Active Inference (Prof. K.Friston)
ฝัง
- เผยแพร่เมื่อ 21 ธ.ค. 2022
- IEEE ASI Webinar: Active Inference
Prof. Karl J. Friston MB, BS, MA, MRCPsych, FMedSci, FRSB, FRS
Wellcome Principal Fellow
Scientific Director: Wellcome Trust Centre for Neuroimaging
Institute of Neurology, UCL
12 Queen Square
London. WC1N 3BG UK
In the cognitive neurosciences and machine learning, we have formal ways of understanding and characterising perception and decision-making; however, the approaches appear very different: current formulations of perceptual synthesis call on theories like predictive coding and Bayesian brain hypothesis. Conversely, formulations of decision-making and choice behaviour often appeal to reinforcement learning and the Bellman optimality principle. On the one hand, the brain seems to be in the game of optimising beliefs about how its sensations are caused; while, on the other hand, our choices and decisions appear to be governed by value functions and reward. Are these formulations irreconcilable, or is there some underlying imperative that renders perceptual inference and decision-making two sides of the same coin.
The applicability of the FEP for psychotherapy, training of psychotherapists is staggering. I'm glad that I've stumbled across this now.
Thank you!
Brilliant
from: th-cam.com/video/Fwa82X7tkUI/w-d-xo.html
the kind of systems that we are
7:10
interested in show this characteristic property that they have characteristic
7:15
states which they keep on revisiting you can describe this um in terms of inland
7:22
dynamical systems a pullback attractor you can articulate this in terms of physics in terms of non-equilibrium
7:28
steady state the key thing is to be something is to have characteristic
7:34
states that I return to or at least the neighborhood of and thought of like that
7:40
we can now read this object as the probability that you'll find me in any particular state if you sample me at any
7:46
random time and that's important it's important because we know a lot
7:52
about the maths of the relationship between the Dynamics the flow the
7:57
amplitude of the random fluctuations and this description of the space system in terms of its characteristic States or
8:04
its pullback attractor
INFORMATION :
other words in order to counter the random fluctuations I have to be flowing towards my artistic States in order to
9:12
stop the probability density changing and that's the key behavior that I want to pursue for the rest of the talk and
9:20
so this point let's put the remark of banking back in play and write down that
9:25
gradient flow in terms of the amplitude the random fluctuations these things denote circular or solenoidal flows that
9:32
give this kind of itinerancy and life cycles and oscillations and it what what the Markov blanket
9:38
tells you is that it is subject to the same law and it means that the internal States
9:45
and the active States will look as if they're trying to increase the log probability of in this instance the
9:52
sensory part of the Markov blanket and I'm going to interpret that in terms of perception and action respectively and
9:59
just ask the question how would I then interpret this quantity here well we've just said that the
10:06
states that I'm most likely to occupy are those are the characteristic of me they are literally the states that
10:12
constitute my attracting set to which I am attracting so they are valuable for me they have meaning for me
10:19
um denoted by them here um so one could read this log probability just as value and what could
10:26
spin off reinforcement learning if you're an engineer optimal control theory if you're an economist expected utility Theory
10:33
um if you're a free energy theorist the negative variational free energy is just a way of writing down this valued
10:39
function um if I just multiply this by minus one we have a complementary perspective that
10:45
people in information Theory will recognize so this is now known as this
10:51
negative log probability that are now looking as if I'm trying to minimize is
10:56
known as self-information information ethereum or simply surprising or surprise it's just a measure of the
11:03
implaus ability that I would sense this given I am me and this is the quantity
11:08
that is bounded by the free energy leading to things like the informatics principle the principle of minimum redundancy and indeed the free energy
11:15
principle this is nice because the average of this thing is known as entropy so the
11:21
expected free energy or self-information is entropy so it'll look as if I'm
11:27
trying to resist the second law by minimizing the dispersion or the entropy
11:32
of my sensed States and of course that's the Holy Grail of self-organization in
11:37
physics and synergetics of the kind described by Hermann haken and indeed if I was a physiologist it would just be a
11:44
statement of homeostasis and I have to to exist I just have to keep my sensed
11:49
physiological States within some viable um bounds that are existentially
11:54
consistent and have meaning for me because they are characteristic states that I occupy
27:53
error that comes from the muscles in my eye that I could predict and I could use a prediction error to infer where my eye
27:59
was currently pointing but there's a much simpler way that I can minimize these prediction errors I can just
28:05
change the stretch of my muscle to match the predictions of the stretch receptors
28:12
and what I'm describing here is it's a classical reflex arc in uh motor control if I was doing interceptive inference
28:19
this would be an autonomic reflex basically actively and reflexively
28:24
um minimizing prediction errors in relation to deeply hierarchically informed predictions that are generated
28:31
by my model but I'm trying to maximize the evidence for um so that's
28:38
um the basic story um I just want to finish the story
28:44
um with saying well actually not quite um before I do that let me just illustrate
28:51
the kind of active interest and the kind of sort of Engagement with the world and one way of establishing a very simple
28:57
synchrony with the world that inherits from the architecture I've just shown you
29:02
um what we did here is basically equip a synthetic subject with a a generative
29:08
model that had Dynamics autonomy at this kind of itinerancy that was implicitly
29:14
referred to yesterday in the form of a central patent generator and then map
29:19
this abstract Dynamic to some point in extra personal space and we told the
29:26
synthetic subject or part of its generative model was that there was an
31:48
the external States and the active States now become causes of my Sensations so now I have a model
31:57
of the causes of my Sensations that include my own actions and it's this
32:02
particular move um that I'm um suggesting or from the
32:07
point of view of simulating these kinds of uh this kind of self-organization may be a necessary move to introduce agency
32:15
in the sense that uh agents will have some notion some oils can be described
32:21
as having some sense of their own agency the consequences of their own action
32:26
because the um their own action is now not directly accessible it can only be
32:34
observed by the sensory consequences that therefore have to be modeled and
32:39
this takes us into a slightly different and um I think richer world where you
32:44
now have the notion of a model that incorporates explicitly the consequences
32:51
of the agent's action I'm not saying at this stage there's any awareness or sentence of that agency but just it is
32:58
there that is then used to plan into the future um to evaluate the different
33:04
consequences of different actions so that you are now implicitly introducing the notion of planning and the temporal
33:10
depth into the Dynamics that then provide the predictions of what I'm