Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right welcome back cohort 5 we are in our second discussion on chapter nine so Andrew or anyone else who wants to write a question or raise their hand let's just jump into it okay I just heard the recording in progress thing maybe there's something about Pete's Coffee that causes a time dilation but either way now we're in chapter nine so where shall We Begin yeah I was going to say if anyone has any particular points they want to get into with this chapter happy to jump into those otherwise um if anyone's like you know only able to make this meeting ween able to make the last one I'm happy to kind of give a general overview of what's going on here uh if you don't mind a general overview would be good to start with if other s c yeah go for it previous meeting is at 11:00 in the at night in my time so I I haven't been uh able to make them so General overview would be really helpful for me if if it's a possibility yeah great um yeah let's do that and there is a lot going on in this you know this chapter this is sort of the aside from chapter 6 which gives us a general recipe for composing uh gener generative models and kind of the more General method you might take here in actually applying active inference in the second half of the book um recall in chapter 7even we're given the um PDP model how to construct that chapter eight we're given more continuous models and then here we bring it back to okay so how do we actually use these models and construct them in a way that we can apply them to empirical experiments where we're actually collecting real data um and so we're we're moving beyond the range of just purely building models and running simulations to collecting empirical data such as like neurophysiological data could be EEG it could be just observing um behavior of anywhere from human beings to to rats or if you're in more of a an engineering setting you might be looking at other kinds of data that's being extracted and you just basically we want to figure out how is that what are the what's the process uh that is generating that data which of course we may never know um but we can attempt to model it so um yeah this chapter it's it's quite dense I'll be honest whenever we incorporate everything um but it does give us a nice full flow of how to apply these methods to empirical data um so first of all in the introduction there's hints at what we'll find in the rest of the chapter um we're introduced to metab basian analysis which is where we actually view this as the agents subjective model versus our own objective model that we're building of the agent subjective model that kind of uh carries into the active inference ontology right just we do have to account for the fact that we are sort of the quote unquote objective scientists in this situation we're attempting to model the behavior um and other parameters of like an a separate being from ourselves like in a teamas we're trying to you know understand the subjective model of a rat which we don't have you know collecting data and applying analysis to see if we can do something like approximate it so here our goal is to what we would say is recover or infer the unknown parameters of they phrase it as the subject's brain its subjective model using our own objective just by drawing on the behavior in our observed data and then we uh invert our objective model to infer those parameters of the subjective model and then finally by doing so we can test and compare hypotheses for example we might find multiple models that we could build we could compare them and um and see which one holds best to like the behavioral outcomes that we're seeing in the data and then they they have another concept called computational phenotyping um this is where we kind of by by recovering model parameters prior beliefs of our agent we could actually like kind of classify um our our the individuals we're studying whether it be multiple rats and a teamas it could be clinical patients and more…