engage with, in
51

## Page 52

terms of educational, commercial, or academic partners. So, you've got to explore the
model space, to find the right generative model of the way that your system or your
organization works.
Find the Right Generative Model!
07:35 The first steps in writing down the generative model are basically to define its
structure in terms of the sort of hidden factors or latent factors and their interactions, and
all that good stuff.
But it has to be done in the simplest way possible. So, what's the simplest way of writing
down a, generative model? Well, it's to write down a Bayesian graphical model. What
does that mean for the actual coding, practically, and the software schemes and
implementation that you would either offer to people or pre-package in terms of user
interfaces? Then it's going to be message passing on those graphs.
'Pack the Simplest Tool-Kit!'
08:18 I'm trying to get back to Bert de Vries's ForneyLab formation. To my mind,
that's the simplest, most generic bit of computer science that you would come across, in
the service of finding the right software tools -- to build absolutely everything! --
Because:
Absolutely everything can be written down as a generative model! If there's a generative
model there, there's a Bayesian dependency graph. If there's a Bayesian dependency
graph, you know there's a factor graph. If there's a factor graph, then you know there's a
message-passing scheme. What is that message-passing scheme? It's just a Variational
Free Energy-minimizing message-passing scheme.
So I would imagine that, as tool development increases, there will be a move towards a
common language that will look very much like Bert's Forney-style message passing.
And within that, which is a good thing! -- Because that, again, speaks to this
minimization of complexity -- and just course-graining the world, and your world at its
coarsest level that will sustain an accurate account, or a precise account of what you
want to achieve.
the tools just have to come in two flavors. They have to deal with continuous state-space
generative models, to interface with the kind you need for robotics.
52

## Page 53

But the other flavor will be in discrete latent state-space models, that you need to do for,
say, computational linguistics, or modeling the climate in various states.
And we know all the message-passing schemes that would be entailed by a commitment
to one of those two kinds of models, in the sense of generalized Bayesian filtering for
the continuous state-space. And by generalized, you include generalized coordinates of
motion, which generalize things like Kalman filtering.
And on the discrete state-space side, you're talking about either belief propagation or
variational message passing.
So, when you just think about it: What you have to do in providing tools of a software
kind, or a simulation kind -- happily there aren't many choices you have to worry about.
[Chuckles.] So, in that sense, all you need to do is to make sure that your tools
accommodate both generalized Bayesian filtering, and belief propagation and/or
variational message passing, and then you're using off-the-shelf technology.
Which brings us back to, "Well, what's the real problem, then?" Well, the real problem is
writing down the generative model!
How would you unpack those problems, in terms of innovation and tool development?
Well, it's solving the model selection problem. When describing the space of problems
that are faced, say, with Generalized A.I. (or A.G.I., [Artificial General Intelligence])
you can unpack them at different spatial-temporal scales into the inference problem --
into the learning problem -- and into the selection problem -- by which I mean using
Bayesian model selection to get the right structure. You know: "Do I use six, or twelve,
layers in my deep neural network?" "Do I use a convolutional model; or do you use a
transformer?" These are basically problems that are solved, if you have a mechanics that
can score the structure, enabling you to select the right form. So that, I think, is going to
be a focus of innovation -- yeah,--> it already is! But certainly in the near future, in
terms of development, and in the sense that the inference and learning problems - they're
solved problems; that you can just go to Bert and get your favorite message-passing
scheme, or you can keep at the level of your educational or academic message-passing
user MATLAB schemes that we generate here in London for toy problems.
And what is not, I think, a solved problem, and will require an innovative solution, is the
structure learning problem, or the selection problem -- Exploring not "the right
53

## Page 54

hypothesis," but -- we're (in the principled way) exploring the space of generative
models you might want to bring to the table. And that has many, many different issues.
And, things that come to mind are, of course, that you could do it in a bottom-up way by
trying out new hypotheses. Where do you get those from? You get them from experts in
the field, because effectively they are bootstrapping themselves on the basis of our prior
beliefs, or of your knowledge about how something works.
You can do it in a top-down approach by having over-parameterized… over-expressive
models - but with very weak, imprecise parameterizations; and then use Bayesian model
reduction to solve the selection problem. These are ways that people are thinking at the
moment. But this thinking is innovative, because I don't think there are any clear
answers.
So how would you use Active Inference to solve the structure learning problem? Well, in
a sense it's already being used in the sense of Bayesian model selection as "natural
selection;" but you really want to speed that up and make it work within your
commercial or academic lifetimes! But I would imagine that exactly the same principles
would be brought to bear there.
.tools Question: 'How can Active Inference Concepts Help Us Design for Complex
Agents Interacting in Complex Niches?'
14:22 That almost answers the next one, "How can Active Inference concepts help us
design for complex agents interacting in complex niches?" You just have to build these
things as a proof of principle and hypothesis testing!
And the nice thing is, you know all the machinery and the tools that would be requisite
in building these things, right from the variational message passing (using, say,
ForneyLab) through to now!
You've got the right fitness function when it comes to using, say, a genetic algorithm to
explore a structure space. And what is that fitness function? It's the evidence lower
bound or the Variational Free Energy.
So, you've got all the maths in place. This is a question of simulating these things and
providing proof of principle.
How you would translate that into the real world? I don't know at this stage, I'm afraid!
[Laughs.] A challenging first step would be to actually use robotics, or in silico, or sort
54

