Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

All right. Hi, everyone. It's cohort 3, 725-23, and we're in our first discussion on chapter 9 here. So let us head over to chapter 9 and begin before we go into any questions with just any general thoughts or comments that people want to bring up about chapter 9. Model-based data analysis. Even just what, however in-depth people have read this, like what would they expect such a chapter to cover? What would somebody want this type of a chapter to cover? What does this title bring up for people? Or the quote from President Obama? Well, I'll just give thoughts, but please write in the chat or raise your hand or... Well, I'll just give thoughts, but please write in the chat or raise your hand or... Feel free to add in more. Great. Yes. Yeah. Thank you. It's totally fine across different cohorts. There's not really a difference. And just it's so many coats of paint that skipping around, but it's not out of line. And also, in a sense, chapter 9 in specific... I mean, let's look at the textbook. Chapter 1 and chapter 10 are very similar, that they're both overview chapters. There's no equations. They get big picture. Chapter 2 and 3, we know the low road. Chapter 4, technical details with generative model. Chapter 5, a lot of neurobiology and examples. Chapter 6, with the recipe for designing active inference models, probably makes more sense after knowing what the generative model is. However, you don't have to know it in technical detail. Chapter 7 and 8, on the discrete and continuous time generative models, again, probably helps to understand the generative model. So 7 and 8 might rely on 4. But 9, in a sense, stands alone. And so I hope that we can dive into some of the intuitive and immediately accessible components of Chapter 9, because it's going to be a very different tack than the approaches from the previous chapters. So let's get into it. In this chapter, we focus on the ways in which active inference can be applied in understanding empirical data. Now, in all the previous sections, the examples, whether it's the frog jumping out of the hand, or the musical notes, or the rat in the maze, or bird song, all of these examples follow kind of a similar pattern, which is like a generative model is specified, and then it's played out. It generates data. That's one reason why these are called generative models, because these are models that generate data. However, in the real world, we want to go both ways. We want to be able to specify a model a priori, and then generate data. But also, we want to have structured data, and then do inference about hidden states, for example, given the empirical data. So we don't just want to spin out mountains of data just to show that we can. Then we want to take the data that we have, which is called empirical, and then do some kind of inference with it. Our general goal is to recover the parameters of the generative model that a subject's brain or mind uses to produce behavior, the subjective model. So if all we were interested in was descriptions of behavior, we wouldn't need sophisticated modeling at all. We would just need summary statistics or descriptive statistics on behavior. Like if all we wanted to do was just, how often does this person make a sound? Again, you could just count. But if you wanted to say, how frequently does the baby transition between happy and unhappy? So in that case, you need to parameterize aspects that are not directly observable, but are based upon observables. And that's subjective, because it's about a subject. So we'll talk about subjective and objective models. And this is going to be introduced in terms of meta Bayesian methods, which sounds very meta, but once we see the figure, it'll actually be revealed to be, in fact, the structure of behavioral observation. And this relates to many different topics we've discussed, like Bayes' optimality of behavior. So looking at the diversity of neural systems or cognitive systems and asking, what is the…