 Good afternoon everyone. Hello, welcome. My name is Dimitri Peroulis and it's my privilege to welcome you to this event tonight. This is the Purdue Engineering Distinguished Lecture Series. This is a College of Engineering series that it was introduced in 2018 and it invites world-renowned faculty and professionals from all around the world from industry and academia to debate big issues, big challenges and big opportunities in our field and beyond. We typically start by presenting a lecture by our distinguished speaker and we will follow this with a panel of experts that will debate a lot of different and interesting ideas that you're going to hear today. So to introduce this event is my pleasure to introduce Dr. Mang Chang, the John A. Edwardson Dean of the College of Engineering. Good afternoon everyone. I realize that we didn't build a big enough atrium in the new Armstrong Hall of Engineering here. Not for events like today's, not for speaker like today's. Dr. Jim DeCarlo from MIT who's a chair professor there and also the head of the Department of Brain and Cognitive Science and our principal investigator of multiple institutes over there related to brain and neuroscience interacting with other views. There are very few topics that are as exciting and unique and curious as today's and very few speakers in the world with as much curiosity to share with us here today than Dr. DeCarlo. And this reminds me in he and his team's quest to reverse engineer how we see, how we think, reverse engineer bring and reverse engineer mind reminds me of this saying that supposedly Bertrand Russell said when asked what is mind said no matter and when asked what is matter then it says never mind. The matter mind duality and interaction is one that to puzzled all human beings for millennia and the interaction between engineering and neuroscience. The interaction also between artificial intelligence and natural human intelligence is absolutely at the forefront and the crux of many of our endeavors. So it is such a distinct pleasure to be able to steal several days of time from Jim's busy schedule and to grace us with his fantastic ideas and discovery here today in Purdue College of Engineering. So a big round of applause to welcome the distinguished lecture today. Thank you. Okay. Thank you for that wonderful introduction. Thank you for all of you for showing up. I'm sorry the standing in the back here. I'll try to be entertaining enough that you won't mind that have standing the whole time. I would like to sort of start by saying I'm actually a midwesterner. I grew up in Ohio. I actually spent summer here in Purdue for learning how to code and do some microbiology. So it's sort of a real pleasure to be back here to be feels like home to me. So again thank you for welcoming me to Purdue. I'm going to try to give you today a spirit of what we're doing which we refer to as reverse engineering the neural mechanisms of human visual intelligence which is really a combination of forward engineering under constraints of brain measurements. So here's the here's the sort of spirit of what we're what we're doing here. Reverse engineering is again the goal is to somehow account for every ability of the mind which which we mean by intelligent behavior. Somehow using only the connected components of the brain. So I guess this is the mind matter divide that was just referred to. For us this means connected neurons. Neuroscientists think the brain is a big recurrent neural network of some type and our job is to figure out which one is actually running inside our heads and to do this in the language of engineering which means that we're not just saying it's a neural network but which neural network and it's able to predict things that enable us to do applications with that knowledge. So these are classically the domains of mind and brain sort of measurements and discoveries from science about intelligent behavior and on the other side engineering about how you might even engineer systems whether it's software hardware robotics. So these are really fields of science and engineering trying to come together around understanding or building intelligent algorithms which again when you're thinking about neural network components components of the mind you're going to end up playing in a space of neural networks which I'll refer to as neural network algorithms. So this is quite exciting right now because both engineering and neuroscience are really working the same class of hypotheses like how do we find better neural networks to do things like artificial intelligence and natural intelligence. So again I'm going to refer to these as the family of deep recurrent ANN models. So the research that I'm going to tell you about today from my lab is really just one foundational component of visual intelligence. It's not all of thinking. It's not certainly not even all of vision. It's just one component but I hope it gives you a sense of this way that you might approach such problems with this a blended idea of going between science and engineering to come up with actual solutions. I have to sort of start with all my talks by acknowledging the folks who do the work. Everything that I'm saying here is really done by the hardworking students and technicians and postdocs in the lab many of which are highlighted here. I will try to highlight them along the way and I sort of with that a nod to all the students and other researchers in the audience. I know that you are the ones that are on the ground doing that just as these folks are in my lab. So let's start back, step back to kind of what is vision and the problem of visual intelligence. So in my start in my lab in 2002 at MIT I thought can we kind of find a way to reverse engineer human visual intelligence. Now what does that mean? It means you look at a scene like this and there's a lot that you seem to be able to intelligently do. You might be able to say what is in this scene? What are the objects in this scene? Where are the cars? Not just what is there but where are those things? Where are the people? What will happen next? Where is it safe to walk? And many other questions like this. These are all questions that I'll put under the umbrella of visual intelligence. We've not answered them all. We don't know yet how the brain does all of these but we've made a lot of progress on the ones on the left. Things like what is in the scene and where are the things in the scene. And I'm going to tell that progress today. But again you should think of this as the beginning of the story not the end of