## Page 55

of hardware, or possibly (a lot of excitement at the moment) using soft robotics. And
actually, you design your niche and see what happens.
And then turn your attention to niche construction -- where you now acknowledge that
the niche itself is also succumbing to the principles - not of Active Inference in and of
itself (in the sense that niches don't plan) - but certainly in the FEP sort of vanilla free
energy minimizing approach.
So, yeah. I haven't actually thought about that before. But that's an interesting
asymmetry, when it comes to simulating multi-agent interactions in the context of niche
construction -- where often it is the case that the niche is just the other agents in an
ensemble.
But if you now actually include the environment as part of the niche - that is playing host
to all the denizens that are the ensemble of Active Inference agents -- then there is this
distinction between the ability to plan the consequences of action, that would entail
optimization of the expected free energy - versus simply reflexively minimizing
surprisal, by minimizing free energy as an evidence bound.
And put that even more simply, more intuitively: You're either with generative models
that support planning, or not. So there's nothing fundamentally different between these
approaches.
If you've got a generative model that is a model of the paths into the future consequent
upon how you act upon the world - that's a much richer, deeper generative model than
the kinds of generative models that would be applicable for a thermostat or an
environment.
And it's likely that the environment that I have in mind here - which is a warehouse, that
you've got a sentient robot going around trying to collect the right things. So the robot
can plan. But the environment, the niche, can't [plan]. It will still conform to the
Variational Free Energy Principle. There will still be particles and things that are
conserved. And they will still fall, and behave in a predictable way. There may even be a
thermostat controlling the temperature - but none of these things are planning.
So, there's an interesting asymmetry that gets into the game when you're talking about
complex agents interacting in complex niches. Part of that complexity has to be a
specification of whether the complexity entails planning or not. And it just creates
different problem spaces , certainly in the context of multi-agent simulations.
55

## Page 56

"carve up" the problem spaces, in terms of, implicitly, problem spaces that will only be
explored by doing. And by "doing" I just mean actually realizing physically these
processes, in the kind of situations that you think are going to be useful for the future.
Moderator's recap: 'When to Select?'
18:51 Friedman:
Thanks for the answer! It's really fascinating about using simulation, so that selection
can happen within the generation of, for example a startup, rather than between
generations. Because, of course, we can let organisms (or startups) proliferate, and then
let pruning occur at the generational scale. Or there could be ways to design so that
selection occurs within a generation, more like learning and development rather than
intergenerational selection.
.tools Question: 'What Areas of Applied Active Inference are Exciting, Promising, or
Important?'
19:21 This could be a broad question; but we're curious: What areas of applied Active
Inference you think just might be exciting, promising, or important?
Active Inference for Computational Psychiatry
19:34 Friston:
My personal usual response to this comes in two flavors. The first is from the point of
view of a theoretical biologist and a psychiatrist. If you can understand how a normal
sentient artifact or person behaves, then that creates a space in which you can think
about false inference and false learning - or certainly suboptimal, from the point of
view of minimizing surprisal or free energy. That's a fancy way of saying understanding
the computational basis of psychopathology. There's a whole literature on using
Active Inference as a normative framework within which to provide an ontology of false
inference , or failures, or aberrant Active Inference.
And why would you want to do that? Well, if it can all be reduced just to the good belief
updating, and the good message passing, we actually have quite a comprehensive
understanding of neuronal message passing, and all its physiology, and all the roles of
various neurotransmitters, and microcircuits and neuroanatomy that underwrite that kind
of neuronal message passing. And implicitly we also then have a fairly fine-grained
understanding of the role of neurotransmitters and the consequences of pharmacological
interventions in the context of experience-dependent learning and an inference of the
kind we've been talking about. So from a translational perspective, literally translating
56

## Page 57

the formalism on offer from Active Inference into the clinical domain - that would be
one motivation for developing this theoretical framework.
Active Inference for Artificial General Intelligence
21:42 The other one is more in the line of technology and artificial general intelligence
(A.G.I.). So then the question is: "Well, I now want to build sentient artifacts -- and not
only build them, but build "brothers and sisters," so they are complex, and interact, and
learn to love each other, in a complex environment that could include me!" And then,
you've got a clear offer from Active Inference as to the design principles you might want
to use to actually build these artifacts.
And then, there are interesting questions about, "What kind of artifact do you want to
build?" And we've already discussed the difference between a thermostat, and a sentient
robot going around collecting your next home delivery. There are different kinds of
generative models. So now you ask the question, "Okay -- what are the exciting and
promising kinds of artifacts, as defined by their generative models, that one might expect
to see in the future?" And then we get into the world of generative models that support
planning, so we're talking about deep generative models where they have a temporal
depth.
Deploying Precision -- Mental Action
23:04 What are the next stages that you might be looking at? Well there's also a sort of
hierarchical depth that would, at some point, first of all include the capacity to deploy
precision. And why is that important? Well, as soon as you have deploying the precision
as a process of inference, you have now a normative theory for this kind of mental
action, or covert action.
One example of this would be -- (I don't know the technology; but I can be assured that I
know what it's trying to do.) –
But thinking about transformer networks, and the way that attentional selection operates
in this context: What you're saying is, you can actually optimize the attention selection
as an inference process, using Active Inference or an evidence lower bound. And where
you're now predicting what things to attend to, and what particular weights to switch on,
and which weights to switch off -- at that point you can understand that as mental action.
So when the transformers or variational autoencoders start to now optimize their
estimates of the posterior precision at lower layers in an auto encoder, it's now acquired
57

