Wellbeing, Modelling, Precision

hello everyone and welcome back to active inference insights today I have the real pleasure of speaking to Ryan Smith Ryan is a principal investigator at The laua Institute for brain research and he's the research associate professor at oxy College of Health Sciences at the University of Tulsa his work focuses on the neurocomputational mechanisms underlying emotion cognition interactions and how they relate to brain body interactions with special focus on planning Information Gain and interceptive inference he is also the primary author of probably the best known and most highly rated tutorial on active inference writing it alongside KL friston and Christopher white Ryan thank you for giving me your time uh welcome to the show thanks for uh thanks for having me excited to get Talk No it's it's an absolute pleasure it's it's really very exciting as I said I mean obviously I'd come across your work uh kind of inevitably for me to have invited you but it was actually a real treat to sort of go back and go through like what feels like a you've done loads and work at this incredible velocity um so it's been really fun to dive in and explore this I guess what felt like a more empirical side um because I'm spending most of my time reading sort of philosophy papers so it's great to see some like traditional statistics and some proper empirical work so so it's been a great it's been great fun on my part I wanted to start at kind of a pretty broad level and and speak about subjective well-being so you wrote this paper in 2022 sort of grounding subjective well-being in the parameters and mechanisms of active inference um you alongside Cas Beal of uh cesp have come up with this really nice way of grounding uh well-being in expected free energy but as you write in that paper you say um while variational fre energy can be understood to track basic biological Fitness the this is neither necessary nor sufficient for an individual to report High subjective well-being so maybe where we could start for the audience is just explaining why the minimization of variational free energy is not sort of the simple solution to obtaining High subjective well-being sure yeah and I mean it's a definitely a very kind of complicated and yeah complex complex topic jump into and it honestly depends a little bit about the exact kind of formalism that you're kind of working in I mean me I almost exclusively work within the um sort of like pomdp formalism um associated with decision-making and planning processes that uh that um you know as the focus for our tutorial and you know the focus of the tutorial was very much like trying to get people to a point of being able to do true kind of experimental empirical applications right I mean that's the final section is all kind of meant to transition into that um just because that's primarily what I do um so so I mean with respect to the kind of um yeah your question here is that within the p P scheme um you know variational free energy is primarily used as a means of optimizing posterior beliefs about States right the state inference um sort of process whereas expected free energy is the thing where you're trying to say okay what do I expect the kind of rewarding of the outcomes to be and how much information do I expect to gain you know conditional and choosing X versus y um so that there is a true kind of veency thing right in the expected free energy because you have this expected Divergence or lack thereof between the things you want and the things you expect right under a policy whereas the variational for energy is much more um a function of it's just a function of the observations you're getting and how you update beliefs about the most sort of appropriate approximate posterior over those States right so in other words you know if I perceive that uh you know the observation I'm get getting is that you know someone's like stabbing me or something right then the the process I go through to minimize variational for…