the story. So just as you should know for those of you who aren't neuroscientists or cognitive scientists you do not see that whole scene at once. What you really see is where you fixate where that red dot is shown you really have high acuity on your retina. So your front end sensors on your eyes are really good at looking at detail really in the center let's say 10 degrees or so. 10 degrees is about two hands at arms length try to illustrate that here so you can get a sense of you're not really understanding that whole scene but you get a sort of phobiation that might look like that. And the way you digest the whole scene is you make rapid eye movements around the scene. Those are called saccades and you dwell at these fixation points that you sample for a couple hundred milliseconds each. You don't even notice that you did that when I put up that image but that trust me that's what your motor system and your brain are doing together at this moment when you look at this scene. So why am I telling you this? It means that we can sort of reduce it a little bit and make it a bit more tractable in the lab. So here's now snapshots taken from that scene just so you can get a sense of what is confronting your brain at the central part of your retina every couple hundred milliseconds and I'm going to show you that here so just try to watch that screen here and see what you can kind of be able to detect. Okay I'm going to do that one more time. I hope that you can even though these are now cut out snapshots for just a couple hundred milliseconds again about the duration of your eyes dwell that you couldn't be able to notice that you could be able to say things like was a car in that image was a person in that image is there a sign or a tree a car or a sign you can make these distinctions very quickly even though you couldn't quite say them I hope that you could perceive them and in fact we can measure that and you are actually quite good at that task and it makes sense because this is the way that you digest a scene like the one I showed you this ability to do quickly be able to say what's there and other things about it we call core object recognition because we think it's a foundational component of visual intelligence and again that's what we've been studying in our lab for the last 15 years or so and that's what I'm going to tell you about today. So now just make this a little more concrete if I was going to test you in the lab I might put you in front of a screen like this I'd say look at that dot and an image would come up and then you'd have to maybe make a choice was it a car or a bird what do you guys think car okay so again I didn't even cue you what it would come but we can sort of try this a little more I have a few more just to give you a sense we give a lot of different kinds of both naturalistic photographs and genetic things they're challenging for computer vision systems that we test humans and primates on as well so here's sort of an idea of how it goes here's some more let's see there's I didn't even see that one was hopefully you guys got it very fast here's another one okay that's pretty easy bird I think okay and then natural image okay hopefully you notice that was a bird okay your ability to do this this is just how we quantify that that was you all the primates in the room here including myself here's another primate this is a rhesus monkey in the home cage doing this and the rhesus monkeys love to do this all day long they get nice juice rewards for doing this this this rhesus monkey is doing what you just did he's triggering the central screen to present an image and then he's choosing among two choices and I hope you can notice there's many objects coming up he actually knows about 30 objects they learn about one new object a day if you teach them so they can discriminate for instance cars from dogs and so on and so forth and and now we got that one wrong and that was a timeout but usually he's doing quite well here's the behavioral data from primate homo sapiens and primate rhesus monkeys what I want you to see on this slide is basically how similar these are in other words I mean like it or not we're really no better than rhesus monkeys and our ability to discriminate in that situation okay that may sound depressing but that's also an opportunity because this means that it makes sense that's that we're evolved from these line of hominids and this this this makes it quite possible that we can now study this these abilities in the rhesus monkey that can get direct insight as to what's going on in our own brains so just you should also notice these patterns of errors which is what I'm showing you here it's not as if we're perfect at all these tasks some things are harder to discriminate than others and that's what's shown in those different kind of colors there just to give you a gist of what's on the spot but the big takeaway is humans equal monkeys with regard to these kinds of tasks okay now that's for one task called core visual recognition and we showed in 2010 that those primates were both way better than 2010 computer vision system so there's some magic some secret sauce if you will in those primate brains that we think is shared among both those brains but is better than what was going on a 2010 computer vision system so that made it a very interesting place to study to say what's what's the special stuff going on there that isn't yet in those computer vision systems so the reason we're looking at the primates is not just because we think monkeys are cool but because we can gain more direct access to what's going on in their visual system and we know a lot about the visual system of non-human primates much more than we know about human brains this so this now we study the primate brain which I'm showing you here and for those of you who aren't familiar with the visual system it's mostly in the back of the head those colored areas of the different parts of the cortex the outer part of the brain that processes visual input you see them there's a V4 and IT and this is called referred to collectively as a network called the ventral visual stream it has both feed forward and feedback connectivity I'll show you that in a minute it ends in a kind of high area called infertemporal cortex or IT lesions and IT result in deficits in these kind of tasks so based on decades of work we already knew in the non-human primate where the kind of main circuitry lived we just didn't know the algorithms that were being executed within that circuitry just end in IT cortex and then we all go home