## Page 58

the capacity for mental action. And it now will pay attention to various representations,
and possibly even various data sources.
That's not magical! We do that every day in the sort of MDP (Markov Decision Process)
and use it to explain a lot of the attentional mechanisms implemented in the brain. If you
can migrate that technology into deep learning, you would have taken one baby step
towards true sentience, which is mental action.
Meta-Inference
25:02 The next step would be: "Okay, so, how can I now minimize the complexity of my
generative model, where my generative model now actually includes this
'meta-inference?'" - in the sense, "I am now providing predictions about my inference,
because I'm controlling the precision of hierarchically subordinate message passing."
And at that point you start to think, "Well, perhaps one way of simplifying the
computational complexity part of the inference, would be to carve up different states of
attentional deployment!" -- in exactly the same way we're talking about carving up
people into Biden versus Trump voters.
A simple, stable, complexity-minimizing carving up - which suddenly suggests to you
that you can now equip an artifact with states of mind. So that they can be in four states
of mind – they can be happy, they can be sad, they can be confident, they can be unsure.
And they will have to infer, given all the evidence at hand, including the message
passing lower in the hierarchy, what state of mind it is in.
And if you now include in terms of the sensory evidence - you know, the voltage on
their batteries, or some measurement of their interoception - you now have something
that's going very, very close to, say, Ryan [Smith]'s notion of emotions. So, now you've
got - a part of the generative model is now inferring, "What state of mind am I in?" as
the best explanation for all these interoceptive, embodied sensations. Not just the
proprioceptive state of my actuators, but also "Are they getting a bit sticky?" -- "Is
there some wear and tear?" -- "Are my batteries charged?" All of these things come
together as part evidence in conjunction with all the usual visual, radar, acoustic inputs,
to actually supply evidence for a posterior belief "I'm in this state of mind", "I'm
anxious", "my battery's running out."
This immediately creates different prior preferences, cost functions if you like, that
would be applied to your policies, because you've got a deep, changing model that plans
58

## Page 59

into the future. So now you've got an artifact that not only has the capacity for mental
action; it's now got the capacity to be in different emotional states.
The next step is to say, "Hang on! So there are these different states - Can I now equip it
with a minimal selfhood?" "Can the hypothesis that 'I am actually an artifact,' provide
empirical priors that reduce the complexity of my message passing at subordinate levels
-- that is inferring the state of mind that I'm in -- that in turn optimizes the posteriors of
the precisions of various likelihood mappings or preferences over policies?"
Self-Awareness
28:16 So, at this point, you're starting to get to artifacts that could have minimal
self-awareness! The next stage would be, "That's only going to be ever useful, when
you consider dyadic interactions again." Because the only rationale for having
self-awareness, is to disambiguate "self" from "other." Which means that there must be
some confusion, or some uncertainty, at hand - in order to justify the resolution of
uncertainty, justify that complexity of the model! Which means that you have to be
interacting with, or exchanging with, things that are sufficiently like you, to license the
inclusion in your generative model, of a self versus other, or that "You are like me!" or
"Not like me!"
So we actually come back full circle to what we're talking about before, in terms of
inferring "Who am I talking to?" So, I think, this is structurally something quite
fundamental about this inference problem: "Are you a creature like me, or not?" -- or
"Are you like one of those?" -- "Are you a pet?" -- "Are you a plant?"
Just being able to carve up this world in a way that is self-referential, necessarily entails
a minimal selfhood in the inferences of these, that speaks to the importance of getting
the necessary evidence from the environment, that would license that degree of
complexity. And the only kinds of environments that can license that degree of
complexity, are when that environment, that eco-niche, actually comprises other agents
like me - that make it, if you like, worthwhile me inferring, "Oh! It's me, not you, doing
that!"
So I would imagine, then, the most promising applications of active influence in
constructing sentient artifacts, pets and carers or people, things that you can converse
with, would be to grow them -- certainly with themselves, but more importantly with
you there, so they can learn by their doing with you there, so they're curious about you,
59

## Page 60

and you're curious about them. And at that point, one could argue that's the only scenario
which you're going to have any empathetic interaction, with these artifacts.
I'm sure there are other applications in terms of climate change, or commerce, or
whatever. But in terms of imagining what you could produce, what you could sell, I
would imagine that a mindful robot that actually is curious, genuinely curious about you
-- because that will teach you something about itself.
Moderator's Recap, 'Exciting Areas?'
31:19 Friedman:
Thanks for that answer! The idea of tools for attention, and of design and engineering
for "regimes of attention" (to use an Active Inference term), is really essential.
And what you were talking about there with the phone: First off, before the Internet,
when there weren't other devices of similar kind, there was no need to communicate out.
And what we've seen, is that - as there's more and more devices of similar or
interoperable kinds, new levels of organization have to emerge.
Technological Self?
31:47 And then - I thought about the anxiety that a person might feel when their phone is
running low on battery. Right now, that sensor reading is getting emotionally offloaded
to the human. So we could have that anxiety on-device -- so let's have a more relaxing
relationship with our phone! And then, as you pointed out, it would be the incipient steps
of selfhood, or perhaps what they could even call a self-phone (if I'm allowed one pun
per symposium).
.tools Question: 'What Kinds of Tools Have Been Most Helpful in Your Work &
Research?'
32:18 The next question is: "What kinds of tools have been most helpful in your work in
research?" -- which includes many areas such as SPM and DCM, that a lot of people
who are just learning about Active Inference might not be very familiar with. And, what
kinds of tools don't exist yet but might be helpful for Active Inference work?
Mathematical Tools for Active Inference
32:40 Friston:
So, the mathematical tools -- you know, I'm often asked this question of students, "Do I
need to be able to do maths to contribute to this field? - and if so, what kind of maths?" I
won't tell you what my answer is, but what I have found useful is certainly mathematics,
60

