The Stilwell Brain

"I think, therefore, I am." But am I? I think. Ha. A single microscopic brain cell cannot think, is not conscious, but if you bring in a few more brain cells, and a few more, and connect them all, at a certain point, the group itself will be able to think and experience emotions and have opinions and a personality and know that it exists. How can such astonishing things be made from such simple ingredients? Well, answering that question means learning not only who we are, but, more importantly, how we are. Today, using what neuroscientists know so far, I am going to make my hometown function like a brain! ( all cheering, applauding )

A single brain cell is tiny, both in size and abilities. But when enough are together, they can do amazing things like be aware of themselves. When the collective power of a group working together is greater than the sum of their individual parts, that is called "emergence." In a similar fashion, we as individuals are connected to the people around us. Those connections form communities that, when functioning properly, can work together to accomplish amazing feats. A great example is "wisdom of the crowds." Even if not a single person in a crowd knows the right answer to a question, collectively, they could all somehow know the right answer. In 1987, economist Jack Treynor conducted the "Bean Jar" experiment. He asked 56 students to guess the number of jellybeans in a jar. Now, as you can probably guess, not a single one of them guessed the right answer. But amazingly, when he took the average of their guesses, what he got was a number within just 3% of the real answer. Now, some people guessed way too high, but others guessed way too low, so all together, their errors balanced out, and from a whole bunch of wrong guesses, the true answer emerged.

What else can a crowd do? If I got a bunch of humans together and had each one of them act like a brain cell, turning on or off in response to the actions of other people, could I make a neural network like the one in our brain? And if I had enough people, could intelligence, emotions, a mind, emerge? If I recruited every single person in the country of China and arranged them like neurons, would the result not only be a simple brain, but something that can think and feel and be aware of its own existence? Well, this is the China Brain thought experiment, first proposed by Lawrence Davis and, later, Ned Block. It's never been done before and, well, unfortunately, I don't have access to everyone in China. I made some calls, and like a lot of them are busy. But the first step is to see what a crowd in real life could even do. This hasn't been done successfully before, but I want to blow a neural network up to the scale of a crowd. And what better crowd to use than one made of the people whose emergent properties made me who I am today? That's right, I am going home to Stilwell, Kansas.

( birds chirping )

Michael: For help designing the brain we would make out of people, I recruited Chris Eliasmith, director of the Center for Theoretical Neuroscience at the University of Waterloo. So Chris, we're headed south, going down to the heart of Stilwell, - where I grew up. - Nice. We're going to do something a little bit weird. Um. I want to create a brain. - Right. - OK? But with a crowd of people. It sounds like a challenge, for sure. I looked into it, and I found that the roundworm has a brain that's made up of only 300-some-odd neurons. - That's right. - We can get 300 people, and where better to get these people to make a brain than my hometown of Stilwell? This was the community that, in many ways, made me who I am.

Michael: This is all downtown Stilwell. Some of my earliest memories are from here. This used to be, and maybe still is, a feed store, and they would have sno-cones during the summer. It was the most awesome, delicious thing ever. But as you can see, a lot of corn is grown in Kansas, but around here, the main thing that I saw being grown - was just sod. - Oh, really? Yeah, There's a famous sod farm around here whose slogan was "High on grass." - ( Chris laughs ) - It was pretty...pretty edgy for the time. OK, so back to the brain that we're gonna make. You know, building brains is in my job description. I wrote a book called How to Build a Brain.

Michael: Chris is known for his neural network, the Semantic Pointer Architecture Unified Network, or SPAUN, which is one of the world's most complex computer simulations of the brain. It uses 6.6 million simulated neurons to perform functions like counting, reasoning, and image recognition. SPAUN is cutting-edge, but neural networks are nothing new. The first was made by Dr. Frank Rosenblatt of Cornell University in 1957. His network, called the Perceptron, was designed for image recognition, and he hoped it would become capable of learning, just like a brain. But the project was only partially successful, and after some controversy, fell by the wayside. It was only when researchers in the 1980s came back upon Dr. Rosenblatt's work, and as computing power increased, that the field of artificial neural networks came back to the mainstream. Today, it is alive and well. SPAUN, and even neural networks used in self-driving cars, are expanding the possibilities of computer learning.

If I want to make a brain out of people, where do I start? That's a good question. I think the first thing we want to do is figure out what we want our brain to do. I would recommend something like vision. Vision. Let's make this brain see.

Michael: Before we can design the intricacies of the brain we're making, let's look at how visual processing works. Let's say we look at a cat. Light information from every point on the cat lands on the retina. This information gets sent to our visual cortex. The visual cortex is structured in layers-- V1 through V6. Each of these layers are made up of neurons activated by specific features, like lines, angles, and shapes. The features that are detected are sent to the infratemporal cortex which puts all the pieces of the image together, and we get our Eureka! moment where we recognize the object we're looking at, what it means what feelings we have towards it. I love cats.