of course IT has to connect to other areas involved in decision and action areas involved in memory and value judgments and those connections of course exist roughly in the areas that I show you here up to frontal and medial temporal lobe structure but we're going to focus on the ventral stream today here's the ventral stream you know as an engineer I like to kind of draw these things not as tangled up cortex which is very kind of complicated looking but think of these unrolled sheets of neurons in each area sort of schematized here with these sort of hundreds of dots or so and you see again both feed forward feedback and like recurrent circuits that are just schematically shown there and so the conceptual way we think about the ventral stream is that an image comes in in the front it of course is captured on the nice camera on the front of the system called the retina the retina transduces the light into patterns of spiking activity at the back of your eye at the center of your brain and so there's a basically that first image is a nicely processed photograph so it's still basically a copy of what's out there in the world but then it's transmitted through to produce successively new neural representations which I'm indicating here with these kind of colored flashing lights meaning that's an active neuron versus a non-active neuron just to simplify it and make it conceptual for you ideas that there's this progressive changes in the neural code from what's on the retina to what's at the end it takes about a hundred milliseconds to go up this chain of processing as I showed you here when you show a new image you get a new pattern of activity up in IT cortex those red dots and if I go back to the old image you get the old pattern back again and you know you can study this all day long which we do in the lab where you can kind of measure the response patterns even when you show images such as this your IT or your homologue of IT will nicely follow with again a lag of about a hundred to a hundred fifty milliseconds it can easily follow at these kind of transmission rates shown here and that shouldn't be surprising to you because I told you at the beginning that you can see vision in a glimpse and you all did it in just a couple hundred milliseconds and that movie was even faster than that yet you can still notice you can recognize objects in that rapid visuals presentation movie. Okay so this is the sort of lay of the land of the ventral stream sort of rough feed forward flow and I'm giving you the backbone of the ventral stream and overview but we know for some time and this is work from 2005 that if you go and record from these IT neurons so now you place microelectrodes you go and record from individual cells the animal is fixating the screen and you show a bunch of images some of which are shown at the top you measure the spikes out of those individual neurons so you can think of this as the transistors turning on and off if you like but just the neurons fire in spikes that are the same as the other ones and you can see that in the image you can see that each is showing a little dot a little blue tick mark here and each of those rows is a different presentation of the image and the different images you see produce different patterns of spikes so I would like you to notice on this plot this is just so you can get your feel for the fundamental data that we collect in this species and notice that each this is a two we see the spike patterns are different across site one and site two and here's site three so it tends to like that third and fourth image for some reason and whereas site one likes the first and second images more just to give you a feel for what's going on also notice that it's not exactly the same on every trial and I'm not going to talk about that today but well just so you can get a feel for what's going on let me then tell you about how we actually mostly analyze this it cortex and we then go ahead and sort of average over and say what's the mean response to each of these images so we can sort of average the number of spikes you can think of counting the spikes in those red windows that gives what's called a spike rate this is you can see this example here these are four spike rates we can now take this as a sort of more compressed measure of how information is being coded these are called rate codes and now we can ask what these rate codes have interesting properties and I'm going to tell you about next I want to say we don't just record one unit at a time but we typically recording hundreds of sites with chronically implanted electrodes that are put in during a surgery which are then under human aseptic surgery and then animal can just be connected to a connector each day to record from these neural activations okay so these and we can record for many months from the same neural sites so now here you are recording an IT cortex of those neurons we're taking a sample of a population of millions or about a thousand or so neurons and now I'm showing you in the color the green I'm showing you whether a neuron responds more or less to a particular image so this show this is called a this is a population response vector out of IT cortex and these neural codes if you will have very very special properties so we measure just one but here's with response to eight images in fact we measure to thousands of images so we get data sets that look like this very high volumes of data and one of the things that we showed about these codes and continue to show is again they're very special in the sense that you can put just linear classifiers basically sums with thresholds on top of each on top of these population codes and be able to produce the behavioral accuracy of the organism on these kind of tasks that I showed you earlier so you replicate that pattern of behavioral performance that I showed you in that color plot and its performance level in other words those codes were far more powerful than 2010 computer vision systems they essentially contain the secret sauce of the solution that the brain is computing to these object challenges that computer vision systems weren't capturing at the time a summary of that all that work really decades of work is that the brain solution to these core challenges is really conveyed in that IT neural population code that's the summary of what I've just told you and the primates are better than computer vision systems because they had this code okay for neuroscientists this has important details about how this we could experimentally manipulate this code and cause