## Page 61

but not necessarily very high end; this is always Wikipedia-level mathematics; and in
particular, dynamical systems theory, information theory, and linear algebra are probably
all you need, to do everything, really. And indeed, you could read most of quantum
electrodynamics as basically linear algebra, with a bit of probability theory underneath
it! So that has been the mainstay. If there is one tool, that would be the tool and the
language , of maths - and relatively simple maths.
'Teaching is Learning!'
33:36 The second thing is, the "Learning is Doing!" -- you know, "See one - Do one -
Teach one!" ethos applies, very pragmatically, in this context. Which means, it's very
useful if you can get students to actually build their own little simulated artifacts; and
even more useful when they can actually code it out themselves. Which means you need
access to a high level, at least third generation, programming language that a student can
get fluent with should they want to. Not only to use the existing tools, but try and write it
down themselves without having to spend years training as a computer scientist!
MATLAB
34:24 I found MATLAB very useful in that respect, not because it's terribly efficient
(although I have to say, some of the matrix operators and under-the-hood tensor
operators are much more efficient than people give them credit for, because it actually
came from X-ray crystallography). However, what's really useful about it, is it uses the
same syntax that you would find in a book on linear algebra, which didactically or
educationally is really quite important when it comes to writing and reading the code. So
we have deliberately stuck with MATLAB not because it's computationally efficient, or
that it's open source (it should be - I don't think it is) -- but simply because it's
configured in a way that people reading standard texts, "101" texts and linear algebra
and the like, would be able to see how it transcribes into a computer language. So that's
been a really useful tool.
High-Level Tools for Generative Models
35:21 And looking ahead, I imagine that one's gonna need open access and possibly
more. I'm just thinking about, first of all, people like Bert and ForneyLab, in terms of
very generic, very high-end specifications of message passing in computer science. It
may be that that's the level you want people to actually compose their generative models
and their artifacts. And they don't even need to know about linear algebra, and even less
information theory. What they need to know is the language of the object relations, and
61

## Page 62

how to specify just different classes of exponential probability distributions, and (you
know) "Is it categorical; is it continuous; is it always positive, or can it be positive and
negative?" And that may be quite sufficient to write down a factor graph, or a generative
model. And then everything else is just off the shelf, and "it'll write itself."
So that would certainly be possibly helpful in the future.
.tools Question: 'What Kind of Tools Don't Exist Yet, but Might be Useful for Active
Inference Work?
36:27 I'm moving on to "What kinds of tools don't exist at the moment." So I'm thinking
of -- (I never used it, but I imagine would be) Bert's ForneyLab facilities, but offered as
an application or a user interface, that allowed you to: Compose a generative model;
Compose a generative process (the "actual world" that's going to be modeled); and then
just click "RUN" and see what happens! That would be really useful, I think.
Amortizing Parts of the Inference Process
37:01 Having said that, the other side to "future-scaping" here, is (I repeat) this sort of
leveraging more specialized or other fields; and, you know: Amortizing certain parts of
inference; or Learning to infer; or (indeed) Inferring to learn; or Learning to plan; or
Learning to infer how you plan…
Which Parts of the Inference Process can be Amortized? - Cerebellum…
37:29 -- or, Starting to see, "What parts of the inference process are so conserved, that
they could actually be amortized and learned?" And certainly that looks as how that's
what the brain has done! For example: there are people who think that the cerebellum
has basically "learned" how the motor cortex does its online KL control or Kalman
filtering, and therefore lends a fluency and a computational efficiency to the message
passing -- which, in its absence - it doesn't mean you can't do something; it just means
you can't do it as fluently, and as gracefully, and as quickly as you could with a
cerebellum. Indeed, when you have a cerebellar lesion all that really happens is you
become a bit clumsy and slow.
So, those kinds of tools - a quick and cheerful integration, or importing various
amortization and deep learning technology into a Forney-style message-passing scheme
that could support any kind of generative model, would be really, really useful.
38:51 Friedman:
Awesome, thank you!
62

## Page 63

.tools Question: 'What Kinds of Tools and Platforms Could Inform Trans-Disciplinary,
Highly Contextual, Team Engaged, Participatory, Action Oriented Approaches?'
38:53 Approaching this nexus from another angle: What kinds of tools and platforms
could inform transdisciplinary, highly contextual, and engaged teams that are working
with these approaches? ActInfLab we hope to be working with others to be developing
the Active Inference curriculum, and Body of Knowledge more broadly. But when teams
are actually using these kinds of approaches, what kinds of platforms might exist to
enable their work?
39:28 Friston:
Yeah, okay -- I have a strong suspicion that you know the answer to this! [Laughs.] So
I'm trying to guess at the answer!
We've already talked, implicitly in the way that you presented the ambitions, and
implicitly sent the questions, and all the answers are there. Whether that's trying to
engage through education, whether it's trying to engage through insight, using (say)
embodied experience, illustrations of the basic principles; whether it's supplying games
or graphical user interfaces to facilitate the designing and enacting and the playing with
generative models and Active Inference. -- I think these are all your obvious and
laudable ways of leveraging what Active Inference has to offer .
Tools for participatory approaches?
40:44 Participatory. Yeah. The "Learning is Doing" thing, and the "See one! Teach one!
Do one!" keeps coming back to mind. And the course completely licenses the
participatory aspect. But what kind of participation did you have in mind? Are you
talking about hackathons? Are you talking about playing games with Active Inference?
Computers that start to hate you, or love you, or -- what level of participation were
you…?
41:18 Friedman:
Yeah… Stephen, do you want to give a quick thought on a few kinds of participation --
or what does that mean to you?
Psychodrama
41:24 Sillett:
One area is quite interesting. In psychodrama, they use action methods like action
sociometry, or spatial activities, to look at how people relate to their experience in a
63

