Realising Synthetic Active Inference Agents, Part I: Epistemic Objectives and Graphical Specification Language Realising Synthetic Active Inference Agents, Part II: Variational Message Updates

All right, hello and welcome everyone. It is October 24th, 2023, and we're an ACT-INV livestream number 55.0 on realizing synthetic active inference agents. Okay, 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 some of the links on this page. This is a recorded and an archived livestream, so please provide feedback so we can improve our work. All backgrounds and perspectives are welcome, and we'll be following video etiquette for livestreams. Head over to activeinference.org if you want to learn more about participating in livestreams or other activities. All right, well, we're in livestream 55 series with a goal to learn and discuss these two very interesting papers on realizing synthetic active inference agents. Part 1 on epistemic objectives and graphical specification. Part 2 on the variational message updates. As with all videos, it's an introduction for some of the ideas, not a review or a final word. We're going to introduce ourselves, then jump into a fairly lengthy background section that will prepare us to ask the questions and get to a place where the paper's contributions can be figured out. So, let us begin with introducing ourselves and saying hi and saying maybe something that was exciting to us or made us want to participate in this series. So, I'm Daniel. I'm a researcher in California, and I was interested to go a little deeper on message passing. It's something that's brought up a lot in the textbook and implicitly in other papers, but this was an opportunity to tackle it head on. And I'll pass to Bert. Yeah, so I'm Bert. I study civil engineering in the Netherlands. And I struggle with the math of active inference. But recently, I picked up reinforcement learning. And together with this paper, I think it really helps. Jacob? Hi, I'm Jacob. I'm also a researcher in California. And I'm really excited about this paper from a number of different angles. I guess the graphical notation and the notation that the paper introduces, I think, can have really profound impact on the field from both a computational and a theoretical viewpoint. And I'm interested to learn more about the implications of the new notation for the research. Yeah, new notation just dropped. Okay. There's a pair of papers, as mentioned, and the information is here. So, each of us can phrase the big question that brought us to the paper. But I wrote it this way, which is, right there in the title, there's at least a triple play, a triple pun tondre. And there's a diad of papers. So, what is this realizing in the context of the title? Well, in one sense, we're realizing something in terms of implementing it or manifesting it. They're deploying something that is being realized. So, there's an accomplishment sense of realizing. Also, the work calls attention to our own realizing process, our relevance, realization, how we come to appreciate and interact with synthetic intelligence. Ours. And then, we're also talking about building agents that do some kind of realizing in themselves, like realizing agents in that sense. So, what kind of inning starts off with a triple play? I don't know. Bert or Jakob, what big questions brought you to the paper? What do you think the paper takes on? I think in terms of creating scalable active inference models that are reproducible across a variety of settings is quite exciting. So, maybe realizing in that sense across a number of different domains is part of the meaning here. And, of course, the triple play that we were exploring with our own work from going from just a simple graphical representation to a mathematical description of the generative model and the algorithm for message passing and updating the generative model through time to then a code implementation that can be deployed in various dynamic settings. That is also what I think is an important part of this paper. Oh, yeah.…