Applied Active Inference Symposium 2025 part 2

welcome back it's day two of the fifth applied active inference symposium november 13th 2025 and the first session will be the future of active inference.jl with peter jonathan and samuel so peter thank you to you thanks daniel all right it's state please mute the youtube youtube sorry i have some sound problems here thanks one second okay here we go again can you see the screen otherwise i'll just reshare it and and there we are thank you daniel sorry for the the sound issues there i had another tab open that i shouldn't have had and also thanks for organizing this symposium which is incredible the work that's been put into it and worth being grateful for so just that's worth saying probably a lot of times during this symposium again and again hi everyone so today i and sam and jonathan will present on the ongoing and upcoming developments of active inference.jl which is a julia software tool for running and applying active inference models to data or in simulation and i'll start out with just a background and a conceptual thing and then jonathan will take over and sam will take over and they will go through code and give a demonstration of how to use the package in its newest version that's not yet released but and there's a link in the youtube channel if someone wants to follow along the code to a github repo which can be cloned and used with this new experimental version of the package so um i'll briefly to start with background we're conceptualizing conceptualizing active inference as a kind of cognitive model and what a cognitive model means is just that it's a formal model that produces behavior of some kind so all of artificial intelligence et cetera can be conceptualized as cognitive models and this is broader framework where you have some agent and you have some environment and the environment produces observations to the agent the agent produces actions that affect the environment and then the agent has a cognitive state called the theta here that changes over time that could be its beliefs about the environment and um the agent has some cognitive parameters that could be for example learning rates or its priors that govern how its cognitive states change and govern how it produces actions and in general cognitive modeling and therefore also active inference has three kinds of uses that i know of one is theoretical modeling so that's when you run simulations to see and what a given agent would do given different parameters so there you know the theta here and you simulate the environment and you want to see what happens and active inference is full of these so all the simulation work where you're going to show what happens if active inference fish swim together etc could be considered theoretical models and much of the backbone of the field is built on this then there's applied cognitive modeling and that's where instead of having the parameters you have the observations and the actions so you have the behavior for example because you recorded it from a human in an experiment and you want to reconstruct the behavior that govern the parameters that govern the behavior you want to find out what was the general model or what was the prior so what was the learning rate of a given participant that's what's used in computational psychiatry but also much beyond that and finally one uses artificial intelligence where again you have the parameters but you don't simulate the environment you have real environment for example you know burning house or whatever and then you want your your artifact your tool to generate the right kinds of behavior for example a vacuum cleaner that's supposed to move in some ways and not other ways you can use active inference in all three ways our lab specializes in applied cognitive modeling so the second of these three and last year we focused quite a lot on how to use active inference for that but this year we will not because instead we will focus on new structure of the package and how that…