## Page 64

dynamic way, so physically. So I've been looking at ways that spatial participatory
approaches can unpack people's relationships to different niches, or different
workplaces, or different types of embodied experience. And then that could be visible, to
be put into Active Inference-type geometries.
42:08 Friston:
I see. Well, there's a great example!
Architecture; Epistemic Affordances
42:12 Two things that I've come across before, are architectural design and the
importance of -- not just pragmatic affordance that says, "Can I walk up?" -- "Can I sit
there?" But also the epistemic affordances: "If I look over there, what would I learn
about the space around me, if I go around that corner?" There is embryonic interest, in
my world, from the architectural sciences and architecture in and of itself, that could in
principle be motivated -- (It's an odd discipline, because it's half like art and half like
science.) But certainly some of their ideas are very much aligned with certain Gibsonian
notions of affordance, and also the affordances, the dual- aspect affordances brought by
expected free energy under Active Inference. So it's not just particular chair?" -- but
also, "What will I learn if I so do?" And so things become epistemically attractive to
engage with.
The Synchronization Manifold
43:32 The other domain is in entertainment and in music; and in particular the joy of
synchronization and mutual predictability or minimizing free energy through mutual
prediction, when singing or dancing together; or indeed interacting (with a slightly
greater asymmetry) in terms of being a member of an audience, watching a band for
example. one of the key things that comes out of that kind of research is ways of
measuring the implicit generalized synchrony that you get from having this information
geometry that I was talking about before, that rests upon there being a synchronization
manifold between the inside and the outside. But if the outside is another inside from
another person's point of view, what you now have is something called a synchronization
manifold. So there's a mathematical image or space, to actually talk about mutual
inference and mutual active inference and engagement and communication - singing
together for example, or diachronically exchanging messages, that does actually
translate mathematically into movement and belief updating on a synchronization
manifold.
64

## Page 65

Measuring Generalized Synchrony
44:55 And that has real-world correlates. You can measure that using kinematic
measurements. You're putting LEDs on people who are dancing together for example, or
measuring heart rate variability or galvanic skin responses, or , doing eye tracking, or
indeed EEG. There's quite a lot of work , in things like hyperscanning, and in ethology,
and dance disciplines, in the arts, in the life sciences - where they do use a lot of these
techniques to quantify the degree of generalized synchrony. What it would be nice to do,
is actually try and model that synchrony, or understand that synchrony, in terms of
movement on the synchronization manifold, which is sort of the mutual belief updating.
Circular Causality
45:53 And one thing which comes out of that, just in discussion if no further, is the
reciprocal [causality], the circular causality, that is necessary to maintain that
generalized synchrony.
The particular synchronization manifold we're talking about, from the point of view of
Active Inference, of course, is mediated across the Markov blanket, as are the active and
sensory states.
But in general, you need to have reciprocal coupling in order to get synchronization. So
directed coupling doesn't work. And if that's true, what that means is that engaging as
an audience, for example, or participating as a spectator, will only really work in terms
of establishing that generalized synchrony that you are chasing, and while you're chasing
it well as soon as you have a generalized synchrony, you've got predictability for free for
all. And that's a good thing, because that minimizes free energy. You know: the more
predictable you can make the world, the better it is, from the point of view of free
energy. But you can only do that if, as a member of the audience, or a witness to
something, you can actually actively intervene on it.
So that brings to mind -- (I'm discussing this with friends of Maxwell [Ramstead]) -- If
you wanted to promote virtual concerts online, for example, during the pandemic --
What you don't have online -- which is what glues things together, things like mosh pits
in carnivals and festivals -- is you don't have the audience participation - the applause,
the roars, the lighter waving, or the light waving.
65

## Page 66

So how would you get that back into a virtual experience? Because that would be
absolutely essential, to actually engage people! Otherwise, you'll be just looking at a pop
concert on television. So:
From Revealed Audiences --
48:08 More than just revealing the underlying correlates of that generalized synchrony
in terms of the EEG traces of the dancers, or doing some sensory mapping from their
motion to auditory input, just making the sensory evidence that supports the mutual
inference more precise and more available - just by having it displayed, say, by putting
motion in sound or sound in motion, or EEG, electroencephalographic, measures of
performance, or the audience, visualizing that (and that has been done by people like
Paul Verschure in Barcelona) --
-- To Empowered Audiences
48:54 -- more than that: To actually enable the audience to change what the performers
are doing - - or perhaps what other members of the audience are doing -- you have to
empower them to close that circular causality, to get that dynamical-coupling play, so
you get the right kind of generalized synchrony.
you know that--> That sort of dynamical-systems perspective on synchronization and
free energy minimization certainly speaks to a particular kind of participation and
engagement, that does indeed rest upon action-oriented approaches. But crucially, it's the
action of the audience on the performers, not the performers' action on the audience, that
is usually what you need to pay more attention. Was that the kind of thing you were
thinking about?
49:52 Sillett:
Yeah! - that's really a useful answer! We were thinking about that, and some
participatory immersive theater type events, and other participation in collective
meaning making. So that's the type of thing that we're looking at.
Community Power'
50:07 Friedman:
And it reminds me of the Livestream affordance, which is relatively novel, but allows
people to be asking questions. And it enables not just efficient production of material in
a one-shot approach, but it allows the feedback. And I can't help but add, that it's that
affordance for participation, for example "speak now or forever hold your peace", that
66

