Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hello, thanks everyone for joining. It's March 29th, 2023. We're in Cohort 3, having our first discussion on Chapter 5. So, before we go into any questions or look at the text, does anyone want to just provide any general reflection or thought on Chapter 5 or just the first half of the book up until now? Just unmute and go for it if you want to. Okay, let's look to Chapter 5. And then I'm sure people have many thoughts as we get going. Any general thoughts on 5 or on this opening quotation about plants and animals? Well, there are plants that move and there does seem to be some thought that there are interconnections between plants, that they have something that, while it is not an animal central nervous system. There are patterns within the architecture that have similar functions and it's the timeline over which they operate as much longer. I very much agree with that. Here's my PhD advisor, Deborah, and a paper from 1989 describing the plant behavior idea. And then, I believe a more closer to our area is this 2017 paper on plant predictive processing with Friston. So, agreed. This quote, it's a little bit of like timescale dependent. It's just kind of like a... It's a fun first pass, but isn't that interesting that it's like... It's actually not so stark. Or the plants that do respond, like the Venus flycatchers and everything. But here in this chapter, we're going to be focused on the mammalian nervous system and body, not the morphology of the nervous system or morphological development. Okay. Anyone, please just raise your hand or write something. Otherwise, let's just go through chapter five and just look at some potentially interesting pieces. This is an interesting concept that not every variable depends on every other variable. And that's sparsity. And so, there's sparsity in the brain. There are many connections, but of the space of the possible, it's quite sparse. And similarly, in statistical models, the sparser they are, the easier they are to fit. Doesn't mean that it's going to be better. And a lot of modeling is about finding the trade-off point between like model complexity and model accuracy. Because if you enable every edge, it's going to be a really laborious calculation. But if you restrict edges that are actually important, you might have a very rapidly fit model. However, it might not reflect some of the important true causal relationships amongst variables. So, that comes up in different ways. Let us take a step back from the technical material of chapter four, everyone, breathe a sigh, and turn our attention to the process theories accompanying active inference. Here, they point to the difference between a principle, free energy principle, which is that systems are described by their dynamics on free energy landscapes, and that persistence is associated with reduction of free energy, to the extent that agency is exerted, and a process theory about how any given system may actually implement those kinds of dynamics. So, a principle cannot be falsified. And this is at the root of a lot of questions and commentary around like the falsification capacity for the free energy principle. And the simplest thing, as far as I've seen, is neither can a linear regression be falsified. It's just not within the space of being confirmed or rejected or falsified by anything at all. It's just a proposed model framework as a principle. Now, any given linear regression, which is a specific hypothesis about how that principle of linear regression or the principle of the L2 norm might be implemented, that given linear model might be adequate or not. More adequate or totally inadequate or however, but it is more subject and answerable to empirical data. So, free energy principle is not really in the space or the game of falsification or not, nor does it even engage with empirical data directly. Whereas, proposed and manifest models are directly addressing empirical data and empirical data speak to their validity. So, none…