Mortal Computation, Cybernetics, AI

hello everyone and welcome back to active inference insights I'm your host darus P way and today I have the pleasure to chat to Alexander orobia Professor orobia is an assistant professor of computer science at Rochester Institute of Technology where he directs the neural adaptive Computing laboratory his work focuses on developing new learning procedures and computational architectures that embody various properties biological neurocircuitry and are Guided by theories of mind and brain functionality last year he published the paper mortal computation a foundation for biometic intelligence with Carl friston which grounded the Mortal computation thesis in the active inference framework in which they argued that active inference might prove useful in guiding the construction of computational systems Alex thank you so much for joining me it's been a little while since we've done one of these so I'm going to be a little bit Rusty um as exemplified by my my pronunciation of the word labor Tre uh lab I normally just go for lab um how are you I'm doing great thank you for having me on and uh no you're very welcome you can call it laboratory it sounds nice and formal I like it yeah laboratory lab um great well before we dive into that paper I wanted to ask you're kind of coming from a computational background um I've never actually asked this question first up to a guest but it struck me that I probably should which is how did you come across active inference as a as a framework oh that's that that's a nice question um well it was a few years ago and the short compressed version of that is that I was working with some uh colleagues in the Imaging science department here at RIT they do eye tracking and uh cognitive science research I also have a cognitive Science Background or part partial background in my Gra grad school years and uh so as we were uh talking about building computational models for uh ey tracking problems uh we we ended up settling on this reinforcement learning Dynamic control formulation of the particular problem we were solving and uh I had been long thinking about biological process models for reinforcement learning control and so I was already aware for a very long time of the free energy principes so I because I I'm a predictive coding researcher I like to also pitch that as something I have quite a body of work uh build been building a body of work on and so active inference is like an extension of that right it's basically say let me put action into this process and actively sample the environment uh to facilitate my self- evidencing and so uh it just kind of made sense that we would maybe look at the free energy principle originally wanted to build predictive coding models of our reinforcement learning system and then we said well wait a minute uh exploration is going to be a difficult component here and that's what active inference brings one of the big things it brings to the table uh to use Carl's phrase uh which is this epistemic foraging I the idea that we're doing intelligent exploration driven by our surprisal and so it just kind of made sense to investigate this framework in a in a in a further further away uh and then since then it what was nice about that group is that uh my my colleagues uh weren't very familiar with uh predictive coding so I gave them a talk and then we wanted to understand active inference so we had a little reading group that we established with the grad students the other faculty and uh you know we had forced ourselves to dive a bit deeper into for example Carl's work um Chris Buckley's work as well so I was already familiar with Chris even before I actually got to meet him years later um and then one of our one of my primary grad students uh George he was uh actually building some active inference models uh and then we effectively settled on a version of that in our it tracking problem and actually that led to a really nice article about last year with George in h frontiers of…