But what should we have our brain recognize? We don't want really high resolution images or images that depend on too much detail, - Ok. - so things like letters and digits. Let's say we use digits. Ok? - Ok. - I want to be the one who draws a digit, and then you will be on the output side. You should be able to determine what I've drawn; not because I showed it to someone and they telephoned it back to you, but because they processed it intelligently. That's what we need to figure out, how we're gonna show an input to our people. So we should take some small number of them and put them at the front, as the retina, and really just show them each a little bit of the image. So if, for instance, we're able to put like 25 people in that kind of front row, the "input" layer, then whatever image we show should be made up of 25 pixels. - Exactly. Right. - Twenty-five pieces. I'm gonna draw 25 people. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25. See? I can count.

These are our retina cells, and each one is an individual person that's literally standing, like, in a field. What do they then do next? They merely need to indicate whether or not their cell is on or off. - All right. - So, they should start firing. They should start spiking like a neuron. What could they do to indicate that they're firing or not? They could jump up and down, they could wave a flag... OK, I like that.

When Chris and I use words like "firing" and "spiking," we're talking about how brain cells, neurons, talk to one another by sending an electric message from one cell... to another. It's called an "action potential," and it travels down the axon of the cell. When the ionic flow into a brain cell reaches a certain threshold, the cell will fire an electronic message down its axon. So a neuron can either be on or off. It's either firing or it's not firing. What we need to find is a way for a person to be either on or off-- raising a flag or their hand should do the trick.

To illustrate our visual input, I will be drawing a number from 0 to 9 onto a grid divided into 25 squares, or pixels. Now each person, or neuron, will receive one pixel. If a neuron receives a pixel with writing on it, it will fire. The V1 layer identifies pixels in the retinal layer that form particular lines in the number, and the V2 layer identifies particular combinations of lines from V1 that form angles. - What does V3 do? - V3 is more sensitive to color. We're only working with black and white in this case. So you're saying we won't even need to have a V3 in our brain? We're skipping V3 altogether. All right, sorry, V3. - So we're gonna go straight to V4? - Yeah.

Michael: V4 neurons will fire when their assigned combinations of angles have been detected. At this point the basic shape of a number is beginning to take form. And so actually the next one is called IT. - Ooh! - And that stands for infratemporal cortex.

Michael: Now don't worry. We haven't forgotten V5 and V6. They exist, they're responsible for higher-level image processing in our brain. But for our demonstration, we don't need them. We do, however, need the infratemporal cortex, which is the final layer needed for visual processing in the brain we're designing. Our IT will consist of ten neurons representing the numbers 0 through 9. They will be looking at neurons in V4, and will only fire when their corresponding neurons fire. For example, if one or multiple V4 neurons representing the shapes of an 8 fire, the IT neuron representing the 8 will also fire. Voila! We just recognized a number.

So what happens after the infratemporal layer? So after that, I think that's where I'll be. It's gonna be me making a decision about what digit I think was actually shown at the end. So I think we're gonna need a couple hundred people, so one question is, where do you put that many people? I would love to use my high school football field. The question is, is it gonna work? - I am hopeful right now. - We got our work cut out for us.

Michael: Knowing how we want to structure our neural network, it was now time for us to head to my high school football field where our "brain" will take form. I spent a lot of time here impressing the world with my body's prowess-- at the clarinet. OK, so Chris, I brought you here because we need to talk about the actual logistics of getting all these people together. So here's the plan as I see it. Everyone is going to be wearing a shirt that is a different color, based on what layer they're in. We're also gonna give everyone one of those, um, like, "I'm running in a marathon" kind of... - The big bibs? - Bibs. Thank you. Yes. I knew that too. You see, I'm always checking, because Chris is one of those nerds who doesn't know what us sportos talk about.

Anyway, the bibs will give every person an individual number, so if something goes wrong, we can target that one brain cell and say, "Are you damaged? What do you need?" Take out the problem. OK, we're here on this 40 yard line. I will be here. This is gonna be the input layer. The retinal cells will be all in front of me, all 25 of them. You are gonna be way down in the end zone on the output side. And I'm gonna be using the scoreboard. - Ok. - So when you make your prediction, based on what you think the brain has figured out, we'll put that on the Visitor's side and I'll reveal the Home number as what I really wrote. - Sounds good. - So literally from here to that end zone is the amount of space we're going to need for these hundreds of people to also have the right eye lines. Just have to make sure that communication lines are open, meaning that it's easy to see whoever you have to pay attention to.

