Return to audio interview | Next episode | Explore UC Science Radio
Molly Magid: Welcome to UC Science Radio, where we interview a range of postgrad students to tune into the fresh voices entering the world of science and learn what sparked their passion. I’m Molly Magid, a Master’s student in the School of Biological Sciences.
Today I’m talking with Josh Mallinson. He's a PhD student studying physics. His research focuses on uses concepts from neuroscience to inform computation, like using artificial neural networks that mimics how our brains complete tasks. This research informs the creation of neuromorphic computer chips.
Kia ora Josh, welcome to UC Science Radio
Josh Mallinson: Thanks very much, it's great to be here.
So you're studying physics but specifically what is your research about?
JM: So our research is about neuromorphic computing and specifically using percolating nanoparticle networks. So a percolating network you can think about if you have a footpath and you draw a line across the footpath and then a meter away from that you draw another line across the footpath. And then if you imagine that it begins to rain in between the lines that you've drawn. Over time, the raindrops will overlap and they'll form connected regions on the footpath. And there's a certain point at which one raindrop lands that will create a complete pathway from one line to the other. So at that point, you have a percolating networks of raindrops on the surface.
And you're using concepts from how our brains work to inform this process, so what exactly are you doing in terms of looking at the brain and then applying it to computing?
JM: If we go back to the percolating networks, for a moment, so if we take this pitch that I've described with raindrops on a pathway, and you shrink it down ten thousand times, and we replace the path with smooth silicon nitride surface and the lines become the edges of gold electrodes and the raindrops are conducting nanoparticles, so they’re metal nanoparticles that have landed on the surface, when we apply a voltage across the two electrodes, then current can flow through those conducting nanoparticles in this kind of complex network.
And so there's many pathways which the current can flow through that network and there are some pathways that just aren't quite completed, so there's gaps. So in those gaps is where the interesting physics takes place. So the electric field can induce the formation of an atomic scale conducting filament, which allows the network topography to change. And when that happens, the voltage across that network is redistributed and that causes more of these filaments to form. And the filaments can also break which has a similar effect—it leads to the breaking and formation of other filaments in the network. And so this is very similar to the way that the neurons in the brain fire, when one neuron fires and it triggers subsequent neurons to fire, and so you get this cascade of events. The way the brain operates is by creating these cascades of neuronal firing events. And we showed in a paper published in Science Advances last year that the statistical distributions of the switching events- the formation and destruction of these atomic scale filaments in our network closely matches the behaviour of the brain in terms of neuronal firing.
So that's sort of how it relates back to the neuroscience, that's why the neuroscience in important to us. So we want to use these chips that have these brain-like dynamics to try and create hardware versions of artificial neural networks which is the goal of neuromorphic computing: to make physical systems that can process information the way the brain does.
And what sort of information processing would be improved by using this neuromorphic hardware?
JM: Yeah, sure, so that's a great question. So conventional computers were designed to perform logic operations. They're very good at doing mathematical calculations. And they were designed that way because humans aren't good at doing that sort of thing, right? So a computer can do maths much, much more quickly than any human.
And so that's why they were designed and now modern society wants more from computers. We want computers that are intelligent, that can make decisions, that can recognize patterns and predict what's about to happen. And so we basically want our computers to do things that humans are good at now. We’ve made computers more intelligent over the years by scaling. So we have computers that can do sequential logic operations so we're gonna scale them, we're gonna make them able to do more of these operations per unit time. And if you keep stacking, stacking, and stacking, making these things better at doing the same thing, you know, more quickly and in higher density, then you can create the intelligent computers we have today.
But the problem with this approach is that it uses a ridiculously large amount of energy compared to the biological counterpart. And so the most advanced artificial intelligent systems use megawatts of energy. Whereas the biological counterpart, the human brain, only needs a sandwich and a glass of water to run. And it does a whole lot more than even the best artificial intelligence systems. So that's the goal, that's why we want to take the inspiration from nature's approach to information processing, and try and replicate that.
And going off what you said about, needing less energy to compute, you and your lab developed a chip that could be used in computers that would reduce the amount of greenhouse gas emissions that come from computing. Can you talk a bit about that?
JM: Globally, computing hardware uses a significant portion of the energy that humans generate. So I think somewhere on the order of 8% or something like that, which is a significant amount, and so, you know, you think about the ways in which that energy is generated often involves using fossil fuels and other methods of generating power which is harmful to the environment. So the more we can reduce the amount of power needed by computing hardware in the world, the more we can reduce those negative effects of generating that power. And so, you know, for the amount of computation that a human does, their greenhouse gas emissions that are a direct result from that computation is very low by comparison to conventional computing hardware.
And what was your path to this research?
JM: I didn't decide- "I really want to work on neuromorphic computing." It's one of those things where it was an opportunity that presented itself to me. So I did my undergrad here at UC, and I finished that in 2012. And then I actually went away and worked in the construction industry for four years. And I always intended to come back, and so when I did that in 2016, I just went to the professor I remembered having the most enjoyable classes with, which was Simon Brown. And he said these are the projects I'm working on, and I said that sounds really interesting, and he said yeah, cool. So I started a master's and eventually was upgraded to a PhD which I've just finished. And so it sort of, it just flowed. It just came about.
What was the most fun or exciting part of your work?
JM: I think it's where you see some strange behaviour in the data you've collected and you see some strange observation, and you get a suspicion about what could cause that. And it might not be anything that you have thought about in the past and then you think of a way to test that theory, and then when you're right, the payoff is one of the most rewarding experiences that I've ever had. I've been fortunate enough to have that happen on a fairly regular basis throughout my research, so it's about theorising and then understanding and being able to show that you're right, I think is the most enjoyable aspect for me.
So my last question is if you could see one big change come out of your research, what would that be?
JM: I would like my research to contribute towards humanity's understanding of intelligence and we don't really know how the brain works, right? And so we're trying to make things that can behave like the brain, and I think that is a pathway towards understanding how the brain works. It's sort of the fundamental question of human, of: what am I? How can I be thinking about myself? And I find that fascinating. So any progress towards understanding the nature of consciousness and computation in the brain would be a good outcome.
Thank you so much for talking with me!
JM: No problem.
Return to audio interview | Next episode | Explore UC Science Radio
