Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

okay it's September 14th 24 or 16th my my clock was looking weird um we're entering chapter 9 for cohort 6 um do either of you want to type or ask any topics you want to cover explore for nine otherwise we can look through the chapter I was actually about to oh I'm sorry go good I was actually gonna ask if we could just like run through the chapter because this was the first chapter that I was like oh man we are jumping into creating this model uh could we just run through uh one time into going through like what was the data and what was the model that we were actually applying as well as just going through the loop of figure I'm sorry I've lost the figure it doesn't really matter for that last part yeah got it okay holding back I think going over the chapter is the first thing that we'll do then there's the questions and some notes and then there's also looking at software just speaking generally maybe we can do that in one or two weeks from now looking at the Jack's branch on pmdp looking at RX and fur and looking at other empirical work let's start with the chapter though okay second half of the book applying active inference coming fresh off of chapter 6 recipe chapter seven and eight giving the continuous and discreet time now getting to using the models that are created for data analysis that's kind of revealing that in in all the previous chapters 2 3 four um not five but then seven and eight when talking about building a generative model it's to produce synthetic data like to run a simulation forward here now we're also or alternatively looking at a setting of having a certain structure of data and then seeking to recover latent parameters instead of specifying the latent parameters and then generating the synthetic observation like data um our general goal is to recover the parameters of the generative model that that system of Interest uses to produce Behavior subjective model 9.2 metab basian methods this is really well summarized in figure 91 it's like zooming out a layer from the figure 4.3 um figure 7.3 type PDP hash model stream 14.1 um PDP that we've been specifying and talking about minimizing the variational free energy and so on zooming out saying okay not just the mouse and the Tas now considering the observations for the mouse are actions for the researcher and then conversely the output behaviors of the mouse are objective observations for the ethologist interesting how it's using the notation of like the subject centered view because it's it's calling them oh where whereas you might also call them actions of the the researcher and then it's like it gives the the hint of the suggestion okay yeah you could make a policy if you wanted to do policy control variable um like another Pi sub scientist and then have that connect to O but otherwise we're just taking this as a given so we're just taking that as the input data for here and so that subjective model which does the forward inference which is like um the kind of perception action Loop that it it's focused on that's happening with a given set of observations in an experimental context and outputting some Behavior so this kind of showing okay the data like the input data here's the the information that we knew on the which where we put the food in the te-as and here's the actual sequence of moves that the mouse made um brief little statistical interlude that starts to to bridge back from this question of reducing uncertainty of the unknown parameters of Interest figuring out the subject's prior beliefs it's like we know where their iades were that's just empirical um or we know which things they bought that's just empirical but then to do belief modeling cognitive modeling Beyond just the pure behaviorist layer then that is reducing uncertainty about these causal PR M of Interest so how are those going to be analyzed um from real finite data um they go into variational Applause this is something brought up in other textbooks um in the SPM textbook okay…