## Page 67

expands the wedding into the community, because there is the opportunity for feedback.
It's not just a breakaway clique - it's actually something that remains integrated through
the affordance for participation.
.tools Question: 'How Might Future Modeling Involve Large Scale Patterns in Social
Data Sets, and Working Back to Infer Hidden Causes? (E.g. Pandemic Modeling,
Governance)'
50:40 (So) I'll turn to the last question for this section: "How might future modeling
involve large-scale patterns in social data sets, and working backwards to infer their
hidden causes, for example in the case of pandemic modeling, governance, economic,
other situations?"
51:03 Friston:
Well, this is a very practical and very prescient question, because of course a lot of
people are asking themselves that now, specifically with respect to pandemic models;
but also the people who are exercised and have the interventional clout when it comes to
COVID, are generally also the people who are invested in climate change problems as
well. So there's a lot of noise out there at the moment about how we can harness the data
assimilation and modeling advances made during COVID-19 and keep the momentum
up to tackle climate change - and not just climate, but the economic structures, and
financial structures, and informational structures, that are deeply interwoven in terms of
climate change.
My answer is going to be somewhat deflationary. I've had this kind of conversation
before, again with Maxwell [Ramstead] and John Clippinger and Kim Jones and related
friends; and I'm due to have another conversation with him on Open World or (I can't
remember), in the near future.
There's a temptation to take all the "high church" of the Free Energy Principle and
Active Inference, and epistemic foraging, and all of that good stuff we were just talking
about, and say "Oh, well, now let's make it work in terms of understanding (say) the
pandemic!" And you don't need to do that. All you need to do is to apply the good,
scientific principles that things like Active Inference appeal to, to the problem at hand.
And it all comes back to the generative model.
So, all you're saying here is, "How might future modeling involve large-scale patterns of
social data to infer the hidden causes?" -- is just a statement of, "We need the right
generative models to make proper sense of the big data at hand!" And in saying "the
67

## Page 68

right generative models," we need the equipment both to invert those models (in the
sense of inferring the parameters' interactions), using the simple tools we've just talked
about. They will just be lifting it from the laboratory, or continuing to use MATLAB.
The Structure Learning Problem
53:37 But the bigger problem is what we talked about, which is the selection of the
structure learning problem. This goes beyond just "How many layers do I have in my
deep network?" Much more important, I think, it's a factorization - it's knowing, "How
many conditionally independent factors do I need to minimize the complexity - to get
the right granularity - the right way of carving up the latent causes behind all the data
that is available to me?"
So I think the pandemic modeling is a beautiful example of this, because: The factors
that determine whether I infect you can certainly be written down in terms of virology
and the ACE receptors, ACE-2 receptors, and base reproduction numbers, and
transmission strengths, and transmission risks, and the spike proteins - but that's only
half the story.
The Social Sides of Disease Modeling
54:42 The other half of the story is, "How likely are you to be at work, or at home, when
I'm at work? Are you likely to be wearing a face mask? Are we going to be one or two
meters apart?" So all these behavioral aspects start to become really important factors.
And even beyond that: When it comes to making sense of the model, the likelihood part
of the model that actually generates the data, can become extremely difficult to optimize
when you start to think about, "What kind of data is at hand?" For example, just
notification rates of new cases per day of coronavirus. Now you might think, "Oh, that's
really great data." It's really difficult data to handle, because the different kinds of tests
not only have differential false positive and false negative rates; but the different ways in
which they are deployed, really compounds that in terms of the selection bias. So: "Are
you testing people who are symptomatic? - What's the probability of being affected if
you're symptomatic?" - Are you not? - Are you doing survey testing? - Are you doing
the same amount of testing this week as you were doing last week?" All of these - what
would be from an epidemiological or a behavioral science perspective really
un-interesting factors, suddenly now become the most important factors in making sense
of those data! But you only know that when you start to do the model comparison, the
structure learning; when you actually commit to writing down the co-generative models
68

## Page 69

And that's certainly what I've learned over the past year, now coming up for a year and a
half.
The Future of Modeling
56:19 The future of modeling -- First of all, it's obvious what the future is: It's just
basically writing down the right kind of dynamical state-space models that account for
data. But the future is really dealing with the problems of structure learning and model
selection for any data, but in particular from the big data at hand in terms of pandemics
or trafficking on the web, or climate change. So it's a really exciting opportunity!
Why do people want to do it? Well, once you've got the most evidenced (i.e the
minimum free energy) model at hand, and you've got posteriors over all the model
parameters and all the right interactions, then you can do all sorts of stuff in terms of
reducing people's uncertainty about the future. Because you've quantified the
uncertainty, and explained to them things that were once uncertain about and what isn't
uncertain about. That has enormous implications for mental health and well-being; and
possibly even feeding back into finance, because you always hear, "Well, the biggest
determinant in terms of the markets is the market confidence. It's all about the
uncertainty!" So if you can do uncertainty quantification in a principled way, using using
this kind of modeling, you've done a big thing already.
'Interventions?'"
57:56 But then you come to monitoring putative interventions! You've now got a direct
handle: posterior estimate on the latent states you actually want to make decisions on. So
it's not the notification rates or the number of new cases in California today, it's a
number of new people that have become infected today. And that's a very difficult thing
to infer given all of these complicated aspects of the generative model. And then, of
course, once you've established the validity of this model, in terms of its construct and
predictability, then you can intervene on it. Then you can say, "Well, what would happen
if I changed this? Or what would happen if I changed that?" And, "What would happen
now? - What would happen in the future?"
Meteorology Beyond the Weather
58:42 So that, you know, you're suddenly in a world of quantitative modeling, where
you can start to ask some very powerful questions, and also share with everybody who
matters, the products of your inference. So you can now start to think about
69

