Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right welcome back coort 7 we're in our second discussion on chapter two so let's head over to two we can begin with anyone's thoughts or Reflections or any question or idea or section of the text they want to begin with OR curiosity from to or just some quote they want to go to anyone can just raise their hand or type it in the chat okay actual go for it sure so one thing that I found a bit interesting it's on page 26 of the book um it says the person can resolve this discrepancy in two ways it's talking about this thing of looking at um like expecting to see an apple when you look and then actually seeing a frog and it says first she could change her mind about what she's seeing so actually that it's a frog instead of an apple um or she can fiate the nearest apple tree and see something that looks very much like an apple this resolv the initial discrepancy but in a different way um so she's kind of doing an action to make the world fit more her model and they they do another example afterwards but I I'm kind of curious about exactly how this tradeoff is done when would you choose to do one thing over the other when would you choose to do an action versus when would you choose to do a um like an update of your internal model uh especially in a situation where you don't know you know perhaps if you've had this situation happen many times you kind of learned that it's important to know that it's a frog uh so you don't want to look away you want to up your model but what if you're meeting it the first time I was a little bit confused about that great question anyone have a thought on this you're likely doing both in parallel much of the time somebody who's too fixated on one or the other that's tends to be a neurotic problem or Worse what that's greate um but there a few few ways to take this first there's the secret third option which is neither discrepancy doesn't have to be resolved but it is kind of like the equivalent of a a ball on a shelf it's like an unresolved potential energy or like an unburnt candle it's like an unresolved Gibs free energy differential however it isn't the case that a discrepancy is resolved one of these two ways however these are the two ways that there can be the reduction or of discrepancy or or discrepancy continue to increase so then the question is like okay in in what ways uh are or could be discrepancy resolved through inference and through through action um there's defin ways to take this um in the in the particular partition so that's splitting up into these four nodes these two nodes that constitute the the two directions of the blanket and then the internal and the external states that are partitioned off from each other from that blanket those are all happening simultaneously so one answer is they're both separately in play so you could have a situation where where neither uh belief updating nor action updating were happening like a rock you could have a situation where both were concurrently um updating sometimes it can be helpful to think about these kind of multiple time scales of what changing the mind can look like the fastest time scale which is kind of equivalent to the forward pass on a neural network is just the the forward inference that's just all the kind of parameters are staying the same and then new observation comes in basy and update cranks then at a slightly slower time scales um corresponding to attention and learning mechanisms there's kind of like those are corresponding to like weight updating in a neural network and so those those are like slower kinds of changing the mind um changing the world through action also like that's very dependent on what the kind of action is for example here the um eyes were moved so it's a muscular action to bring the observations into alignment with the expectations however there wasn't a change in the world so visual activity is kind of a special kind of action because it doesn't modify the world but it is still a way to reduce…