Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hello, everyone. Thanks for joining. It's October 28th, 2022. ActInv Textbook Group Cohort 2. We're on the 9th. Oh, we're in the... The weeks are shifted back by one. Is that true? There's two 21s. Or there's been some kind of other time glitch. But today is 1028. Anyways. We're discussing Chapter 4. Continuing our discussion of Chapter 4. So, where would anyone like to start with? Does anyone have a general Chapter 4 remark? Or a particular question that they want to look at? Or anything else? Daniel, it's Bronwyn. Mm-hmm. There's just a lot in Chapter 4, obviously. And it sort of... I think it takes in a lot of different aspects of active inference and then sort of piles them all in there and then throws out a whole lot of maths. So, you know, it's unbelievably dense. But... And, you know, I don't know any maths. But interestingly, as I read it, there's some sort of themes that come out that are sort of simplistic. And I'm wondering if somehow that can be sort of... If you can sort of summarise that today, if that makes sense. It's just sort of like a... A capsulated paragraph of the process it goes through. That was one thing. The other thing was... The Box 4.1, the Message Passing and Inference. I didn't quite get the difference between Variational Message Passing and the Belief Propagation. Whether it can talk a bit about that. It's about... Yeah. Yeah, that's about a bit early. Okay. Does anyone have any... Thought on this in chat? Or just anything else to add? Or we can... Try to address these. Okay. Okay. Well, one way to look at the theme of Chapter 4 or the kind of looking beyond all of the technical details. And the technical details, it's kind of like an iceberg because they could show a little or they could show a ton. Every time they show one, they could also show a whole set of ways to get to it or not. So it's kind of like... Maybe... It's like a variable of the text. How much of the math they just mention versus explicitly write out. Yeah. Because there's a lot of papers with no equations at all, like a philosophy paper, that will still say, because of the variational free energy or because the generative model is this way. So kind of referencing the same topic. And so... But different texts differ in how much formalism they have. What is the role of the generative model in active inference? What does anyone think about that? Or in the textbook, where is any quote where they use... Where they say, a generative model is blank. Or this is the role of the generative model. We can look for some too. Or any other thought someone has. I mean, I think it says that they can vary depending on the inference problem. That's one thing about generative models. Okay. Yeah. Another way to... That variation point's really important. The variation is sometimes described with like fine-tuning within a given inference problem, which is sometimes framed as perception in the most fast timescale and learning and memory in a medium timescale. Yeah. And then also different implementations or different like model structures can vary across inference problems. There's also different ways to like write and see the generative model. Can anyone just like list or suggest how are generative models shown or represented? So is that the factor graphs? Yeah. The 4.3. In which section? Figure 4.3. Is it... 4.2. So how... Yeah. One view of the generative model is like this figure 4.3 kind of classic slash common representation. The Bayes graph view with the nodes being a random variable and edges being relationships among variables. And so like if every variable were connected to every variable like every time point were causal on every time point every hidden factor influence every hidden factor you just have like a fully connected all by all causal association matrix. so that's like the weakest world model. And then you can say well the present is like the only intermediating Markov blanket between the future and the past. Or this is the you…