## Page 70

supplementing the weather forecast with an epidemic forecast, you know, "The virus in
your area: And tomorrow we expect…!" You know, you can also do that for the markets.
And these kinds of things, I think, are going to be more important when people, or when
the current generation (your generation, I guess!) start to wrestle more with climate
change, because they're going to want to - not just know whether it's going to rain
tomorrow; they're going to want to know, at the level - not just the weather, but, "The
climate - what are the indicators?" Because those indicators really contextualize, and
inform their generative models about their place in the world, and that global scale. But
to provide that kind of weather forecasting, that meteorology beyond the weather, you're
going to need to have these state-space models properly optimized, and in a
first-principle way in relation to their marginal likelihood on their evidence bounds.
[subSection "Future Modeling for Governance"]
60:12 And "governance." Governance is just policy decision-making based upon
counterfactual outcomes. So that is always underwritten by these Bayesian beliefs. But
you can't get the Bayesian beliefs, unless you've got a generative model. And that has
the consequences of action in the future. There would be also interventions, either
politically, or financially, or or otherwise.
Final words from Active Inference Lab
60:38 Friedman:
Thank you so much again for joining this Symposium! It was really a special moment
for the Lab, and we look forward to continued interaction.
Thanks for everyone who's watching, and we hope that you participate in ActInfLab.
70

## Page 71

Supplemental Lists:
1. Subjects
2. References
3. Mentioned Names & People
4. Other Resources
71

## Page 72

1. Subjects
4E Embedded, Extended, Embodied, Enactive
5HT-2a receptors
a priori hypothesis
action-value functions
attract
attracting set
Active Inference
Active Inference framework
Active Inference Lab
active learning
active perception
active sensing
agent
ambiguity
amortize
attentional selection
Bayesian belief updating
Bayes optimal
Bayesian mechanics
Bayesian Reinforcement Learning
belief-based schemes
bounded rationality
causal coupling
circular causal
co-transformative space
death
deep learning
coarse graining
continuous state-space
directed coupling
discrete state-space
dyadic
dyadic interaction
dysmorphophobia
enactive perception
72

## Page 73

epistemic foraging
epistemic trust
ethology
Expected Free Energy
false inference
false learning
factor graph
fault tolerance
ForneyLab
Free Energy Principle
gain
generative model
hierarchical depth
hierarchical nesting
hermeneutics
hyperscanning
ideomotor
in silico
infer
information gain
information geometry
integrative framework
interdisciplinarity
Kullback-Leibler (KL) Control
language
least action
lexicon
loss function
machine learning
marginal likelihood
Markov blanket
Markov decision process
mental action
metaphor
mindful
neural network
neuroectoderm
73

## Page 74

normative theory of sentient behavior
objective function
objective functional
optimal Bayesian design
optimization
partition
pathological hypothesis
perceptual priors
partially observed Markov decision process (POMPD)
passive inference
plausible hypothesis
posterior belief
probability transition matrix
process
psilocybin
psychodrama
psychotherapy
principle of least action
process theory
pullback attractors
purpose
reinforcement learning
renormalization
retinal slips
robot
risk-sensitive control
saccadic eye movement
saccadic suppression
self-awareness
self-information
self-organization
self-referentiality
sensory modality
sentience
separation of states
Shannon information
situated cognition
74

## Page 75

soft robotics
somatosensory dimension
sparse (dynamical) coupling
state-action value function
state-space
statistical manifold
sub-personal prior
surprisal
surprise
synchronization manifold
synergy
teleology
trajectory
transformer
vanilla free energy minimization
variational autoencoder
Variational Free Energy
variational message passing
variational principle
variational principle of least action
visual illusion
75

## Page 76

2. References
Aumann, Robert J. 1976. “Agreeing to Disagree.” The Annals of Statistics 4 (6): 1236–39.
http://www.math.huji.ac.il/raumann/pdf/Agreeing%20to%20Disagree.pdf
Bouizegarene, Nabil ; Maxwell J. D. Ramstead; Axel Constant; Karl J. Friston; Laurence J. Kirmayer (2020)
"Narrative as Active Inference: an integrative account of the functions of narratives."
https://www.researchgate.net/publication/342828597_Narrative_as_active_inference
Constant, Axel; Andy Clark; Michael Kirchhoff; Karl J. Friston (2020) "Extended Active Inference: Constructing
predictive cognition beyond skulls." https://doi.org/10.1111/mila.12330. Accessed 2021-12-14 at
https://onlinelibrary.wiley.com/doi/10.1111/mila.12330
Engel, A.K.; Friston, K.J.; Kragic, D. (2016). The pragmatic turn: Toward action-oriented views in cognitive
science. LCCN 2015048280. ISBN 9780262034326.
Engel, A.; Maye, A.; Kurthen, M.; & Konig, P. (2013). Where’s the action? The pragmatic turn in cognitive
science. Trends in Cognitive Sciences, 17(5), 202-209. doi:10.1016/j.tics.2013.03.006
Forney, G. David (2001) 'Codes on graphs: Normal realizations.'IEEE Transactions on Information Theory 47.2
(2001): 520-548.
Forney, G. David (2012) 'Codes on Graphs: Fundamentals.' Accessed 2021-11-28 at
https://arxiv.org/abs/1306.6264
Hartley, Matthew; Neill Taylor; John Taylor (2005) Knowing your place: Subfield specific involvement in
hippocampal spatial processing. In Neural Networks 2005.
academia.edu/776884/Knowing_your_place_Subfield_specific_involvement_in_hippocampal_spatial_processing
Smith, Ryan; Karl J. Friston2; Christopher J. Whyte (2020) "A Step-by-Step Tutorial on Active Inference and its
Application to Empirical Data." 
Vasil, Jared; Paul B. Badcock; Axel Constant; Karl Friston; Maxwell J. D. Ramstead (2020) "A World unto Itself:
Human Communication as Active Inference." Accessed at https://arxiv.org/abs/1906.10538
Vyatkin, Alexander, Metelkin, Ivan, Mikhailova, Alexandra, Cordes, RJ, & Friedman, Daniel Ari. (2020). Active
Inference & Behavior Engineering for Teams (Version 1). Zenodo. 
https://doi.org/10.5281/zenodo.4021163
76

