Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right welcome back to cohort Andrew take away really quickly can you all hear me okay perfect um yeah so hey everyone welcome back uh cohort 6 just resuming here Midway through July uh we're going to be starting the second half of the textbook uh first half we were introduced to a lot of the underlying ideas and mechanics behind active inference uh Bas role Basi and brain hypothesis uh hidden States being inferred from observations uh action as a means of trying to change the hidden State through transitions uh a lot of really kind of foundational ideas uh towards understanding active inference as well as following along uh the low road active inference starting from basian can statistics as well as the hod active inference following the uh principle of free energy minimization so the second half of the textbook I think is um it's great for better understanding how to more directly apply active inference so chapter six a recipe for Designing active inference models right off with that kind of gives you a sense of what the rest of the textbook is actually going of potentially be about we'll be learning how to design models uh chapter seven and eight will be about uh putting together discret and continuous time models respectively um so that's great for people who want to um you know figure out which kind of model they need for their particular use case it's nice that there are rather expansive chapters on both types of those models chapter 9 uh gets more into analysis so once you've set up a an active inference simulation uh it starts to discuss like how you can actually interpret your results um depending on your use case like how you might set things up depending on what your particular question or inquiry is whenever you're setting up the simulation and finally chapter 10 is a nice uh wrap that is maybe slightly out of place compared to the rest of what I said about uh the second half of the book but it's uh it's good it's it's a rather holistic kind of uh conclusion type chapter both wrapping up the book comparing active inference to a variety of other disciplines and Fields like control theory uh other social sciences psychology uh reinforcement learning as another approach these kinds of problems and so on so uh just a quick overview on chapter six again a recipe for Designing active inference models uh the book the uh the intro to this chapter uh friston at all WR uh the challenge is not to emulate the brain piece by piece but to find the generative model that describes the problem the brain's trying to solve the model provides it so that refer two is the model provides a complete description of a system of interest and the resulting Behavior inference and neurodynamics can all be derived from a model by minimizing free energy all these ideas should be familiar from the first half of the book but it's just more clearly stating yes active inference is heavily inherited from or derived from uh different ways of approaching computational Neuroscience that said we are are not trying to recreate you know um billions of individual neurons and all of the kind of crazy uh sparsely connected architecture underlying say the human brain or otherwise the empasis upon developing generative models um you know this kind of gets into the difference between complexity and accuracy found in uh any machine learning related problem re enforcement learning problem so on so again not trying to recreate the entire brain uh just to where we can get things accurate enough to where may start to track and potentially could even be fitted to empirical data um so whenever uh we'll be introduced over time to more Exemplar kind of models but we can adapt the form of the models presented in this textbook to our respective problem of Interest we can change their form uh should they be shallow or hierarchical we can modify the variables involved I.E uh the beliefs so what are the states to infer what are the observations that the agent receives what…