Learning in Physical Systems

Hello, welcome. It is November 5th, 2025, and we're in Active Inference Guest Stream 123.1 with Marcelo Guzman, learning with physical networks. So Marcelo will give a presentation and then I will read some questions from the live chat. So thank you for joining. To you for the presentation. All right. Thank you, Daniel. Thank you for the invitation. So I'm Marcelo Guzman. I'm a postdoc at the University of Pennsylvania. Today I want to talk about learning with physical networks. So one such example is the one that you're seeing to your left in your video. This is an electronic circuit. Okay. Seems kind of complicated, but this electronic circuit has the property that's able to self-adapt. And by doing that, it can learn machine learning tasks like a computer, like an artificial neural network. The topic of my talk is going to be centered around two papers that appear this year. One is microscopic imprints of learn solutions in two novel networks, open source in PRX. And the other one is physical networks become what they learn, which is in physical review letters. And in a nutshell, what I want to tell you is how physics shapes the way these machines learn. All right. And hopefully by the end of this talk, you're going to understand this picture that is appearing to your right. And I'm going to explain it to you later. So let's begin with, with learning. So when we think about learning, we have these two big paradigms. Right now we have this very trendy paradigm of machine learning. Okay. And on the other hand, we have our brain or what happens in biology, what I call biological learning. And these two paradigms are very powerful, although they are very, very different. Okay. And so here I'm listing some of the differences that, that, that, that are in between these two. For example, in artificial lunar network, what you usually do is to modify the weight parameters, current technologies, current, uh, large language models have around 10 to 12 parameters. Everything is digital. Everything is happening in your computer or, uh, in, and digital centers. It's very energy demanding. Everything is centralized by CPUs or GPUs, and it lives in the virtual world. There is no physics. Okay. On the other hand, if you consider the brain, what you do with learning is you modify this synaptic strength. You have around 10 to the 14 synapses. Everything is analog. And from time to time, you go to the, to the digital domain. You have these spikes, very, you know, efficient. Okay. So to operate the brain, you just spend as much power as a light bulb. It's distributed. So, that is, you don't have any, uh, any CPU or DPU and it's subject to physical loss. So, but learning is a much broader phenomenon. Okay. And what I want to introduce now is a third paradigm that we call physical learning. So these are two, these are, this is one example of a physical learning system. So this is a resistor network. Okay. So all the edges are resistors. This is the experimental realization. And what this resistor network does is to modify its own contact lenses. All right. And there, and therefore by that learning, uh, machine learning task, it has around 10 to the two parameters. This is a very novel technology. Uh, it's completely analog. It's energy efficient. It's fast. It's distributed. And what I want to highlight is that it is also subject to physical loss. All right. You cannot circumvent those. So why do we want to study this kind of third paradigm? And there are two main reasons in the community. One reason, which is more on the engineering side is to improve the current technologies of machine learning. Can we make, uh, artificial neural networks more energy efficient? Uh, can we increase the speed? Can we make it more robust to damage, but there's a fundamental part of it. And this is, uh, what, uh, what I'm interested in is that maybe by studying this kind of systems, we can understand some more complex biological phenomena. These physical learning…