Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hi, thanks everyone. It's 4th of July, 2023, and we're in our second discussion on Chapter 7. So we'll head back over there. Does anyone want to type a question or add any reflection or thought or just question or area they want to look at today? Okay. Well, I could say something. Hi there. Yeah, please. So last time I joined was four weeks ago, and I was saying about while I was advocating having Python code as examples, I went away and just coded up. The bits at the very beginning of Chapter 7 of the notes, the musical notes, which were sort of transferring, converting the text into Python with these comments. It didn't really get me. Well, it showed the reasons why the matrices contain the numbers that they did. It didn't really say much more than that. But I've just discovered Appendix C in the book. And in a way that goes through something similar. Although it doesn't explain the matrices, the precise numbers in each matrix in quite so explicit a way. And then later on in Chapter 7, there's the maze example. And again, these are all filling in with the application. But then it's just formatting the data to the point where it can then be given to an SPM12 function. Yeah. So MDP there without. So I would like to see what's inside there and then be able to graph, for example, various metrics like equations. Things like mutual information, novelty, to help get an intuitive feel. I think Appendix C says about being able to play around with the numbers, get an intuitive understanding. But I want to see more than just playing around with the matrices, the inputs. So, yeah. So that's just an update on where I am, really. Thank you. Yeah. Appendix C is very sparse. Mm-hmm. And it doesn't. I mean, they would have done well, potentially, to have just simply provided the reproducibility code for each figure without interspersing it with plain text. Mm-hmm. Because then it's kind of this reverse engineering process. But what you said is absolutely right on, which is the entire POMDP, the entire figure 4.3 discrete time model. It's all the letters that we talk about. A, B, C, D, and then E, which is just given the letter V here. And then the true starting condition. And then you just let it run. Now, here's the SPM, MDP, VB, X, variational base, X. I'm not sure exactly what the X is. But I've heard that this is sort of a, like, not inscrutable, but it's not simple to translate this function just straightforwardly to Python, which is one reason why, like, many people do continue using the MATLAB variant, because this contains a lot of, like, accelerated methods that are – and it contains, like, a lot of extra bells and whistles to generate the figures, and so that makes it quite easy. But in general, it speaks to exactly what Chapter 6 highlights, which is that active inference modeling is about representing the agent and the environment, and then the methods for letting the agent come to its own solutions are straightforward. Rather than needing to do a secondary – and at some computational scale, yes, further techniques are needed. But to a first approximation, the work of understanding a system of interest is in exactly specifying the dimensionality of the matrices. Same as in pi MDP in model stream 7. But yeah, how was it – I mean, did you see the matrices any differently or think about the models any differently? No, what I coded up, it was really just – well, to get an idea of how things might be coded up, and this wasn't to have something that – certainly not anything that was simulation efficient, it was something very clumsy. So, for example, what you're showing there with the B matrix, it would be explaining why the 97s are there in those particular places. So, starting off with an all zeros matrix, or maybe an all ones in this case. And I know it's normalized there. But so that you can – you're then describing the matrix B directly in terms of the application. You don't need the separate bits of text to explain what there is…