perceptions you want to think about like the matrix those are things that we could talk about that I won't talk about today but just for today I want to kind of focus on the question of we have these very powerful codes in IT but the key question for us was even though you have these codes you haven't told me how to compute them I've just shown you where we can measure them how do you compute them from the image that is what are the algorithms that intervene between the pixels and the actual codes was the holy grail question you could also sort of phrase this as what are the intermediate codes produced along the way in these other areas that I briefly alluded to up to IT cortex okay so this was the setup and this is mostly one I want to spend the rest of the talk telling you about is that we had a breakthrough on this understanding in around 2012 and I want to give you a sense of that right now so in 2012 the responses of these neurons were a complete mystery I showed you some of these responses I showed you some example images they were a complete mystery in that if you kind of plotted them here's a bunch here's one neuron's response to 1600 test images grouped by the category that generated the images you see four different images here there were a mystery in that you couldn't just look at this and say oh it's responding because there's a chair or there's a line or there's some feature in the image that you could easily name they were complicated in some way that we weren't able to describe there's those images blown up here's another classic famous neuron that responds more on averages to faces than to non-faces but it doesn't respond to all images of faces and it also responds good to some images of images that are not faces so they were complicated responses over the image but we knew as a group they were very powerful as I said so now how do we make progress and this is where the breakthrough comes in science was sort of telling us kind of what was going on in this system we already knew a lot of background in the monkey that I sort of referred to direct connection to today's AI for vision and more broadly so neuroscience knew things like each area processing has these kind of spatially local filters these little so called Gabor patches that are kind of applied to the image applied to the inputs is a good model of its processing that that process isn't just at that one spot that I'm showing there in red but is implied across different spots equally across this image so it's repeated which can be modeled as a convolution over the image and it isn't just one oriented filter there's a whole bank of oriented filters like the ones I'm showing you here and this blue one would also be applied across the image in different spots so these were kind of some of the things that were known we also knew things like beyond that first pass linear operator there was these threshold nonlinearities that I'm showing you on the right of that plot and there were normalization circuits in place as well these were things that we knew at the time we also knew that there's this kind of deep stack of areas in fact I introduced the whole ventral stream by showing it to you as a deep stack of areas this is textbook for neuroscience and now you see also that we thought that v2 is operating on v1 we knew anatomically that it was also doing probably the same operation just repeated over space and I'm showing in black because we didn't really know what that operation was although we had ideas but so this is sort of the framework on which people were building kind of starting to build models of what was going on so one thing I we also knew is that there's this very fast feed-forward sweep that I sort of implied that you quickly can propagate spikes up this way and that you have these distributed codes that I showed you were quite powerful so this is just a sort of summary of some of the things that are very important that we knew and as I mentioned this sort of turned into efforts to build models so as far back as 1980s people were building kind of brain style vision models like this one from Fukushima which tried you can see almost looks the same as the drawing I've shown you below and then my colleague at MIT Tommy Poggio built a class of models called HMAX that again you can see has a similar structure they were trying to sort of model what was going on in this system and solve kind of object recognition challenges and then some folks in my group started searching a larger family of models here in this example from Dave Cox and Nicholas Pinto but the thing is this all felt a bit incremental we knew the model should look like that but the key problem was that even with all those things I've given you on the left that's not enough to tell an engineer what you should build into all the parameters of these models you can imagine many of the filter types that I showed you in V1 as I said those weren't really known in the higher levels there's a lot of unknown parameters that aren't on my list here that we're not determined by any of the science experiments and so here's where the breakthrough kind of happened and this was worked by Dan Yamans a postdoc in the lab at the time who's now a persistent professor at Stanford and Ha-Hung a graduate student we had a breakthrough with a model that we called HMO it's an artificial neural network just like the other models I showed you but we made we did a sort of trick that allowed us to make the breakthrough and I want to tell you about that here so this ANN model now here's a series of kind of simulated neurons that you can see in the lower right there's neurons that are meant to have like these linear operators followed by some non-linearity with some normalization just like the brain had been telling us they're applied over the whole image so that convolution idea and that there's a stack of different areas calling V1, V2, V4 and IT and you see it's all feed forward in this model here at the moment so that's an example of an ANN and then we sort of took this ANN and that ANN guided the architecture to said this is the kind of models that you want to be building and here's the trick we said let's not just try to build that bottom up from neuroscience let's try to get that model to do something and what we thought we'd get it to do is do core recognition the thing I showed you earlier was to recognize objects across transformations and things like position scale and pose and then the real power came when we just took the tools of engineering and said let's optimize the parameters of