Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Okay, welcome, thanks. It's March 15th, 2023. We're in meeting 20 of Cohort 2 in our first discussion on Chapter 9. So, as we begin this discussion of 9, is there any general comment or question about 9 that anyone wants to bring up? Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. Okay. One part on 9 is it at least points the direction towards an avenue that many people who are interested in learning and applying active inference want to go down, which is to make generative models, yes, that help us understand different cognitive systems, and also have models that can accept empirical data so that we can enact the tail of two densities. So, we can use the model in the generative direction to go from specified parameters to output data sets of synthetic data, kind of like generative AI, but then also to recognize outcomes in the world or data and then update or parameterize our generative models from those data. And it's that tail of two densities, the generative density and the recognition density, the ability to run these models both ways to generate valid synthetic data from a generative model. And if you design your experiment so that those data which are generated are also actually being generated in the real world, then you can just instantly plug it back in. And we'll, I'm sure, unpack and explore more. So, we'll go through the chapter, but is there any just general comments or general thoughts on chapter 9 first? Okay. So, let's look at one previous question. So, how did chapter 4 on the generative model relate to chapter 9? As we're getting close to the end of the book. Chapter 6 was a recipe. It was about applying active inference. Chapter 4, we were learning the necessary prerequisites. And chapter 1 and, or 2 and 3 were the roads to chapter 4. Then, chapter 5, we got the neurobiology and some examples. Chapter 6, we got the recipe. Chapter 7 and 8 was the discrete time and continuous time generative models. And now we're in chapter 9, where we're learning how to solve problems. And it's about getting one's hands dirty with case studies and problem solving. So, we'll see this in the figure upcoming. A key distinction is between the subjective model and the objective model. There's different readings of the first sentence and in some readings it could be seen as playing fast and loose with instrumentalism and realism. So, the subjective model though is that which the subject is, that's that which is subject specific from the subject's perspective. Now, it turns out that it is a map that we're making from the subject's perspective. So, it isn't the actual territory of the subjectivity, but it is from the subject's perspective. So, it's the objective model, which doesn't mean that it's the one and only. It means that it's our view from the outside seeing that thing as an object. So, this entire setup with the tail of two densities and the meta-Beyesian enables us to make Bayesian cognitive models and think about experimentation in a Bayesian context. So, this is a way to go to Bayesian context, which lets us talk about Bayes' optimal cognition in different systems. And one area that that has been especially relevant is in computational phenotyping. This chapter deals with the utility of active inference formulations in analyzing data from behavioral experiments. It's going beyond the proof of principle simulations. So, in the proof of principle domain, we have the jumping frog in the hands. We have the rat in the teammates and so on. And here, active inference is going to be used to answer scientific questions. Since the subjective GM is the generator of behavior in active inference, then the scientific hypotheses that we're going to entertain about the causes and consequences of phenotype must be in terms of hypotheses about alternative generative models. So, the challenge is to, just like an abductive logic challenge, generate a portfolio or spaces of…