Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hello, welcome, thanks everyone. It's February 1st, 2023. We're in Meeting 14, Cohort 2 of the Paradol textbook. We're on Chapter 6. So, today we can look over any and all of the questions on Chapter 6 that have been raised. We can also go to the text. But just to begin, does anyone want to bring up any quote or topic or anything about Chapter 6 or any section from the text or question that someone wants to go to first or a general point about Chapter 6? Ali, and then anyone else? Well, I think now that the white paper from Versus Lab is published, it might be a good idea to read Designing Ecosystems white paper along with Chapter 6 because it can provide a much wider perspective and see how the research in developing these kinds of active inference based AI is currently developing and to see how the plan for the future developments have already been discussed, especially in the case for sympathetic and shared intelligence, which is kind of long-term plan to develop those kinds of intelligence systems. So, yeah, I think this paper can immensely help to see the big picture related for designing these kinds of active inference based models. Awesome. Thanks. Yeah, the first 20 pages of the paper have great points, but just to jump to the timeline part. It's 2023. This was released at the end of 2022. And in the coming years, here are some doubling timelines that we might expect or prefer. So that does help set the stage for Chapter 6. Thanks. Anyone else? Just unmute or just raise your hand. Chapter 6 is a recipe for building active inference models. So it's not the fast food restaurant yet, but it's a recipe for the home and potentially industrial chef. It covers the essential steps to design an effective model. And that is going to be taken in like a staged way. And starts on page 105 in the textbook. So shall we look at the questions? Does anyone have a specific question they would like to jump to or any general points? does anyone have a specific question? Otherwise, let's look at the recipe and then see if there's things that we can add. And then in this active inference model recipe, let's see if we can flesh it out with what we're seeing and learning from doing modeling. And also how it's similar and different from other modeling recipes. How different is this than doing a linear regression model? How different is it than doing some other type of modeling or analysis? So give me six chapters to model active inference and I'll spend the first four on the generative model. So, Carl Lincoln. Which system are we modeling? What is the appropriate form for the generative model? How to set up a generative model? How to set up the generative process? So those have been copied out here. Does anyone just want to give some comments on these stages or how would one just briefly summarize what each of these four stages are and what they do? Or just any feature about these four steps? I guess one thing is they need to be discrete these steps and that's already a decision that it's we made. Do you mean the steps are separate from each other? Yeah. And we need to have categories that's if we have these hidden states in the Markov model, these are separate states. And I guess that's a kind of a problem and a restriction also. Are there other modeling approaches that you think don't have such restriction? Yeah. For instance, in neural networks, these transformer models where you can type in a whole picture, for instance, or something like this? I may be wrong. So maybe there's categories too. Yes. Well, maybe a related question is when we specify what is a hidden state and what is an observation in our generative model? Are we kind of locked into that? Or is it possible for something like a pre-POMDP that describes structure? So potentially this is between steps two and three. The hidden state is going to be an image and then it gets set up with a certain dimensionality. So, anyone can raise their hand. Let's just look through the…