this network to try to solve this problem and then what we were able to do is this is like evolving networks with inside a computer so we're evolving trying to find a good visual system that solves this task so you see this blend between science and engineering and our question was could it evolve to be a brain so we're evolving this artificial neural network under some guidance from neuroscience with a cognitive science task with a lot of engineering tricks to try to get it to be like a brain so this gives us a specific artificial neural network now we've got an actual model and algorithm and then we could compare that because it's a neural network neuron by neuron with the actual brain data that we were recording so that's what that then allowed us to ask questions like for each artificial model neuron the neuron within the brain and it's the best match and ask how well they match on new data so being able to predict the neurons response to held out images and that's kind of what I've just said here and the upshot is this worked remarkably well so we had built this network that was trying to do vision roughly in a bioconstrained space and now we asked it to say look do you have an IT neuron that looks like this one I showed you earlier that I said was mysterious and we searched through the model and we say look here's the prediction of a neural response out of this model these are predictions on new images and they're not fits to the data and you see how well it predicts a lot of the detail responses to these individual images that before were hard to kind of capture in simple models or human words and you see how good the match between the red and the black line was and this was again this was the real breakthrough for us suddenly we had models that were in the right function space and started to look like the high level functions that we were measuring at the neural level in the brain and this not just for this neuron but here's that face neuron you can see it matches in great detail the output of that neuron across a range of images if you look at all those images which are shown in the horizontal there across that space okay so this model then this was this is the breakthrough that I just described now we have this model that mostly explains these responses the big picture here and I'll show you how we're using this model next but the big picture of this model is that we were what we had done was we had kind of looked at the model families and that's what shown in these blue dots and you see that there's a correlation between models that are good on core recognition tasks and models that how well the neurons in the model match the IT level how well the internals of the models look like the internals of the brain that's what this correlation shows on this plot and you could think of one as a neuroscience goal and the other as an engineering goal so science essentially inspired the family of models but then engineering optimized the parameters to do well on these core recognition tasks on large numbers of images and that led us to a higher performing model which then when we compared it with the brain in around 2012 was explaining the brain data much better than any previous model which is what's shown in this plot so there's a virtuous cycle going on here between neuroscience and engineering to both get better performing engineering models but also to better explain the brain and that is what I think is a quite exciting loop that can be carried into other parts of our field now a question for discussion is can brain scientists just wait for engineers to build more accurate models and hope they're going to be perfect copies of the brain I hope you see the answer to that is probably no and I have things to say about that but so you have to kind of keep this play going both directions so look what's cool about this is we can now just say look we've got a pretty good match up in IT about 50% we can start to compare all levels before we get a pretty good match up at behavior we get an even better match and at low levels as well other labs have shown quite good match in V1 and of course this is ongoing work as to how well these models match but a key message for the engineers in the room it's these models were matched mostly because we got them to perform well they're running standard gradient descent for the aficionados in the room currently is what deep models of vision do but these initial models weren't actually running SGD so deep learning if you will led to high performance but this doesn't necessarily mean that the brain is doing deep learning to get to these models just want to make the connection for those of you playing with deep learning models today it may be but this data do not kind of imply that that is necessarily true the big picture that I hope that I kind of convey to all of you on what I've sort of said here about this story is that there's this interplay between science and engineering and this is playing out not just in vision but in a bunch of other fields of neuroscience and I list some of the key folks here that are working in other areas with this kind of similar approach now and there's a review if you'd like to read in 2016 if you want to read more about that so let me sort of let's see so let me just I want to sort of check time here let me see so okay so 15 minutes good okay now I want to show to show you that was really to me by way of background or just to bring those you don't know the kind of history of where vision is gone in these things I want to now took a little more forward as to what's happening right now and what's likely to happen next in this area hopefully to inspire some of you and think about your own work so what I told you so far I would call a glass half full half empty story it's half full that we have these kind of an ends that are now the leading models of the primary visual system they predict the internal neural responses quite well they predict the behavioral patterns quite well I didn't show you that but they do and these brain like machines are also among the highest performing computer vision systems available today for those of you who haven't followed the deep learning and vision literature no deep and it passes all our tests by the way and I'm not gonna I won't be able to show you all that today but that's the half empty part so again think of it as like we kind of made some progress but there's still work to do even in these vision problems so what I've kind of shown you so far is this sort of this inner this kind of intersection between