This is a nice explanation. Good, multi-level prediction of likely sensory inputs followed by searches for unexpected deltas is an extremely powerful data reduction technique.
Could the FEP formalism just as easily be interpreted as saying that an organism is an environment’s model of itself? This seems plausible particularly if we are going fully embodied and enactive with our FEP by rejecting the Cartesian idea of internal mental representations.
To flesh this out a bit: Organisms embody generative models that correlate with the statistical regularities of their environment, with evolution and development sculpting them so that their very physiology down to the chemical level embodies the constraints and patterns present in their earthly and heavenly surroundings. Rather than serving as a separate internal model, an organism’s structure and behavior are deeply entangled with environmental dynamics, effectively making them a natural instantiation of the environment’s own structure.
Have you read Laukonnen & Chandaria’s paper “A beautiful loop”??? Relationality and recursivity being key aspects of “self-ing” within a FEP framework.
I’m so glad we’ve got your mind to help with this Matt. 🙏🏽❤️
I think the links between Laukonnen's work on meditation (including some his earlier work, e.g. "From Many to (N)One") and the FEP framework provides a very intuitively fitting way of conceptualizing the way that mystical traditions have long encouraged folks to experience existence. The paper with Chandaria is a fun extrapolation of those ideas.
I love Ruben. We share many of the same obsessions. And I think his Eureka heuristic is a really important contribution to cog sci. I don’t quite share his confidence in the FEP perspective but he’s pretty convincing (and he actually gets the maths, which I do not 🤪).
Nice! I tend to be interested in the links between various psychological and philosophical models and various mystical traditions. To that end, Laukkonen’s Eureka heuristic seems like it may be in a similar spirit to how Andrew Newberg relates moments of “little e” enlightenment to the “Big E” in some of his research/writing.
How does agency play a role in this theory? Does this imply that we take actions which will minimize our internal surprise? Are there other motivating mechanisms in the many layered thing that is our brain? I assume so, but I'm not sure how to think of the interplay between FEP and, say, value systems. How many conceptual layers exist in between such ideas..
EDIT: one of the conclusions of this view of agency is that I would do best to sit in a sensory deprivation tank until I die. I can predict that future quite well! But we don't, there are obviously other things at play as well. That's not to say this theory is wrong, but it does suggest that this is just one aspect of the overall picture.
You're exactly right. I think that's the case, we take actions expected to minimize future surprise. We want to test our hypotheses in controlled ways. You might like Jürgen Schmidhuber's ideas as well. He talks about the drive of all intelligent agents, compressing experience. Value systems come from our need for survival.
RE: Jürgen Schmidhuber's ideas. Will do! These are new to me. I'll look into it :)
EDIT: I don't see much about agency linked from his Wikipedia page. Can you explain further which ideas you think are interesting? I found on his home page the sentence "Schmidhuber's curious learning agents like to go where they _expect_ to learn something" which is quite interesting, but this is also a different claim than the FEP.
From the above link you can find his actual paper. In the video I go through the interesting parts.
I interpret FEP through machine learning which I'm more familiar with. I could explain this a lot better by drawing diagrams. The key is also the difference in probability distributions as Curt outlined, but hey equations are tough for almost everyone.
FEP talks about surprise. Schmidhuber talks about curiosity. Perhaps curiosity can be defined as the expectation of surprise? Idk how mathematically based you are, but Schmidhuber actually defines curiosity and rewards in terms of the first derivative of compression, basically how steep or how fast you can compress experience.
All this to say with agency: when we consider the incredible freedom we have it's amazing we are not overwhelmed by choices. Surprise and curiosity are a way for us to rank our choices.
We have a hypothesis in FEP, really this is what we expect the data to output from a given input. The difference between the expected output from actual output corrects our hypothesis. This is surprise.
We have a sense for what experiments are likely to give us major refinement. That's how scientists pick which experiments to run. Some change the course of the field of study!
Honestly I think is similar to Alfred North Whitehead's theory of perspective of perspectives from your ''Process Philosophy: From Plato to Whitehead and Beyond'' with the difference that it is materialistic. It presupposes the existence of the brain the matter of the brain and as such it is presupposes mathematics. Materialism goes hand in hand with the apriority of mathematics. Whitehead's idealism presupposes this subjective consciousness in fact an emotional consciousness that interacts with other such "thing" and in this interaction the world ensues.
As an idealist I'm not impressed. I would be impressed if the kernel of objectivity would be somehow derived.
