Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right welcome back cohort 6 Andrew however you want to begin this discussion on chapter two go for it sure um probably yeah pulling up the textbook might be a good move um I know that this is week two of chapter two so I don't want to uh reintroduce the entire chapters if you haven't already spent a week on it but I thought it'd be nice for anyone who wasn't able to attend last week to just kind of quickly skim what's going on um so yeah this is chapter 2 the low road to active inference in chapter one we were introduced to the idea that there's a high and low road to inference so um the low road is more about the um sort of the the the what or the the how of of active inference and um as opposed to the high road will where there will be more disc discussion on the why and the why of like why is this happening why are agents doing this how do they survive that's kind of the questions getting at here this is getting much more into sort of something like the mechanics of the process um and it picks up uh with the helm Holan perspective which is a 19th century late 19th century tradition uh that just began on viewing uh perception as unconscious Ence uh this idea is picked up more in ideas such as the Basi and brain hypothesis later on in the 20th century and from just an act uh active inference perspective we're treating not just perception but also action planning and learning all as basian inference problems as well as deriving a variational approximation um for these problems to overcome intractability problems um with the helm holian tradition the kind of Distinction about this versus other historical cognitive science Traditions is that rather than viewing perception is only a bottomup process of sensory States being fed into uh being turned into internal representations of the outside or some such thing like that instead perception is viewed as an inferential process that combines topown prior information about the most likely causes of Sensations with bottom up sensory stimuli so we actually have this as opposed to um perception being more of an outside in process it's more of a it's it's both right it's bidirectional um we're introduced to baz's rule which is just um incredibly crucial to active inference as well as just many other kinds of probability problems I strongly recommend those who are wanting to get into the maths to kind of work on um just kind of understanding baz's rule almost as like a sort of Cornerstone from which to begin um it can be viewed as a posterior uh equaling and uh a fraction that is uh the numerator is a prior and a likelihood and then your denominator is evidence and so what can happen is that you can update any of these Quant quantities in order to solve for another quantity right so this this this can be updated over time by taking in new information that then update everything else you can update your priors your your posteriors which are your beliefs at the end like at the end of the process um and so what's nice is that we can actually map all these quantities uh onto the modeling process and active inference it's so a generative model in our case like an agent um it could be a human being it could be uh some other biological organism or otherwise the generative model is just a combination of PR uh Pride PRI and um likelihood mappings so priors would be something like probability of X in the textbook um in often times uh notation can also be like uh probability of O which stands for or excuse me S which stands for latent States or hidden States um likelihoods are probability of Y given X so y condition on X or in other literature can be viewed as um observ ations o conditioned on S hidden Laten States so we have these other quantities um involved and yeah all of those are updated using bases rule um we have some additional um information in that section box 2.1 goes over gives a brief refresher on probabilistic reasoning some Rule and product rules for anyone who wants a little bit…