Michael: Our human brain will have a couple hundred people spread out in five layers across half a football field. Every single participant will be assigned to only react to certain neurons in the layer ahead of them. And it's complicated, so their positions on the field had to be carefully chosen so that every neuron has a clear line of sight to the neurons they are connected to. In a way... something will be born on this field tomorrow. ( laughs ) - We shall find out. - We'll find out!

All right. ( crowd chattering )

Michael: So what does it take to turn Stilwell into a brain? Well, seven tents, 550 chairs, twenty gallons of coffee, three hundred flags, t-shirt and hats, our drone operator Jeff, this cute little Gator, two hundred cinnamon rolls, and of course, our medic, Brian. Now all that's left is to pull this all off. A community is something that is bigger than the sum of all of its parts, and so is a brain. Now, today, I'm feeling pretty excited about the neural network we're gonna build out of people, because there's a zero percent chance of rain, but a 100% chance of brain. OK, we better get started.

The gates are open, and our neurons are filing in. First, they're all given color-coded t-shirts associated with the layer of the brain they will represent. Just want go in the center?

Michael: And then they will take to the field to get in position. You all in the orange shirts are the retina. Your job is to say, "Is there writing on my square?" or "Not." If your square has any writing or black marks on it, raise your flag-- oh, and stand up. Now I'm sure all you mega-brainiacs out there remember every detail about how this brain is going to recognize numbers. But just in case you don't, here's a refresher. I will draw a number on a 25-pixel grid, break the squares up, and hand them out to every person in the retina layer. The retinal neurons will only fire if they have writing on their pixel.

The people in the yellow shirts, you guys are V1. Each V1 neuron will be watching three retinal neurons in front of them and fire only if all three of their assigned retinal neurons fire, revealing lines that make up the number. You guys are V2. You're a bit more advanced. You're looking for combinations of features that make, for instance, angles. The V2 neurons will be watching the V1 layer, and fire only if their assigned V1 neurons fire, revealing angles. V4 neurons will be looking at V2 neurons. Their firing reveals combinations of angles that begin to form the number. Finally, the purple shirts. You all are infratemporal cortex. Extremely important role, not more important than the others, though. Part of the IT's function is to inhibit incorrect results it receives from the V4 layer. For example, if V4 neurons are indicating both a 6 and an 8, an 8 will outrank a 6 because an 8 has more features. Chris will determine the number by interpreting the results from the IT layer. Got it? Good. Because it's happening now.

Michael (over loudspeaker): All right! It is time for me to draw my first numeral. Stand by. There it is. Now it's time to distribute these pixels to the photoreceptors in the retina. 25, 24... 19... 13... All right. I have distributed the input to the retina layer. Is everyone ready? ( all cheering, applauding ) Three, two, one, think!

And they're off. The retina layer has fired, passing off signals to V1. V2 sees V1 firing, and also fires, cuing V4 and IT. There's a lot of flags on the play, folks. Look at all this processing. Good work, stay up, I'm now walking back to Chris where he will tell me what you guys have processed. All right, Chris, it actually happened way faster than I thought. It took me forever to get over here. They were already done "thinking." It was really interesting to watch. We had a little bit of noise in the system, for sure, because we actually kind of have two answers at the end. So I'm going to be doing something that brains do, which is kind of make a guess sometimes based on the best evidence.

Michael: All right, Chris. What numeral do you think I drew?

Chris: I think you drew a 3. Let's get that up on the Visitor's scoreboard. The numeral I truly drew... was a 3! ( all cheering, applauding )

Chris: Nice work. There was some noise in the system. I think we can perfect this a little bit, because it wasn't a confident 3, but Chris did still get it right. And by "Chris," I mean all of you.

Michael: We gave each of our IT neurons a clear tube and plastic balls to make their job easier. Every time they see one of the neurons they're watching fire, they put a ball in their tube. If a neuron they're inhibited by has more balls than they do, they stop firing.

Michael: Our model had a mistake. You should have been inhibited by 254, meaning that if 254 is firing, and puts a ball in the tube, - you just sit down. - OK. But we didn't have 254 written down into your code. I'm gonna do that right now. With that kink worked out, it was time to try again. Take that one. Thank you.

24... Michael: Three, two, one... go!

Michael: Ooh. There's not much happening in V2 and V4. Or IT, for that matter. Looks like our brain died.

Michael: Something definitely went wrong. The processing stopped in V2. So, I think our brain broke. Not a single person in V4 or the infratemporal cortex has been activated.

Chris: Seems like something very strange happened when the light went through the lens and got to the retina. OK, but what you really mean to say is that I handed the pixels out to the wrong neurons. - It's lookin' that way. - Wow, who would have thought that the worst-working part of the machine would be the actual people who get to be people - and not brain cells? - I know, right?

