NeuroAI for AI safety by Patrick Mineault | Devcon SEA

[Music] [Music] Patrick Mino and I'm a computational neuroscientist of uh Vision originally uh I was also a data scientist and I've worked on brain computer interfaces and Industry and I'm currently working at the amaran foundation on a brand new ambitious project uh to make AI safety a reality so the am found was founded by James fle who's one of the pillars of the ethereum community and it funds ambitious research in longevity and Neuroscience and uh is now working in AI safety so of course AI has really changed the way that we interact with uh with text we've also seen tremendous advances in the applications of AI and biology of course Alpha fold recently won uh the Nobel Prize uh for physiology uh and we've seen applications also in entertainment uh including these uh uh dogs which uh which are puppies playing in the snow which are figments of sora's imagination uh but of course there's much to worry about when it comes to AI because it does pose a potentially existential TR so powerful AI uh could pose a risk of uh Extinction and so we saw in this recent petition that everyone from the fathers of deep learning to AI safety researchers to CEOs of AI compan companies as well as members of the Dak and eak communities uh really congregate around this idea uh that we should really mitigate that risk and towards that end uh we really want to create uh safer technologies that can replace uh this technology that is potentially very powerful but is also um understood uh to pose significant safety risks now why would we have a path towards safer AI by learning from Neuroscience well current AI systems are really trained on Behavior so if you look at an llm is trained on the hard work of hundreds of thousands of people that have poured in um their work to text and we learned to imitate that but that doesn't contain all of the nuances the processes that were able to create that text we could instead uh learn information directly from the brain and learn to imitate the activity of the brain in order to create potentially safer AI systems so let's break this down a little bit um we know that there are existing neurot Technologies uh which have enabled large scale uh Neuroscience uh things like neuropixels um the neuralink implant device functional ultrasound is coming on board uh things like diffuse Optical tomography uh new advances in in conomics and these might allow brain imitation learning now we know that humans know how to explore safely they're robust whether it's adversarially or out of distribution they cooperate sometimes they know how to reason about other humans our world is built for humans and those capacities are really a function of the inductive biases which are embedded in our brains so if we could back this out uh we could uh potentially better constraint AI systems that we would be able to using Behavior alone so in order to build a path towards brain imitation learning I like to think about three distinct systems sensory motor and cognitive so on the sensory side it's been known for a long time that the representations inside of convolutional neural Nets which are trained for image recognition are remarkably similar to those that we find in visual cortex in particular in a ventral visual stream however they do diverge in important ways in particular when it comes to out of distribution inputs and adversarial inputs so we funded work at UH Stanford University which is called the Enigma project to really build digital twins of the visal cortex so animals are trained to watch movies for several hours and uh using a lot of this data it's possible to stitch together a model that captures the relationship between the input that's coming on the retina and the representations which are computed by the neurons in that brain and therefore distill a more adversar robust out of distribution robust uh digital twin uh that can be used in lie of the current image recognition systems that we use when it comes to motor cortex uh we've seen tremendous advances in virtualizing bodies uh so people use Technologies like uh X-rays and microct and com focal Imaging uh to create basically parts list of uh of animals and then to painstaking work reconstructing their skeletons as well as the their tendons eventually allowing them to virtualize them in environments like moku so between that and the advances that we've have seen in robotics it has become much more possible uh to create a link between the kinds of cognitive representations that we see in the brain uh and eventually the output of the body and its interaction with the environment so finally on the itive side uh there are hundreds of thousands of hours of available data from across all kinds of brain modalities on archives like uh Dandy Hub as well as open neuro which brings about this possibility that was floated recently uh by Andre kathi which is what you would really want to build an AGI is the inner thought monologue of your brain if you had a billion hours of that AGI would be here roughly speaking and that AGI would have learned uh the same kinds of properties which are so crucial for the brain that allows them to be safe and therefore would form a safer kind of AI now if we want to uh have an idea of what this world will look like uh look no further than fly Neuroscience so over the past few years uh people have been able to freeze the brains of flies to chop them up in thin slices scanning them segmenting them reconstructing the tracks inside of their brain s virtualize their bodies and track the entire visual input from uh the first synapse uh down into the central parts of the brain to really understand what's going on inside of the F's brain and that's what we think of when we think of a virtual fly but imagine if we had that for virtual humans that would have the Dual property of really advancing Neuroscience as well as really um creating a new kind of AGI now it's not going to be an upload like would see in The Matrix or in San hun pero but would rather be an amalgam of the activity of many different systems which are wired up uh together but would nevertheless be a crucial artifact to build a safer path towards AGI so to build this we need lots of brain data a zoo of virtual bodies and environments to leverage the existing data uh as well as bringing uh extra data from closed archives and to accelerate C cutting data acquisition Technologies including conomics and wireless electrophysiology so we have a road map that's coming up in uh about two weeks um that's called Nur AI for AI safety it's 100 pages of deep technical Deep dive uh into this field if you want to hear about its release uh please uh follow me on X uh for uh for news on that I'd like to thank my co-authors Nicolo and