Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hi, everyone. It's August 1st, and we're in Cohort 4, having our first discussion on Chapter 4. So before we go into more detail on Chapter 4 and everything, does anyone have any just overall thoughts or comments on Chapter 4? What did they expect it would be delivering as a chapter? How was it? Or what was any aspect that people remembered or anything? I found this chapter to be rather mathematically dense, I suppose, and the authors do warn of that at the beginning of the chapter. I mean, I think it's great so far as it starts to introduce how they're using Bayes' theorem, how they kind of exploit Jensen's inequality to move in the direction of calculating free energy, using logarithmic functions in order to do that. You know, the initial models of the static perceptual tasks also seems like a really nice introduction to moving on to POMDPs, as well as the continuous models used in active inference. So, you know, in a way it felt like it kind of started with nice first principles and built up from there. But that said, I can start to see maybe the necessity of having a math learning group for some folks here, given this chapter. Thanks. Yeah. Anyone else? Any, like, just overall thoughts or? Well, I would just totally agree 100% with what Andrew shared, basically. It's just similar thoughts. Yeah. Yeah. I mean, basically the, the, the, the maths, uh, are very well explained and it goes gradually, but I think it would be beneficial to, yeah, to go even deeper and slower through them, maybe separately. Definitely. Any other thoughts people have? Yeah. Yeah. Darius. So then anyone else? Yeah. Daniel, it's less a point about the chapter, although I agree with the previous sentiments that it is mathematically dense. It's more of a point about the math learning group. I think myself and several others just having trouble locating it, um, in terms of where it actually is on the discord. So the past couple of weeks I've been trying to get into some of the voice chaps and it doesn't, there doesn't seem to be anything kind of happening on them. So there might be a bit of confusion. Yeah. Thanks. It is, uh, people can use the general voice chat, just the top voice channel. And, uh, you know, if there's no one there, just, just still jump in and stick in there. Okay. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. I'll try to join when I can, but it can, I mean, there's so, so much work on the more introductory side, like kind of just going from the fundamentals on through all the work that Jonathan and Ali and others are doing too. So there's no lack of things to do in math learning group. I hope people do like connect on it. Younghoon? Wait, I don't hear you. I can see you're unmuted. Maybe frozen though. Can anyone else hear or no? Okay. Okay, Younghoon, when you're back, feel free to continue. Any other general thoughts on four or anything writing in the chat? Okay. Just another small specific observation that's a, as a kind of a practicing entry-level data scientist, the way, something that helped me understand the static perceptual models earlier on is that it most feels like it's just describing something like the process of binary classification. One of the models in the figure 4.2, the very first one. You have the example of like your prior being prevalence of a disease. Is there a disease? Yes or no. You have a true positive rate, true negative rate. You end up with the result of your test. I found that to be rather interesting and useful. Just are we trying to predict does someone have a disease or not? There's more that goes into that step by step. Yeah. Thanks. Thanks. That's one of the classic Bayesian examples. Like there's a test that's 99% accurate and there's a disease that one in 10,000 people have. What is the probability that you have the disease if you got a positive test result? And the kind of snap answer…