Untitled

Okay. Hello, welcome everybody to the Active Inference Institute's second episode of our Morph Stream series. So these sessions are intended as a space to showcase and discuss the intersection between neuromorphic computing, so brain-inspired computing, and active inference. If you're new to neuromorphic computing, I did give a short primer in the first session, which is available on YouTube. So the format is that our guest speaker will present in the first part, then there's a Q&A, and then a wider conversation for the rest of the call. So our first ones featured David Cappell, who's kindly joined us again today. Thank you, David. And he gave an absolutely fascinating talk about a free energy model to uncover the role of noise in the brain. So it's like active inference at the synapse. And today's episode features Alon Loeffler. He's a postdoctoral scientist at Cortical Labs in Melbourne, Australia. Cortical Labs being the company behind the very interesting DishBrain paper, which recently came out. So he specializes in synthetic biological intelligence and has a PhD in neuromorphic nanowire networks from the University of Sydney. He also has almost three, sorry, over three years of experience in designing learning tasks and algorithms for brain-inspired systems. And today he's going to be presenting some really interesting research on neuromorphic nanowire networks. And so if you want to share your screen. Absolutely. Thank you. All right. Can you see this? All good. Yep. All good. All good. Fantastic. So thank you very much for having me. And thanks for the wonderful introduction. Not much more to say about myself about, so that was a perfect summary. Yeah. Today I'm going to talk mostly about my PhD research, which I finished last year from the University of Sydney, which was in neuromorphic nanowire networks. And that's quite a mouthful. We'll definitely get into what that is. Don't worry. Hopefully by the end, you'll be experts. And I'll also touch a bit at the end about what I'm working on at Cortical Labs at the moment. And there's quite an interesting link. They're slightly different, but hopefully we'll get through that. So just a bit of an overview of how today is going to go. I'm going to give a very brief background. I imagine a lot of the people who are interested in this kind of topics know quite a lot about AI and the state of AI at the moment. But just in case you don't, I'll give some background and go over that. Explain what nanowire networks are. Explain what we actually do, how we model these systems, and how we actually apply these systems. And then at the end, just give a brief overview of what I've been working on for the last year or so. So let's start with a bit of a background. So the brain, the original inspiration for artificial neural networks, AI, kind of this machine learning idea behind AI, was the neuron. The most basic ANN is called a perceptron. And it was sort of in the 50s and 60s. It was created to mimic the idea of this inputs into a neuron from this sort of dendrites into the, through the cell body, into the axon, and then across as an output into different neurons. And that's sort of the idea that we have. That's sort of what became this beginning of artificial neural networks and AI in general. And so, but there's quite a few key differences between neurons and the structure of neurons in the brain and how we implement our perceptron slash ANN interpretation of these neurons. Mainly, there's the idea behind von Neumann architecture, which is what, how we pretty much implement all of our computing these days, compared to how the brain is structured. So in most modern AI systems, we implement this software of AI on hardware, von Neumann hardware, where we have this separation between computing, processing, and memory. And we have to bus information back and forth from the CPU to the memory, which is quite a big bottleneck. It takes quite a lot of time to do that. And that's pretty much…