Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right hello it's February 15 2023 we're in cohort 3 of The Textbook group and we're in our first discussion on chapter two of the textbook so there's a lot to discuss chapter two goes hard before we go into the questions that people have raised does anyone want to just share a first recollection something that stuck with them or a friction or just a sentiment or a feeling that they had when they were reading even part of chapter two you can just unmute and go for if you want to share it's awesome that so many people added questions there's going to be a lot to discuss anyone just want to give a first thought what did they what did they hope chapter two would do was there any section or like quote or what stood out when people were on the low road to active inference okay let us skim in a few minutes the whole chapter and then we're gonna turn to the questions and just go based upon which questions people have voted as interesting but let's just take a quick uh flyby of the low road so chapter two is going to be coming on the low road so just to get a visual reminder of what we're talking about with the high road and the low road from figure 1.2 the low road is like the how showing how from Bayes theorem which is so simple and tautological and contentless we get a how that meets with a y which is this imperative for persistence and self-organization and they come together in the generative models used in active inference which we're going to get to in chapter four so the low road is going to talk about a lot of the how and some of the what I guess as well chapter three is going to come from The High Road perspective and chapter four is where we're going to come to the heart of active inference with the generative models thank you Maria we'll get to 2.3 and and then feel free to address that section 2.2 is perception as inference and this section gives a little bit of an overview on how even before the Bayesian brain was described there was concepts of perception as inference with helmholtz and even further back as action Maria shared in live stream 43.0 and what active inference is doing is extending that inferential framework Beyond perception to also include action that is in the title it's active inference it's about inference on perception and action this is going to be operationalized with modern statistics throwba is called a generative model we're going to come back through this so just moving rapidly probabilistic reasoning is described by this equation which is um it'll be awesome I think as a question asked like to have this in plain language what is being described or what is being done with basic question an example is introduced where a person is holding either a frog or an apple in their hand and then the object is going to either jump or not jump the person has some prior beliefs about a priori prior beliefs How likely frogs and apples are in the world and they have some beliefs about what is likely to do what action an observation comes in and then one is able to have a posterior belief or a posterior a posteriori of How likely they think the object is to be one thing or the other they've updated their prior beliefs through observation so like the Bayesian moment is this critical moment where a new observation comes in and that in the context of a likelihood model updates the prior into a posterior here's an exact Bayesian inference on that um scenario and for simple settings exact Bayesian inference is totally fine it turns out that for larger and more challenging statistical areas one has to use approximate Bayesian inference but what approximate Bayesian inference approximates is exact Bayesian inference so this is kind of like the core of what you could do if you had infinite computational power and you can do it for simple examples there's a discussion of probability distributions because there's multiple probability distributions as we're dealing with probabilistic variables people may be familiar…