Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right welcome everyone we're in cohort 7 the second discussion on chapter8 so there's a few ways we can go we can look at chapter eight in the context of the book um in terms of continuous time modeling uh following up on chapter 7 from discreet time modeling also uh after or just going to it I can share some updates on continuous time modeling from this morning in the RX and fur group where we um developed on the new example from the RX infer team with the lorenza tractor and an example with the neural network so we could talk about that and uh modify it in the context of continuous time modeling um first though let us just go to the chapter does anyone want to just share a section of chapter 8 or any question or quote or or image or anything like that I have a question okay Andrew then John uh yeah so they talk about um this attractor like this point that things tend to but they seem to talk about it on two different levels like one in reality and then one in the maybe the generative model I guess like and so if you could clarify that distinction um how those work and what's their relation yeah so when you specify the entire simulation you will have let's just say we're dealing with with some with temperature in the room we could have situation where just to to get into this Y is our observable X is going to be our latent State x dot is the rate the change landscape on the Laten state so this is like the thermometer measurements coming in on Y and the Laten State on X and the change in the temperature um you could have a room with a real attractor setting like there's there's actually a variable uh v such that when the true latent state is exactly V the the um rate of change is zero and when the root temperature is higher it drops back to the point of tractor and vice versa for being lower then you have the agents model of the external process that could be built to be structurally identical like they're both dealing with a point detractor in which point uh the agent model will tend to converge pretty simply to the correct parameters but you could also have a situation where like the real room is on a limit cycle so it's just oscillating or it's by stable but the agent's belief is that there's a point attractor so to answer it this is just one example of attractors there's Point attractors limit attractors uh and so on and in practice you end up specifying the actual generative process that gives rise to the agent's observations and then you specify the agent's model of that process and those can be structurally identical or not so so the attractor is basically a construct of the a of the organism in the way that you're describing it like it's something it would wish for you know which may not exactly even exist there but may be useful nonetheless is that yeah I mean on on multiple levels there's there's the the modeler making the temperature model of the room and possibly just choosing to use an attractor style dynamical model just just as a map and then there's the agent within the simulation using her istics or constrained families to do variational inference in and and that could be simplified on up to any point and but the crucial thing is that is the attractor in the agent's mind that that there's something there they ch chasing like they're trying to tune into and at least of understanding and for me that's that's very uh relevant uh because I'm trying to do this like two mind uh dialogue where like you'll have a a continuous mind that's having some kind of affordance or ability and then the attractor would be like a concept and then that could feed into this like discreet mind let's say uh that you know has connections between Concepts let's say but so it seems like an attractor can serve as what I would call a concept and the process of approaching that would be like some kind of skill or ability or affordance well yeah I mean in figure 86 this is like a classic hierarchical architecture where the lower level…