From pixels to planning: scale-free active inference

Hello and welcome. This is Active Inference live stream 58.0. October 29th, 2024. We're discussing the paper from pixels to planning scale-free active inference. So welcome to the Active Inference Institute. We're a participatory online Institute that is communicating, learning, and practicing applied active inference. You can find us at the links on the slide. This is a recorded and archived live stream. Please provide feedback so we can improve our work. All backgrounds and perspectives are welcome and will follow video etiquette for live streams. Head over to activeinference.institute to learn more about the Institute, get involved with projects like live streams, and a lot of other things. Okay, before we jump into the paper, let's say hello. I'm Daniel. I'm a researcher in California and really looking forward to this discussion. Arun, want to say hi, give any context, and thanks again for contributing to the dot zero. Hi everyone. My name is Aaron. I'm a software engineer in the UK. I did my PhD a good few years ago in which I touched on dynamicals of modeling and active inference. I've been thinking about it for a long time since then. I came across this paper and thought it was really hard. And so I requested, guys, can we do a live stream on this? I'd really love to understand this paper properly. And now I've somehow ended up contributing to the live stream. So yeah, all good. Okay. So we're discussing this paper from July 2024, from Pixelist to Planning, Scale-Free Active Inference, by Carl Fristen and colleagues. The rest of this video is going to be some background and context and an introduction and an overview of the ideas. It's not a review. It's not comprehensive. It will be part of an exciting series where in the coming weeks in the dot one and dot two, we discuss with the authors and the dot two is going to fall during the applied active symposium. So that'll also be really fun. Okay. Let's go to the big questions. What are your big questions or whether it's about the paper specifically, or what are the big motivators that made you excited and want to pursue the these affordances? So I think the big one is, you know, the title, right? And I think the question of scale is one that's plagued a lot of active inference, implementations and previous papers. In the past few years, everyone's like, Oh, yeah, I'd love an active inference agent that can do stuff. But if I have more than a few states, I get a combinatorial explosion. And things are fall over or they take a very long time compared to alternative methods that are out there. So the question of scale is a really important one. And I think this paper nominally looks like there's a way around that. Also, the renormalization group as a way to do that is, I think, a very interesting one. It draws from a lot of physics that I actually studied a very, very long time ago. So that really spoke to me and I thought it was really, really nice to see some things that apply to also like icing model of ferromagnetism, some other key parts of thermodynamics that now come into AI, which active inference has been doing for a while with the free energy minimization part. But I think the renormalization group is a really interesting part of maths that seems to get you a lot further. I also had some like much more technical questions, which maybe don't fall under the big questions, which will definitely come on to later when I was reading this paper and going, I'm looking at it and I go, I don't understand actually what's happening here. So I'd really love to dive into those and particularly catch up with the authors in the later episodes to get some of the more technical answers on how these matrices and tensors play together, which seems to differ a little bit to the textbook by par and some of the other formulations by the RxInfer group as well. So actually getting into the nitty gritty of how we implement this will be really fun for me. Cool. Blue also contributed to…