Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right hello everyone coort 5 it's our first discussion on chapter 4 so does anyone want to give just any thought on four overall just any any aspect of what they expected or got from it or however far they got or some section that was just interesting to them I have a question um so one kind of generative model that I am familiar with is um one way of representing a gener model is using markup decision process or partially observable Mar markup decision process uh this is for discreet um the discreete case of generative modeling I'm wondering what else what other what are other very popular or common uh structures that are used to uh write down or Express generative models anyone want to give a thought I would say one alternative way to write down parts of generative models is the continuous time but are you asking is there another way to represent discrete time generative models other than with a PDP right to begin with because of course there's the the whole other uh area of continuous time but um there's one important aspects to the whole um question of um of basian inference which is or some some agent some organism that's implementing basian inference is how does it uh not only learn the parameters of a generative model but how does it also learn the structure of that model um and this is connected for example to questions about nature versus nurture and and uh development in in humans or animals like what are the um the initial conditions what kind of um either inductive priers or like um structure biases doeses our neural architecture uh provide us with from the very beginning and whatever it doesn't provide us with how how do we um figure out what kinds of structure is appropriate for for modeling a particular domain all you have any thoughts uh well first of all uh about the initial uh states of any system uh that I mean it it doesn't always um in active inference modeling we don't we don't always assume U any agent like tabul laasa like I mean we don't always begin from scratch so uh in most cases we necessarily impose some initial conditions uh in order to make it more efficient to model the exact behavior of or phenomenon of interest uh because otherwise we won't be able to um to model the whole process I mean all the emergent properties that um the agent have acquired the the agents have acquired through their lifespan uh so it is uh it wouldn't be feasible to try to um put everything in in such an encapsulated form of um generative model in order to Encompass uh the agents uh entire u i mean experience and lifespan so it is necessarily uh more focused than that uh and uh IT addresses a specific problem uh we're trying to uh we're trying to model so um I guess uh it's not uh a prorally it's it's not something a priori uh and it isn't U any prerequisite to um model uh any situation but uh when it comes to how an agent acquires those uh initial learnings those initial States well obviously there are many many ways to uh address that uh problem U with with various perspectives but um of course F or active inference uh in my opinion is not um a theory or a framework that can answer every um every question regarding uh to the behavior of the agent uh in every possible situation so we need to look for uh other or other granular problems um in order to uh account for the lifespan experience of an agent got it uh thank you so um just to kind of reiterate and clarify this your response for myself um in my mind I kind of related a bit to on the one hand this distinction we we often need to make and keep in mind between an uh a realist interpretation of active inference and an instrumentalist one uh where the latter is us taking this this mindset of like okay uh there there are agents out there in the world there are things out there in the world and and they do stuff uh we can't be completely sure of exactly what they're implementing uh mechanistically but but we can develop a framework for uh making some SE sense of…