science and engineering of course these networks as I sort of referred to have fed back into engineering in all kinds of disciplines I mean it was around this time that deep learning essentially has taken over all of what's called AI at the moment of course AI is much broader than deep learning but deep learning is now being applied in those architectures of those styles are being applied in many problems and the rules of gradient descent are being used on a lot of problems as well and I'm not going to give that talk today but I'm sure many of you in this room are familiar with that progress what I'd like to do is turn to the other side and say once you have these models you have this feedback from the models to actually even deeper things you can do on the brain than we're able to do before we had the models so that's this other arrow back to brain science if you will from these neural network algorithms so a key question in our field is look we built these models you can call them sort of simulations of the visual system does this count as understanding should this count as understanding it's sort of a simulation of the system that's a debate in our field and I'm not going to settle it for you today instead what we did as engineers we said let's just be practical if this really should be understanding it should allow us to do something what can scientists now do with this model understanding of the visual system that we couldn't do before and I would sort of stick in here that really the major goal of the human species is somehow to do some sort of control of the world now I don't mean that in a negative way I mean it in a way that you have once you have models you can do things like design drugs to better fix a brain or to land something on a planet these are things that you do with models that sort of our evidence of your understanding sort of control of what and prediction of the world so this paper we published from our lab this year from these talented postdocs in my lab neural population control via deep image synthesis that I want to tell you briefly about here to give you a sense of how we've turned these models around and sort of let's go back to the brain and do something with them so once you have a model of the visual system you can now you can take this predictive model and you can sort of say what if I want to put the brain in a particular state if I want to turn the neurons on in a particular way I'm going to call this state so I have a model of the visual system now I want to sort of go use that to turn the neuron to a particular state well we can now use the model to automatically design what we'll call controller images patterns of pixels if you will that when shown to the eyes are predicted by the model to put the neurons into that state to sort of activate the neurons in a particular way remember the model is the one making this predictions without the model there's no way that we could do that so what we did in this study was we actually work how well can we actually use these models of the visual system to turn them around and sort of control the neurons in a way that we determine so you can think of this as control using luminous power on the eyes to then influence the brain in a predictable way and so let me show you sort of progress on that so this is an example neural recording this is an area v4 which is the input to the IT area I talked about at the beginning what you see on this plot is measured predicted firing rate on the y-axis so the model is predicting the x and the measured is on the y and each of those dots is a particular natural image and you see one of the better images at the top better meaning it drives the neuron really well and you can see there's a zoom in on the kind of portion of the image that is most on the neuron so-called receptive field the part of the visual field that it's most responsive to and again you could look at that and say it looks like an airplane wing maybe there neuron likes the airplane wing in terms of descriptor but it's at least kind of just showing you what the neuron likes now notice how good that line is how those dots are lining up kind of well on that line that means the model is really good these are not fits to the data these are predictions from held on held out images from a model that's been mapped to a particular neuron using other images that I'm not showing you here so that fit is essentially the story I told you a minute ago that we have models of this visual system then you can see that quite well again this is back to like 2013 that was pretty well but now the cool part is you're trying to use the model to say can you now drive this neuron out to some place you haven't seen it go before can you sort of predict an image that will super activate this neuron in this case this is one example neuron so goal one target state of the neuron turn this neuron on to a very high state higher than we'd observed over the natural distribution of images shown in blue and so now remember we invert the model we ask it please find me a model a pixel pattern that will activate here's the examples of five pixel five neurons being optimized for different five different neurons trying to find some images that will act super activate those neurons they're fun to look at they look a little bit different from each other and you know this is just gives you a sense that you kind of what the model thinks the neurons want in area v4 in this case now here we go back and close the loop and sort of test those images on the same neuron while we're still recording from it and here's now five synthetic images to this from this model for this neuron so again this is the model saying I predicted that these images will drive this neuron very high because that's what you asked me to do and I've now given you these images notice they look similar to each other but not identical and here's the result from this neuron you see that the red responses the model predicts the neuron should be very high on the X axis the neuron does respond higher than anything we'd observe it respond to before so I we call that a victory in some sense the model wow the models actually move the thing somewhere that we weren't able to get it to go before and I say mostly because the engineer in me says look those dots are not on the line which means there's something still not quite right about the model so again glass half full half empty or already able to do things that we couldn't do before in this one example neuron let me now show you kind of another version of this control