Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

hello and welcome everyone it's March 1st 2023 we're in cohort three and we are having our first discussion today on chapter three chapter three um is the high road to active inference previously in chapter two we talked about the low road and today we're talking about the high road those two roads are laid out in figure 1.2 we took it from Bayes theorem to active inference in chapter two and chapter three is going to be starting from the free energy principle and also taking us to octave inference so before we go through the chapter or look at any questions does anyone want to just give any reflection or thought what was some experience they had in reading chapter 3 or a quote or an idea or something that stood out for them you can raise your hand or just go for it okay so feel free to raise your hand or write in the chat if you ever want to add anything um chapter three takes a different attack than chapter two [Music] anyone want anything or we'll we'll start exploring through the chapter and then coming to the questions that people have raised um today or in the next week okay we'll move through chapter three hopefully relatively quickly to get to the first of the questions and then we'll have seeded some questions for coming to our discussion next week so in section 3.1 in chapter 2 free energy was motivated as a bound in a way that we can do approximate Bayesian inference coming from the low road to active inference and you can do free energy based or energy-based methods like a variational autoencoder you could do that on any Bayesian statistical problem at all and it's used commonly it doesn't necessarily mean that it's connected to action selection at all in chapter 3 they're going to start from The High Road which is this Central imperative that organisms or really just things must maintain their existence which is going to be operationalized in a way that's compatible with the surprise minimization that was brought up in chapter two and just like in Chapter 2 where free energy was proposed as a way to bound surprise which is like what we would really want to know coming from the high road we also want to bound surprise and it turns out that minimization of free energy is again going to appear as a computational computationally tractable solution to the problem the chapter is going to describe the formal equivalence between the minimization of variational free energy and maximization of model evidence or self-evidencing in Bayesian inference and then active inference is going to be brought up because we're going to head into chapter 4 which is going to be all about the generative models at the heart of active inference active inference is about how organisms or adaptive systems maintain their existence by minimizing surprise via more proximally this tractable proxy variational free energy and the two ways to minimize variational free energy are perception and action change your mind change the world integrated imperative that supports this unifying perspective on systems high road is going to start from the premise that any living organism or any persistently observed thing has to maintain itself in a set of preferred States if it doesn't maintain itself in a set of expected slash preferred States it just isn't that kind of thing that is going to apply to everything from an oil droplet that's diffusing or not to more sophisticated cognitive goals humans physiological entities have to stay within their physiological ranges in reward learning it might be proposed that there's this value function where acceptable ranges are more rewarding and then rewarding Paths of action would be selected in contrast in active inference we establish the expectations around those physiological variables in terms of a preference and then we reduce our surprise about realizing those preferences slash expectations that allows us to take the most likely course of action to maintain being the kind of thing we are rather than proposing this…