Sensory Prediction Error Signals in the Neocortex with Blake Richards

Automatic TRANSCRIPT

Am with Blake. Richards Blake Blake is an assistant professor in the school of Computer Science and the Montreal neurological institute at McGill University as well as a core faculty member at et Mula and you've also got an affiliation with far yes. I'm a candidate C.. For a chair and a member of C far's learning machines and brands program. Fantastic fantastic passing. Well Blake welcome to the PODCAST. Thank you very much for having me. Also you are doing a talk here on sensory prediction error signals in the neo cortex yes Let's just jump right into that. What's the talk about sure? So a lot of people have postulated for a long time that our brains and in particular. The NEO CORTEX the region concerned with higher order thought or functions if you will is effectively. An unsupervised unsupervised learning machine. It is there to make predictions about incoming stimuli and that it would use. WHO's this differences between those predictions and the actual data that receives in order to learn about the structure of the world and develop a good internal model? And although Oh there have been many computational studies that have postulated this and this idea has also informed artificial intelligence a great deal. The fact act is that there isn't a lot of direct evidence for it in the brain. There are a few initial studies but myself and my collaborators Joel's Albert Virga York University Yatra Banjo also at Milan University. Molly AL and Tim Lily. Crap at Google deep mind. We put together a proposal to the the Allen Brain Institute a couple of years ago to run some experiments to explicitly look for some of the sorts of prediction dictionary signals. That these kinds of models of unsupervised learning in the brain predict would be there so the institute has been running a series of studies. Doing doing what's called two photon calcium imaging and mice is basically a way to record the activity of many hundreds of neurons at once as well as their dendritic processes in a live animal. And so we've got recordings of the brains of mice primary visual cortex. Well we expose them to new stimuli that follow particular statistical patterns which we then violate occasionally and we have found evidence for very clear sure and really strong responses to those violations of the expected stimulus and additionally there are some interesting kind of breakdowns in terms of where those signals appear in the critical circuit. And also ask some interesting data in terms of the way that it seems to be something that the animals actually learn over multiple exposures to the stimuli See your conditioning mice to expect some kind of response then you kind of take a left. Turn when they're expecting right so to speak and and you're observing. What's going on in their brains as a result and so that's not what you've found is what so what we've found is we've Zam? I'm in two different types of stimuli. One which is where you've got a consistent visual flow in in the screen as it were so there's these bricks that kind of drift left across the screen in a particular direction and they're always consistent in that movement and then occasionally some of the bricks will start moving in the different direction than the expected one or like pretty clearly. You notice it when you watch the stimuli yourself very much so And for those stimuli. We see a massive response. In a particular part of the neoclassical micro circuit the cells. Go nuts in response to the stimulus and this seems seems to happen right off the bat with no training so that suggests that this particular type of violation of expected stimuli is something that the circuit is hardwired wired to detect but we also have another set of stimuli where we present basically these random patches of edges That are all sampled from where the orientation of the edges are all sampled from a particular distribution and then occasionally violate that expectation by sampling. I'm from a different distribution for the orientations and when we present these stimulated the animal at I don't see any responses to the unexpected head orientations but over multiple recording sessions. We start to see huge responses to the unexpected annotations. So what's interesting about this. This is suggests that the circuit is able to learn as it were to be surprised particular types of stimuli and it might at the same time be a hard coded to respond to particular other types of violations. We hypothesize that this might have to do with the evolutionary purpose. This of most visual cortex one of which would obviously be to help the mouse avoid predators. And so it's very important that you detect violations of visual flow in your visual field if you're trying to avoid predators because that's something you want to avoid potential. You WanNa see that Hawk flying above or right exactly but for the other type Stimuli where that we're showing them where it's these oriented edges. That can violate these patterns. What's interesting is that they don't show that response right away but they learn to show it and so that what is some evidence that they're neo? CORTEX is in fact a sort of generative model that can learn the data distribution over time and learn to he surprised when data doesn't actually adhere to that distribution in the first case where you've got more of a stark difference in the visual pattern. Turn do they become desensitized to it over time. We don't actually see any evidence for desensitization which is interesting The signal continues to be very robust over three different days of recording sessions and each session is an hour long so even after many repute exposures of this stay still seem to signal this very strongly which again suggests that for that particular type of stimulus. This is a hard wired component which is very consistent with the evolutionary right. Like if you got desensitized to hawks probably be a bad thing if you're and so there was another element of this work ERC or at least one that you haven't gone into this level of detail yet but is talking about the hierarchical nature of inference in these kits. Is that kind of an ancillary result or is that core to the model that you've developed to understand the stuff yeah so The thing that I mentioned is that what's is interesting. Is that that second type of surprise signal that we see that the animals learned to be surprised to the orientation of edges in that occur in an expected elected way we actually see that signal not in the neurons themselves but in the dendritic trees of the neurons and in particular part of the dendritic trees that is the area drives being like the fingers that we see in our nerve. Yes that's right exactly. All those little branches is that comes out of the Dan right out of the neurons. Those are done rights and those are the sites of synoptic inputs to real neurons but real oh pyramidal neurons in the neo cortex which is a particular type of neuron comprises seventy five to eighty percent of the neurons in the CORTEX in. It's the kind of key the information Presence Cell type in this circuit these cells have a one unique dendritic. Nick process called the April damned right which they send up a vessel yet April okay and they kind of like the tree because what they do is they send it up to the top the surface the brain almost like what the trunk of a tree does the leaves up to the sunlight but in this case they ended up to the surface of the brain and what they receive at this location occasion are the top down in puts so higher order information from other parts of the brain and our data suggests so that is he's actually where we see those surprise signals that are learned and from some of the brain or in this structure overall in this dendritic structure. That is up at the top of the brain and what that data suggests and some of our other analyses suggest is that this surprise signal. Is this you know. Oh that violated my expectations signal that the animals learn is being driven by top down inputs. So that suggests that the entire model model that they have for the world that they're learning is a hierarchical model where it's actually the higher order parts of the network. If you will for machine machine learning people you can think of it. As the upper layers of the network that are actually detecting the violation of the expected statistics and then they are communicating that back down the hierarchy hierarchy to the lower layers of the

Coming up next