idea at a scale that I think is even more impressive to me and is going to be more important going forward so rather than turning on one neuron in the brain what if we set a whole set of them and said let's try to control them as a group so now it becomes even more important that the model you know this is hard to do from a human point of view but the model can try to do it for us so try to drive one target neuron high while drawing all other neurons low was one of the goals we had so here's this kind of before recorded population say we're recording from 40 neural sites right here really 38 in this point and we say look I what I want and it's just something we pulled out of pulled out of thin air we could pick anything but what we picked is let's take targets target neurons like 12 and turn it on I want a stimulus that will cause neuron 12 on and all other neurons to stay off right so this is a request of the model please control this portion of the brain in this manner please design an image to do that for me those of you who know vision these neurons have overlapping receptive field so that means it's very hard to just put energy on the eyes in particular physical spatial location and do this that's what shown on the left so this makes this a challenging task to do but we said model please try to do it now what's a baseline here well here's searching through a whole bunch of images trying to find one image of the set that we had tested that would drive this neuron well and this is the best image we could find it tended to drive that neuron site 12 really high and but you notice it also has all this like so-called off target activity it's driving all the other neurons high too so it even though we tried to find the best neuron for both these goals it was not able to do this but the model says here's what I want you to do show this image to the to the monkey this is a kind of synthetic image it's not something we would of course found on our own it looks like it has some structure in it that looks vaguely like that chair up at the top image that we did show that natural image if you if you look at it the red dot is the receptive field of the neuron again where it tends to like luminous energy to drive it and so here's what actually happened here's the result for this neuron and you can see with that synthetic image it really drive neuron 12 high and the other neurons are much lower not perfect again but much better than we had before so again half full half empty we got a lot of control power that we didn't have before out of this model in this population state I'm really excited about this not just because it's sort of validating the models are reasonably good because it makes me start to think about could we control neurons in downstream areas that are involved in things like emotional state or things that might improve human health so and you would you could imagine being able to control this without even being invasive in the brain you might be able to better design ways to essentially put energy into the brain to do these kind of things as I've sort of shown here with this neural control in a deep visual area v4 okay so that is kind of the state of the art of what we're able to do with those models at the moment of course that work continues in the lab I'd like to sort of end the sort of the end part of the talk and I've got five minutes here to say like what's gonna happen next with these kind of models I told you no deep DNA and yet passes our tests so what's next well really the race is on to find better models I've been saying that all along we found a good model got to find a better model and the way we're doing that is break current models then use those breaks to guide better models for this and then come back and build and break the next models this is really kind of the loop of science if you will but we're in very complicated model spaces we built a platform called brain score that's helping us do that in the interest of time I'm gonna sort of sort of skip through that but basic idea is that it is just a way to measure how a model is doing on many many aspects of neural matching to the ventral stream it's sort of lots of matches at lots of levels of detail and again if you're interested you can go to brainscore.org again it's a way of taking models converting them to brain style models testing matches to the models and and then assessing their activations and this is in collaboration with my CBRIC collaborators here at Purdue so you can go to brain score you'll see the current leaderboard of models of the ventral stream these are some of the folks that have worked on that we can add more models we can add more benchmarks and this is sort of ongoing work so let me give you one quick sort of vignette about this Cohedage Car was a post doc has recorded neurons in IT not just at these kind of chunky time windows that I showed you earlier but a very fine time resolution here for each image and one of the things that he reported recently is that the models the brain spits out solutions to these object challenges at different time points which is what's shown on this plot here and you would see those lines reaching these kind of top these kind of accuracy decoding accuracy level threshold levels at the range of times that has inspired a whole new family of models that we built in the lab called cornet with some of the folks in the lab these now are recurrent structured com net models that we did not have before but the sort of brain score guidance through those kind of data has provided us direction to those models those models show a very good transfer performance and accuracy given the their overall depth if you will and this is something that we've recently got out of this that's now being presented this year at NERIP so we're continuing that line of work but that's the bleeding edge of where we are so I want to end by giving you some take home messages for those of you who maybe didn't catch that whirlwind of stuff the main message is one main message is the brain's ventral stream produces an IT neural population that carries linearly decodable image generalizable solutions for all tested core recognition tests that was essentially background I gave you and if you optimize deep artificial neural networks for core recognition tasks they lead to internal representations that are remarkably similar to those that are in the brain and I showed you that and that's sort of older work to me at this point but I think that's a quite powerful result it's consistent with but does not imply that