OK, I think I'm gonna do an 8 this time. And I'm gonna put it kind of up in this corner, or along that side. This is pretty weird, it's not centered, it's not filling up the whole space. Let's see if our brain can recognize it. Each pixel of our image needs to go to a particular retinal neuron. To make sure I didn't mess it up again, we put numbers on the back of each one. Is everyone ready? - ( all cheer ) - OK. Three, two, one, go!

Michael: Oh, yeah. Whoa! That was fast.

Michael: Blink, and you'll miss it, so let's take a break, because this is the Mind Field Play of the Game. The 8 I drew contained 13 pixels, and bam! the 13 retinal cells connected to those locations are firing. Now, that's what I call a sensation. V1 reads the formation perfectly. They don't even know it, but each one firing means a horizontal, vertical, or diagonal line has been caught. Now look at neuron 40's speed. It's sensitive only to a horizontal line low and to the right, which my 8 had. If retinal cells 23, 24, and 25 all fire together, such a line has been sensed, and watch this--boom! Champion reflexes there, folks. If the V1 neurons a V2 player is watching, fire, they stand, and that means that the lines V1 caught made some corner angle. Standing V4 neurons are shapes made by those corners, but it all comes down to IT. A bunch of them fire. Lots of numbers contain the shapes I drew, but there can only be one MVP, and, dang, look at this teamwork! 3 is inhibited by 8, 0 is inhibited by 8. If 8 is getting this much activity, sit down and let that neuron score! This, my friends, is what we in the cognition sports biz call...a gr8 play.

Chris (over loudspeaker): The purple people neurons make me think that you wrote an 8.

An 8? Let's put an 8 up on the scoreboard. And as it turns out, the numeral that I did write... was an 8! ( all cheering, applauding ) - Nice job! - That was good.

Michael: This was our first definitive success. Now it's time to really put the system to the test. So I'm gonna draw a 1, but then I'm gonna add a line down here and I'm gonna do a dot right there. And we're gonna see if that noise trips up our brain.

Three, two, one, go! My guess is a 1. The number that I drew was, in fact, a 1. ( cheers, applause ) 7...and I'm gonna put a line through it. Go! Oh, yeah. Man, these people are good.

Chris: The brain thinks it's just saw a 7.

I wrote a 7. - ( cheers, applause ) - Nice work. Nice work.

Michael: With two successful results in a row, for the last test I want to see what will happen if I really mess with our brain. I'm not even going to draw a number; instead, I'm going to fill in every single cell. This means that every single neuron in the retina will fire, will stand up and wave. Let's see what happens. My prediction, of course, it's gonna look like an 8, because an 8 is a numeral that fills in a lot of the cells. Are you all ready? - ( all cheer ) - Good. - Chris, are you ready? - Ready!

Michael: All right, I'm ready. Three, two, one, go!

Chris: What?! ( laughing ) - ( all cheering ) -

Michael: Look at that. - Michael: Holy cow! - ( Chris laughs ) It looks to me like you essentially opened up the eye and shone a laser right into it. Right, yeah. So what I did is, I didn't even draw a numeral. I just scribbled all over the whole thing, I filled in every single cell.

• So what does the brain think that I drew? - The 8.

I had no idea the inhibition would work that well. We were able to single all of that mess down into just one guess, and it really was the smartest guess. Right, it was the one with the most features in it.

Michael: Congratulations to the entire infratemporal cortex, V4, V2, V1, retina. You guys have been amazing. Great work. ( cheers, applause )

Today, I was a neuron. My favorite part about the brain was that it actually worked. My favorite part was just the whole experience, doing science, meeting cool people. It's just a really good simulation of how the brain works, and it was just really cool to take some information from that.

Michael: And as always, thanks for...being a brain. ( all cheer )

Michael: Our small-town brain did work as predicted. Made up of only a couple hundred neurons, each one with no idea what number I was drawing, it was nonetheless able to process the image and determine the correct answer. We created a living, breathing model of a part of the human brain. And our demonstration was a new way to illustrate and share how the human brain processes visual information. We were able to watch it process...think... in real time, and that's amazing.

Its success shows what we can achieve by working together, and we only used a small fraction of the number of neurons found in an actual human brain. So imagine how powerful the connections of not a few hundred, but a hundred billion human neurons could be.

Now, interestingly, a hundred billion people is about how many humans have ever existed in the history of Earth. There are only a few billion alive right now, so I guess that means get procreating...please. I want to make a bigger superhuman mind. No. I'd like to thank every neuron from my hometown of Stilwell, and the entire community that supported us. Because without them all working together, none of this would have been possible. And as always... thanks for watching.