Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hey, everyone. Welcome back. It is September 26th, 2023, and we're in cohort four, meeting 14 on our second discussion of chapter six. So before we jump over to the text, does anyone want to just give any thought or take on six? Yeah, briefly, just Charles, it alternates time. So the earlier time, now we're in cohort four, chapter six, and vice versa, but it's all different angles on the same thing. So it's not like there's something that won't make sense in six because... That's cool. I'll just hang around and carry on. Thanks. Yeah. So you want to give a thought on six, or we can look at the questions that haven't been addressed yet. Okay. Let's just go to a question. And then if somebody has another question, they can write in the chat or raise their hand. Okay. The authors talk about language. But in language processing, one usually assumes recursive processes, e.g. a context-free grammar. How can FEP active inference account for recursivity? In my understanding, hidden Markov models implement types of regular grammars. I assume that hierarchical extensions, i.e. section 642 does not fundamentally change this, does it? Did anyone here ask the question or does anyone want to give a thought on this? I'm not exactly sure what the question is getting at. However, some senses of recursion can be accommodated by hierarchical modeling. So active inference models of reading have, for example, letters nested within words, words nested within sentences, and so on. So this hierarchical model, embodies the sparsity associated with this kind of a recursive process. And then another kind of related notion of recursion is closer to re-entry, where something is more like self-reflexive. And so that is less referring to the hierarchically nested, like kind of nested for loops notion, and more having to do with re-entry of input and reference, which we might point to, first off, just the fundamental re-entry of the consequences of action through perception via the niche for the agent. And then a little bit more on the sophisticated cognitive entity, the recursive identity concept and like identity re-entering in reference to itself. Ali, go for it. Yeah. Ali Kowalikov, Well, perhaps the question also, I'm not sure, but maybe one angle to answer this question is to answer whether generative models are exclusively recursive models or not. So I believe in active inference framework, and especially its current formulation, there isn't any restrictions for generative models to be exclusively recursive. So they can be applied for both, I mean, nested and recursive models. And also in some situations, we can have purely, I mean, non-recursive or sequential models as well. So yeah, it's in its current standing, it is much more broader than only a recursive kind of structure as HHMs that we're dealing with a couple of decades ago. Great point. And it kind of returns to the theme of chapter six and the questions that are asked in chapter six. It's like, is, or what kind of recursive phenomena is being desired to be modeled? And then there's sort of like what you intentionally include in your model. Like, you know that you want to have this phenomena and you're going to engineer and implement motifs or mechanics to get this phenomena. Like it's your target phenomena. You want to have a phase transition associated with people like leaving a crowded room. And then like, you want to like, kind of like build that into the model super explicitly. And then a little bit more indirectly would be you want to have some phenomena arising, but you don't want that phenomena to be like explicitly a parameter in the model. Um, but that's the chapter six question. What do you actually want to model? What system of interest do you want to model? And, and that's non-trivial to identify the boundaries of the system and to subset what you actually want to model about it. And what's interesting, what questions you want to model? Because otherwise you're just making a map…