By the way "Fichte'' is pronounced with an H like in Bach. CH is always H in german language unless when followed by an s. Then it is pronounced x. ch=h; chs=x
Fichte means Spruce and ficken ... well check google translate. It is hard to believe an german speaker wouldn't know or even an yiddish speaker.
Regardiing constantly testing hypotheses, the superego is constantly testing impulses from the unconscious to verify those impulses match one's self-identity. Cognitive dissonance.
Can you please expand on how the FEP is also 'likely' about the universe itself? As per my understanding, you need both internal and external environments (separated by a Markov Blanket). By stating that the FEP could also be applied to the universe as a whole, it seems you are presupposing something external to the universe. I perhaps could see this working if internal and external are more loosely defined within a metaphysics similar to that of Bernardo Kastrup. Is this what you were going for, or were you thinking of something else when you said the FEP could also be applied to the universe?
I read something similar about vision, whereby there is an average 300ms delay between the time photons hit our retina and the time our brain/body needs to process them to perform a physical action. To account for this difference, our brains build a model of the physical world ~300ms into the future, so we can react to dire circumstances in time. The example given was that without this "Virtual Reality Simulation" going on, for instance, a professional baseball player would never be able to hit a ball, as the ball would be way past them before they had a chance to react. The philosophical impact on me was pretty big, as at that moment I realized we are all living in a virtual reality simulation of the future, created by our brains and possibly based on the "Free Energy" principals you are referencing here.
Aww man. I will try to read this article. It will be tough for me. I know Curt likes feedback and comments of all kids, so I will endeavour.
Take this:
_"Your brain is continually and unceasingly building a model of the world."_ — no it is not.
That is an assumption of Friston et al. There is no model to point to in the neurology. What we have is qualia in our _minds_, not our brains, and that fools you into thinking it is in your brain, so you get fooled into saying the brain is doing the modelling.
The prejudice of Friston is obvious, he is equating mind with 'brain processes' somehow. Materialism. Total prejudice and an unwarranted assumption. He can use the FEP for decent neurology and psychiatry without presuming materialism, without ever mentioning the mind and subjective, but to do so he needs to accept that when he writes "the brain is building a model" that this is itself a mere model of mental and physiological events. It is a figure of speech. The brain does not build anything. It doesn't "think", it just "biologizes". 🤣 (I like coining new words sometimes.)
The point about language and ontological presupposition is fair enough. But it's worth noting that the FEP framework itself has proven useful even without materialist/physicalist presuppositions. It's proven useful in a variety of research focused on meditative traditions by researchers who, from my impressions, are not presupposing a materialist ontology.
Thank you for the article. I haven’t heard of this until now. I have been entertaining myself by searching along similar lines of thought. There are indeed analogies with the FEP and thermodynamics. What you are describing seems to be related to mutual information, which in turn is related to the free energy of measurement, see doi:10.1038/nphys3230. An information theoretic term would probably be better than Free Energy Principle.
I wonder if we can identify a law of nature that is general enough to encompass biological and information systems whereby systems are compelled to reach low entropy, useful states. Knowledge and usefulness are inherent in many discussions about information theory and related theories. The philosophy of this is something I’m very interested in.
Similarity? “We are all hallucinating all the time, including right now. It’s just when we agree about our hallucinations, we call that reality,” Anil Seth
Is there a 'highschool math-level' explanation of the FEP? I'm guessing something important is lost without the math, but Friston's papers assume too much prior knowledge (math chops).
In your April 2021 conversation with Mr Friston you talk about The Free Energy Principle and how it relates to the idea of the "Self fulfilling prophecy". I listened to this bit so many times to attempt to get my head around the ramifications that this might have for our subjective perception of the world and the deeply entrenched beliefs we have about ourselves. The idea of "action as Inference" seems like a very important idea for Psychology. I was comfortable with idea that the brain can alter its model, its perception, to see what it wants to see but the idea that the brain can also act upon the world through physical action in ways that make its predictions more likely to be true is an incredible idea.
I once thought that the FEP would be a foundation of sorts for AGI but it doesn't seem like it's held in high regard in ML circles. Maybe it will be more useful for "naturally" agentic models, if current methods don't pan out.
This is a nice explanation. Good, multi-level prediction of likely sensory inputs followed by searches for unexpected deltas is an extremely powerful data reduction technique.
You're a good writer, Curt. Good to see you branched out to substack.