the brain uses some form of classical stack prop my preferred interpretation is that these are just examples of convergent evolution the brain has evolved the way of solving this problem and a kind of computer vision using deep networks has evolved the problem that a way of solving it that are not that far apart but there again they're not perfect matches at the moment and that's an area for improvement in building better models of the stream is how we train those models I showed you that in this control work that these same even the models even though they're not yet fully correct the half empty part they are half full in that they can be used to guide the construction of novel synthetic images to drive the brain in the states that we decide we want it to be in and I showed you evidence for that being true recently published and then no existing model is yet identical as I said if we think we don't yet have the right model yet we made a lot of progress we don't yet have the right model one differences recurrent structures that are missing from current models and and that's again some of the evidence I briefly showed you and we built some models that starting to incorporate those recurrents so this is sort of the leading edge of where we are in those kind of things but if you take nothing else from this talk it's this interaction between science and engineering engineering with neural network models is driving science and science we think can drive the engineering around those models and we are rallying around that around brain score in my lab and if you're interested in that please reach out to me I would love to hear from you we think there's many applications to this neural control AI, BMI those are the things that we're dreaming about as we build these scientific understanding of what's going on in the ventral stream and with that I'd like to say thank you thank you very very much it's our pleasure and thank you this is wonderful in comparison to the Rayus Monkeys brains as to guide your experiments I was wondering how similar two Rayus Monkeys brains are to each other it's presumably not all brains are exact carbon copies of each other right now that is a great question the question is how similar are brains to each other what's the individual differences we don't yet know the answer it's one of our ongoing projects you need a lot of data to make those comparisons we'd like to think they're linear rotations of each other so it's a feature space and you is a linear rotation to me that's going to be true for low level visual areas the quit at some point it's going to stop when we start getting to memories yours are going to obviously be different than mine and IT is right at the edge of memory areas so it's a very interesting open question it also relates to the we're thinking about sort of you know if you wanted to could you control sort of let's call it an adversarial attack on monkey a that doesn't affect monkey B sort of could you control things differently those are the directions that those questions had these are forward looking great questions that were engaged on but we don't yet know the answers to there's a question here and then back there please or vice versa there are many technological advances which are very different from the biology works like the way we fly or the way light bulbs work are very different from how the fireflies emit light so why do you think it's important to reverse engineer the brain to achieve a functionality I think the question is about do you need to reverse engineer the brain to do engineering I think that's the question and I would be the first to say I think we're going to discuss on the panel you don't necessarily need to what's the evidence that it's useful well all you know basically the leading computer vision systems now are styled after neural network models right so that's evidence that some of those ideas have had payoff but the art is choosing which ones to use going forward and you know you don't want to build planes with feathers is what we like to say right but that finding that balance is a part challenging I think we'll discuss that here but I will say unless you do this my talk was hopefully not just about AI but once you have a model you can do things for human health and BMI that you're not going to be able to do unless you make those mapping so our mission is broader than just this isn't just an AI mission for me this is like find a model of the system which will have AI payoffs but also many other payoffs again BMI better ways to to do human health and I refer to some of the inklings that we have for that but it's still early days I mean we have a lot more to learn about the brain but we need to think of the brain as an engineered system to make those products those gains even outside of AI so I hope if that's a key message that I hope that even if you don't think we need to build do this for AI there's all these other reasons that we definitely need to do it the AI one is more of a bet and we'll discuss that here yeah and just one last question here before we start the panel and then the remaining questions will be for the panel with the growing innovations in technology such as AI it seems to be an increase in fear in the workplace such as factories and steel mills and whatnot that these technologies will sort of take over the human jobs will make the human workforce redundant how substantiated are those claims do you think well I'm not sure I'm the best I'm not really the best person that's a sort of labor economy question I'm a neuroscientist I don't know if I really want to speak to that although I think it is something that is something we most have to pay attention to maybe sooner that's much sooner than worrying about terminators coming up and taking over so I think you're right to be asking those questions a lot of folks at MIT are asking those questions too and I think it's a broad discussion across many everybody who's working in AI and maybe that will come up in the panel where we'll have other expertise I think also to speak to that I'd like to think we need to understand how our brain works for human health that is going to be a good thing to do isn't about labor markets again it's not just about AI for me I'm here on an AI panel but this really to understand how the brain works for many other reasons besides AI which could be used for not could end up making some jobs go away and you're right and that's an important point of discussion but I hope that we can discuss that more together on the panel if that comes up yeah