Active Inference and Artificial Spin Ice: Control Processes and State Selection

All right, hello, welcome. It's August 19th, 2025. We're in active stream, guest stream one 17.1 with Robert Stamps, who will be presenting on control strategies and predictive coding using magnetic noise. Thank you for joining Robert and looking forward to this presentation and then discussion with... Close the YouTube on your side, Robert. Okay. Yeah. Oh, shall I just begin now? Okay. Yes. All right. Well, thank you for this invitation. Maybe I should give you a little bit of a background of how I ended up doing this research, getting interested into this problem because I'm primarily a condensed matter physicist, computational in theory primarily. And my area of specialization has been in magnetic materials and spin systems in general. But I've been doing this for quite a while. And about, I would guess, 25 years ago, I was more or less starting my career. I was in Australia at the time. And I also came across a paper. It was about the theory of flocking. And I found that very interesting. And that led to one thing and led to another thing. And eventually I found myself co-supervising a PhD student who was working with an expert in psychology on psychophysics. And there's looking, doing experiments and analyzing experiments on primate visual system. And what I found striking was some of the mathematics we were doing. We were heavily influenced by, I think about that time, in fact, just recently, a Rowan Ballard paper had come out. And it was heavily influenced by the structural form of the mathematics we were doing. And it mapped directly back on to some of the things I was thinking about doing with magnetic systems. At that time, people were just starting to develop a fabrication technique which would allow them to make new kinds of magnetic materials. And that's what I'm going to discuss today. And it occurred to me at that time, I remember very clearly that when that technology matured, it seemed to me like it would be an ideal one to try and implement some sort of neuromorphic designs in. And indeed, that is exactly what is panned out. And so I'll tell you a little bit about that and what the state of the field is now. And that's, but that explains more or less how I began my interest in this kind of cross-disciplinary field. And just to put the story back into context, then about three years ago, I had a sabbatical, and I decided to go back to Australia and visit that group and see what had happened. And that's when I came across Friston's papers and active inference. And I actually didn't make head or tail of them. And so I thought, well, I better put it into some context I can understand and try and understand exactly what predictive coding and active inference are about. And that helped me. And that's where this work emerged from, trying to understand that. So let me, let me begin then. So I'll just explain these pictures. This is just a schematic of using nanomagnets with an active inference to control some process and then you feed back. And this is a schematic of the sort of nanoparticle arrays, which are in this control box. This is my idealized spin system. And what this is, at the very end, I could show you, it's, it's a model, which is a bit more sophisticated, but it actually tries to model properly the properties of real magnetic materials. And the point of it was to demonstrate that, you know, one could implement certain algorithms inside a magnetic context like this. So to begin, let me introduce myself, the people I work with here, focus primarily on studying both theoretically and experimentally nanomagnets. The experimental parts are led by my colleague, Professor von Lierup. And these are students who involved one way or another in some of the works, some of it overlaps with what I'm going to show you today. And these are just illustrations of some of the different systems which other people and which we also have looked at over the years. These, each of these represent a little magnet. But before I…