Could the FEP formalism just as easily be interpreted as saying that an organism is an environment’s model of itself? This seems plausible particularly if we are going fully embodied and enactive with our FEP by rejecting the Cartesian idea of internal mental representations.
To flesh this out a bit: Organisms embody generative models that correlate with the statistical regularities of their environment, with evolution and development sculpting them so that their very physiology down to the chemical level embodies the constraints and patterns present in their earthly and heavenly surroundings. Rather than serving as a separate internal model, an organism’s structure and behavior are deeply entangled with environmental dynamics, effectively making them a natural instantiation of the environment’s own structure.
Have you read Laukonnen & Chandaria’s paper “A beautiful loop”??? Relationality and recursivity being key aspects of “self-ing” within a FEP framework.
I’m so glad we’ve got your mind to help with this Matt. 🙏🏽❤️
I think the links between Laukonnen's work on meditation (including some his earlier work, e.g. "From Many to (N)One") and the FEP framework provides a very intuitively fitting way of conceptualizing the way that mystical traditions have long encouraged folks to experience existence. The paper with Chandaria is a fun extrapolation of those ideas.
I love Ruben. We share many of the same obsessions. And I think his Eureka heuristic is a really important contribution to cog sci. I don’t quite share his confidence in the FEP perspective but he’s pretty convincing (and he actually gets the maths, which I do not 🤪).
Nice! I tend to be interested in the links between various psychological and philosophical models and various mystical traditions. To that end, Laukkonen’s Eureka heuristic seems like it may be in a similar spirit to how Andrew Newberg relates moments of “little e” enlightenment to the “Big E” in some of his research/writing.
I haven't thanks for the tip!
How does agency play a role in this theory? Does this imply that we take actions which will minimize our internal surprise? Are there other motivating mechanisms in the many layered thing that is our brain? I assume so, but I'm not sure how to think of the interplay between FEP and, say, value systems. How many conceptual layers exist in between such ideas..
EDIT: one of the conclusions of this view of agency is that I would do best to sit in a sensory deprivation tank until I die. I can predict that future quite well! But we don't, there are obviously other things at play as well. That's not to say this theory is wrong, but it does suggest that this is just one aspect of the overall picture.
You're exactly right. I think that's the case, we take actions expected to minimize future surprise. We want to test our hypotheses in controlled ways. You might like Jürgen Schmidhuber's ideas as well. He talks about the drive of all intelligent agents, compressing experience. Value systems come from our need for survival.
RE: Jürgen Schmidhuber's ideas. Will do! These are new to me. I'll look into it :)
EDIT: I don't see much about agency linked from his Wikipedia page. Can you explain further which ideas you think are interesting? I found on his home page the sentence "Schmidhuber's curious learning agents like to go where they _expect_ to learn something" which is quite interesting, but this is also a different claim than the FEP.
Okay I'm just going to plug some of my own content 😉
Here's a short video on the compression drive: https://youtu.be/W0DA_D31MwA
Some writing from my page: https://www.philosopherscholar.com/meaningOfLife_4/
From the above link you can find his actual paper. In the video I go through the interesting parts.
I interpret FEP through machine learning which I'm more familiar with. I could explain this a lot better by drawing diagrams. The key is also the difference in probability distributions as Curt outlined, but hey equations are tough for almost everyone.
FEP talks about surprise. Schmidhuber talks about curiosity. Perhaps curiosity can be defined as the expectation of surprise? Idk how mathematically based you are, but Schmidhuber actually defines curiosity and rewards in terms of the first derivative of compression, basically how steep or how fast you can compress experience.
All this to say with agency: when we consider the incredible freedom we have it's amazing we are not overwhelmed by choices. Surprise and curiosity are a way for us to rank our choices.
We have a hypothesis in FEP, really this is what we expect the data to output from a given input. The difference between the expected output from actual output corrects our hypothesis. This is surprise.
We have a sense for what experiments are likely to give us major refinement. That's how scientists pick which experiments to run. Some change the course of the field of study!
It doesn't explain almost nothing. Just hype.
Honestly I think is similar to Alfred North Whitehead's theory of perspective of perspectives from your ''Process Philosophy: From Plato to Whitehead and Beyond'' with the difference that it is materialistic. It presupposes the existence of the brain the matter of the brain and as such it is presupposes mathematics. Materialism goes hand in hand with the apriority of mathematics. Whitehead's idealism presupposes this subjective consciousness in fact an emotional consciousness that interacts with other such "thing" and in this interaction the world ensues.