## Page 77

3. Mentioned Names & People
John Henry Clippinger
MIT Media Lab
https://spatialwebfoundation.org/
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=824432
Bert de Vries
https://biaslab.github.io/
https://kdvsfoundation.org/medical__board/dr-bert-de-vries-md-phd/
G. David Forney
https://ieeexplore.ieee.org/author/38558627400
David MacKay
http://www.inference.phy.cam.ac.uk/mackay/
https://www.inference.org.uk/is/
A. A. Markov (1856-1922) (Andrei Andreyevich Markov "senior")
Beren Millidge
Judea Pearl (b. 1936)
Maxwell J. D. Ramstead, PhD, University College, London
https://www.researchgate.net/profile/Maxwell-Ramstead
Claude E. Shannon (1916-2001)
Ryan Smith, PhD
https://www.laureateinstitute.org/ryan-smith.html
Paul Verschure
https://specs-lab.com/portfolio-items/staff-paul-verschure/
Wanja Wiese
https://predictive-mind.net/ppp-contributors/Wanja_Wiese
77

## Page 78

4. Other Resources
Blake, A. and Yuille, A. (Eds.) (1992) Active vision,
Cambridge, Mass.: MIT Press. 
Churchland, P., V. Ramachandran, and T. Sejnowski. 1994. A critique of pure vision. (In Large-scale
neuronal theories of the brain, ed. C. Koch and J. Davis. Cambridge, MA: MIT Press.) Accessed
2021-12-12 at https://cogsci.ucsd.edu/~nunez/COGS1_F07/Sejnowski_rdg.pdf
Clippinger, John Henry ?(1999) ed. The Biology Of Business: De-Coding The Natural Laws Of
Enterprise. https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=824432
ForneyLab
https://biaslab.github.io/project/forneylab/
Friston, Karl; Conor Heins; Kai Ueltzhöffer; Lancelot Da Costa; Thomas Parr (2021) "Stochastic Chaos
and Markov Blankets." https://doi.org/10.3390/e23091220
Karl Friston, Thomas FitzGerald, Francesco Rigoli, Philipp Schwartenbeck, and Giovanni Pezzulo
(2020) Active Inference: A process theory. Neural Computation, 29(1):1--49, 2017.
https://www.fil.ion.ucl.ac.uk/~karl/Active%20Inference%20A%20Process%20Theory.pdf
Friston, Karl J. et al. (2020) Dynamic Causal Modelling of COVID-19.
https://www.fil.ion.ucl.ac.uk/spm/covid-19/
"Links"
"Cybernetics is the science of defensible metaphors." – Gordon Pask (1975), Conversation Theory.
"David MacKay's work"
Accessed 2021-12-12 at http://www.cs.toronto.edu/~jessebett/CSC412/content/week1/lecture1.pdf
"Research Debt"
https://80000hours.org/podcast/episodes/chris-olah-unconventional-career-path/
78

## Page 79

Hohwy, J. (2018). The predictive processing hypothesis. The Oxford handbook of 4E cognition,
129-146.
Hohwy, Jakob (2014) The Predictive Mind
"No real latitude for interpretation"
Leibniz, G.F.W. (1678) "Nature, proposing some end to itself, chooses the optimal means." in
"Definitiones Cogitationesque Metaphysicae"
Lilly, John C; Antonietta Lilly (1978) The dyadic cyclone : the autobiography of a couple. Accessed
online 2021-12-15 at https://archive.org/details/dyadiccycloneaut00lillrich
Millidge, Beren (2021) Applications of the Free Energy Principle to Machine Learning and
Neuroscience. arXiv:2107.00140v1 [cs.AI] 30 Jun 2021. https://arxiv.org/abs/2107.00140
Newen, Albert; Leon De Bruin; Shaun Gallagher (eds.) (2018) The Oxford handbook of 4E
cognition
Noë, A. (2004). Action in Perception. Cambridge, Mass.: MIT Press
Papo, David et al (2014) Complex network theory and the brain.
https://royalsocietypublishing.org/doi/pdf/10.1098/rstb.2013.0520
Pezzulo, G.; Donnarumma, F.; Iodice, P.; Maisto, D.; & Stoianov, I. (2017). Model-based
approaches to active perception and control. Entropy, 19(6), 266. https://doi.org/10.3390/. Accessed
2021-12-12 at https://www.mdpi.com/1099-4300/19/6/266
Schwartenbeck, Philipp; Johannes Passecker; Tobias U Hauser; Thomas HB FitzGerald; Martin
Kronbichler; Karl J Friston (2018) Computational mechanisms of curiosity and goal-directed
exploration. DOI: https://doi.org/10.7554/eLife.41703
79

## Page 80

Sejnowski, Terrence J. 1994. A critique of pure vision.
https://cogsci.ucsd.edu/~nunez/COGS1_F07/Sejnowski_rdg.pdf. Also see
https://cogsci.ucsd.edu/~nunez/COGS1_F07/Sejnowski_slds.pdf
Sutton, Richard S., & Andrew G Barto (2018) Reinforcement Learning: An Introduction, 2nd Ed.
https://mitpress.mit.edu/books/reinforcement-learning-second-edition
Varnes, Erich W. (2004) Noether's Theorem. Accessed 2021-12-14 at
http://www.physics.arizona.edu/~varnes/Teaching/321Fall2004/Notes/Lecture14.pdf
van de Laar, Thijs (2018) Julia Toolbox for Factor Graph-based Probabilistic Programming.
https://www.youtube.com/watch?v=RS4hJ4oBr9c
Wesolowski, Lukasz et al (2014) TRAM - Optimizing Fine-grained Communication with Topological
Routing and Aggregation of Messages. http://charm.cs.illinois.edu/newPapers/14-18/paper.pdf
80


---
*Extraction method: pymupdf*