Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

All right, greetings, everyone. It's June 20th, 2023. We're in our second discussion on Chapter 6. So we'll first just have any general comments or anything. Then we'll turn to the questions table and look at what questions we didn't get to last time and just kind of revisit, maybe condense some questions or just see what we can do with what is here. So anyone just want to make any general comments on Chapter 6? I believe this is the second session on Chapter 6, right? Correct. Or maybe I must. Okay. No, no. Yeah, it is. Yeah. There was only one onboarding needed for the welcome back here. So we started one week faster into the rhythm. Okay. So it looks like these were the questions. I remember we left off with this one. Let's just develop this into what was into a question and then we'll continue on. Okay. Is there a common or good representation or rubric for evaluating generative model, generative process? Let's just change it to generative models and generative processes. Anyone want to give a first thought or just some other related question? Okay. Yeah. Well, go on, please. Sorry. No, that's okay. That's okay. Yeah. I don't know. I'm not sure about what's the exact criteria for evaluating generative models or generative processes. But one thing is for sure that, well, I mean, one of the main, at least when evaluating generative model, one of our main criteria should be how closely it tracks our situation of interest and specifically how relevant. And specifically how relevant it is for addressing the question we're trying to, I mean, we're trying to examine. So it's not, I don't think it's something clear cut or, I don't know, written in stone. And based on the context and the situation, the evaluation criteria rubric should be different, I suppose, both for generative model and generative process. Yeah. So again, we'll revisit this generative model generative process question later following the recent live stream. But generative process, working with generative process as the underlying process that gives rise to the observations, this would seem to be more adequate to the extent that it can better describe the measurements. Generative model similarly is being evaluated based upon its ability to fit to the generative process as well as phenomena of interest. So not just to fit visual data, but maybe to model something like some illusion in visual. Then there's a wide range of more general statistical modeling techniques. So first is like the ultimate grab bag, which is just how relevant is the overall modeling. Then there are statistical evaluations of model adequacy and model selection. IQ key information criterion, Bayesian information criterion, Bayes factor, hierarchical likelihood ratio test. If you have a parametric model, bootstrap and non-parametric statistics, just general modeling. And then taking that more out of the statistics into like the engineering model lifecycle, then that's where you can think about validation, verification, validity, quality, et cetera. Systems engineering model lifecycle. Any other thoughts or questions on this? Yeah. Okay. So I'm coming from kind of an early career data scientist background here, but I mean, as far as I mean, what I find attractive about active inference and doing things as generative models is the attempt to, you know, describe or explain what's going on in the data. And that being said, you know, in terms of a lot of many applications of like data science, machine learning, the emphasis is upon prediction. And so are these listed statistical metrics you have here, base factor, BIC, et cetera? I mean, are those basically like your error measurements? Are they your, how to say, just attempted at measuring like accuracy of prediction or, or is there some trade-off between explainability and prediction? Maybe that makes sense. Yeah. Yeah. Just to kind of summarize these. So when you have nested parametric models, then you can evaluate whether two models are better…