Biological Neurons Compete with Deep Reinforcement Learning in Sample Efficiency in a Simulated Gameworld

Hello and welcome. This is Active Inference Guest Stream number 86.1. Biological neurons compete with deep reinforcement learning in sample efficiency in a simulated game world with Faro Habib-Balai and Moeen Khashnijad. And thank you both for joining. Really looking forward to the presentation and take it from here. Thank you very much. Yeah thanks Daniele. I guess I'll just open up. Hello everyone. Thanks for joining. I hope this will be interesting and fun for everyone. So today we are going to talk about a collaboration between Cortical Labs here in Melbourne and Monash University. We will mostly emphasize on the details of the project as titled Biological Neurons Compute with Deep Reinforcement Learning in sample efficiency in a simulated game world. But before going into details of the findings that we had in this study I'm just gonna talk briefly to our setup, our system of in vitro neuronal cultures that we are interacting with and we are embedding them in simulated game environments to be able to do all these cool tests and which has led us to these findings. So to start with I'm gonna just tell you what were our motivations in Cortical Labs and what's led us to where we are here today. Without getting too philosophical though we can say that we as humans need we are known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be known to be Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. So, yeah, without getting too philosophical here, we can say that us as humans, we need intelligence in some form to be able to distill meaning from the phenomenon around us. To be able to make more intelligent decisions by computing more data faster and with better outcomes has actually become one of the major goals and the holy grounds of machine learning and AI in the past decades. Now, in search for different ways of intelligent computing, one way for us to categorize all of our recent efforts could be to talk about these categories like machine learning and AI, quantum computing or neuromorphic chips. Obviously, there are other ways of looking at these and other tools that we've developed, and they are all the right tools for the right purposes. But they could all also have their own pros and cons. And what motivated us to look for new alternative ways of computing was that, for example, in machine learning and AI, we are facing challenges such as their high computational power, high power use, their catastrophic forgetfulness, or the fact that they require very large samples and very long training times. And they are also very susceptible to errors. Or if you're talking about quantum computing, again, we have extremely high power usage. These are very isolated from the external world and very prone to errors, and they are not very adaptable. And also, if you're looking at neuromorphic computing and the whole field around it, there's also limited error correction capabilities. We have a lot of variability between circuits and also, they still require some sort of programming to some capacity to be able to do what we are designing them to do. So basically, in search for this alternative intelligent computing system…