As an idealist I'm not impressed. I would be impressed if the kernel of objectivity would be somehow derived.
By the way "Fichte'' is pronounced with an H like in Bach. CH is always H in german language unless when followed by an s. Then it is pronounced x. ch=h; chs=x
Fichte means Spruce and ficken ... well check google translate. It is hard to believe an german speaker wouldn't know or even an yiddish speaker.
Regardiing constantly testing hypotheses, the superego is constantly testing impulses from the unconscious to verify those impulses match one's self-identity. Cognitive dissonance.
Really helpful, thank you
Can you please expand on how the FEP is also 'likely' about the universe itself? As per my understanding, you need both internal and external environments (separated by a Markov Blanket). By stating that the FEP could also be applied to the universe as a whole, it seems you are presupposing something external to the universe. I perhaps could see this working if internal and external are more loosely defined within a metaphysics similar to that of Bernardo Kastrup. Is this what you were going for, or were you thinking of something else when you said the FEP could also be applied to the universe?
I read something similar about vision, whereby there is an average 300ms delay between the time photons hit our retina and the time our brain/body needs to process them to perform a physical action. To account for this difference, our brains build a model of the physical world ~300ms into the future, so we can react to dire circumstances in time. The example given was that without this "Virtual Reality Simulation" going on, for instance, a professional baseball player would never be able to hit a ball, as the ball would be way past them before they had a chance to react. The philosophical impact on me was pretty big, as at that moment I realized we are all living in a virtual reality simulation of the future, created by our brains and possibly based on the "Free Energy" principals you are referencing here.
Aww man. I will try to read this article. It will be tough for me. I know Curt likes feedback and comments of all kids, so I will endeavour.
Take this:
_"Your brain is continually and unceasingly building a model of the world."_ — no it is not.
That is an assumption of Friston et al. There is no model to point to in the neurology. What we have is qualia in our _minds_, not our brains, and that fools you into thinking it is in your brain, so you get fooled into saying the brain is doing the modelling.
The prejudice of Friston is obvious, he is equating mind with 'brain processes' somehow. Materialism. Total prejudice and an unwarranted assumption. He can use the FEP for decent neurology and psychiatry without presuming materialism, without ever mentioning the mind and subjective, but to do so he needs to accept that when he writes "the brain is building a model" that this is itself a mere model of mental and physiological events. It is a figure of speech. The brain does not build anything. It doesn't "think", it just "biologizes". 🤣 (I like coining new words sometimes.)
The point about language and ontological presupposition is fair enough. But it's worth noting that the FEP framework itself has proven useful even without materialist/physicalist presuppositions. It's proven useful in a variety of research focused on meditative traditions by researchers who, from my impressions, are not presupposing a materialist ontology.
Thank you for the article. I haven’t heard of this until now. I have been entertaining myself by searching along similar lines of thought. There are indeed analogies with the FEP and thermodynamics. What you are describing seems to be related to mutual information, which in turn is related to the free energy of measurement, see doi:10.1038/nphys3230. An information theoretic term would probably be better than Free Energy Principle.
I wonder if we can identify a law of nature that is general enough to encompass biological and information systems whereby systems are compelled to reach low entropy, useful states. Knowledge and usefulness are inherent in many discussions about information theory and related theories. The philosophy of this is something I’m very interested in.
Similarity? “We are all hallucinating all the time, including right now. It’s just when we agree about our hallucinations, we call that reality,” Anil Seth
Is there a 'highschool math-level' explanation of the FEP? I'm guessing something important is lost without the math, but Friston's papers assume too much prior knowledge (math chops).
Nice. I especially liked “reality-whatever that is...”.
In your April 2021 conversation with Mr Friston you talk about The Free Energy Principle and how it relates to the idea of the "Self fulfilling prophecy". I listened to this bit so many times to attempt to get my head around the ramifications that this might have for our subjective perception of the world and the deeply entrenched beliefs we have about ourselves. The idea of "action as Inference" seems like a very important idea for Psychology. I was comfortable with idea that the brain can alter its model, its perception, to see what it wants to see but the idea that the brain can also act upon the world through physical action in ways that make its predictions more likely to be true is an incredible idea.
I once thought that the FEP would be a foundation of sorts for AGI but it doesn't seem like it's held in high regard in ML circles. Maybe it will be more useful for "naturally" agentic models, if current methods don't pan out.