Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

right greetings everyone Thanks for joining we're on eight eight in our second discussion on chapter four so one question was uh brought up that will definitely come to but is there just anything else that anyone else wants to add about for any just reflection or any question that's kind of surfacing that we want to add to the stack all right let's go to the message passing okay so in message passing is there a decay of information as the distance between the variable X and individual Markov blanket constituents increases is implementation of information Decay and message passing an option for model implementation who wants to give a a first thought on that or any any other kind of curiosity around it um so would it I'm not sure if I'm understanding it very well other than intuitively but would it I mean would it be the case that uh like information Decay wouldn't necessarily um I suppose makes sense here given the the models they're already working uh in in terms of like time steps and so you're already kind of um you know you're you're recomputing uh your your gradients and your posteriors and so on in every given time step such that um you know the idea of some kind of uh holistic piece of information that's being carried uh from time step to time step doesn't quite making sense that is each time step involves an entire set of like recalculations does that make any kind of sense yeah um so just to start with one uncertainty is I'm I'm not exactly sure about message passing in the continuous versus discrete time setting um again hopefully something that we we can unpack and talk to Magnus kudal at all since I know that he's an expert on the continuous time um setting but any other thoughts on this or I'll I'll try to give uh an answer Ollie uh yeah just one quick note about message passing is I mean before going into the details of how exactly message passing happens in active inference framework uh it's probably worth noting that uh it's inversely proportion proportional to the robustness of any uh self-organized self-organization systems because you see uh if we consider an agent and uh it's um I mean peripheral um systems or peripheral um let's say limbs as a kind of central like fully centralized system uh in which uh I mean the message passing happens uh almost uh in a fully centralized way from uh the brain or I don't know the central processing unit onto the limbs then this system would not be as robust as the system in which uh this kind of message passing is is distributed among um among the elements of the system so I think that's one of the reasons why we see uh signal Decay or I mean the loss of information as we go up to the hierarchy and uh I mean the more we go up to the hierarchy the less precise the information becomes uh but in order to have more and more precise information we need to somehow acknowledge a kind of autonomy onto the peripheral um nervous system as opposed to a central nervous system so uh that's one of the ways uh I mean in humans and nervous system is organized in which this kind of distribution of uh agency I mean autonomous agency is distributed among both peripheral system and the central system so [Music] um and coming back to active inference framework um I also believe in this framework it beautifully captures this kind of hierarchical model because obviously as we move farther away from the peripheral elements of any system the elements which have direct access to the direct access but closer access to environment then the the accumulation of probabilistic Errors becomes larger and larger and that's why um in order to have a reliable information we need to discretize these kinds of information at some point without uh necessarily having to take it into account every single bit of information that we receive so uh on the lower levels of the uh I mean when we're dealing with the lower level systems we mostly deal with continuous time models which allow for more precise kind of…