Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hey everyone, it is October 21st, 2022. We're in Cohort 1 of the textbook group. It's Meeting 21, and we're having our first discussion on Chapter 9. Chapter 9 is Model-Based Data Analysis. Let's look at what the section headers are. There's a short introduction, a discussion of meta-Bayesian methods, which is going to be very interesting, and in some ways is even like an entry point to thinking about entity modeling in active inference. And so that'll be kind of fun. We'll return to Variational Laplace, resonating with our just-completed discussion on Chapter 4 and Laplace. Then, Section 9.4 and 9.5 are going to help us see where data, in terms of like gigabytes of actual, good day, Jakob, actual data from measurements and so on, where these come into play with models. So how does one go from a furnished, trainable model to a parameterized, specified model, which actually is explaining variance in real-world datasets. And then there's some examples of GMs and some models of false inference. Well, I added only one general question from a lighter reading. I think as we all here move through this, we can generate a lot of other key points and questions. Ultimately, the models described in this book are only useful if they can answer scientific questions. Let's just, we could rehearse the introduction again, but let's just go into it. So, Meta-Bayesian Methods. This chapter deals with the utility of active inference formulations in analyzing data from behavioral experiments. So, one could imagine all kinds of bodily, verbal, digital behavior. This goes beyond proof of principle simulations we've seen in previous chapters and instead, exploits active inference in answering scientific questions. Broadly speaking, there's two related reasons for fitting a computational model to observe behavior. The first is to estimate parameters of interest. The second is to compare alternative hypotheses. So, to parameterize from data is one opportunity. That is to make some model of brain function and then understand within one person or one group or two groups how you can consider those parameterizations to be phenotypes. Phenotype is something of a biological organism or system that's measurable. So, like, femur length is a phenotype. But also, a phenotype doesn't have to just be something that's measurable on the body with a ruler. Like, phenotype might be distance ran in the first 50 seconds after the hawk flies through the sky on a cloudy day. And so then, that is still a measurement that could be inferred or discussed or made. And we're talking about computational phenotyping because the data and observations for sure are like a basal phenotype. Like, the button being clicked was measurable. Pheno means to show. But also, we could talk about, like, the precision variables in our cognitive model as being phenotypic. And also, model comparison can be used. So, whereas this first modality of parameterization is like, the dataset is fixed. We collected 240 fMRI datasets. Now, we're going to parameterize. So, we're going to give some plasticity to our model. And we're going to fit parameters so that our model is, like, the best resembling it can to these 240 fMRI datasets we have. In the second modality, we're treating, in some ways, the models as fixed. And then, evaluating to what extent different models stack up. And that can be used at a very fine scale to look for different models that fit better to a given dataset. But also, this is where we can talk about, like, actual biological explanations and predictions. So, like, to give an example from SP... Oh, yes, please, Ali, first. Sorry, a related question to this whole discussion is... I think it was in section 9.2. It says, this goes beyond the proof-of-principle simulations we have seen in previous chapters, and instead exploits active inference in answering scientific questions. I didn't quite understand this statement here, because in the previous chapters, we were also engaged with…