Prof. Russell Hanson at the Finnish Biohacker Summit 2016, Brain Backups

Hello, it's really nice to be here. It's my first time in Finland, so it's nice to meet all the Finnish people. So I'm going to be talking about neural modems and non-destructive connectomes. Connectome is the wiring diagram in your brain. It's what makes you you. A little bit of background about me. I've been giving talks at hacker conferences, not so much biohacker conferences, for quite a while. So I guess to me hacking is more like getting root access on a server or something like that that you own. My day job these days is as a genetics professor in New York City. And where we do this type of research. Firstly, I just wanted to explain what is a brain backup and what is it good for. So Brain Backups is actually a trademark name for the connectome, and it's the name of this company that does connectome imaging. And it's the neural network of the connections in your brain. And not only is it the neural network of connections in your brain, it's the network of connections in your brain with some metadata about those neurons, like what is the subtype of the neuron, what types of proteins are on the surface of the neuron. The analogy that I often use is that the human connectome is much like the human genome. So you've probably heard about the human genome. There are lots of services these days like 23andMe that allow the service of providing information on your ancestry and information on, on what genes you have and what risks you have for different diseases. So our goal with, with connectome imaging is to bring the price of the connectome down to about the same price as the human genome these days, which is about $1,000. A question that we often get is, what would you really do with a connectome? And we see many applications of the connectome. First and foremost, the applications are in health. So the connectome can help image different psychological diseases or do better memory experiments, like Julia Shaw said earlier today. And really, almost anything that's involved with diseases of the brain can be reduced to aspects of brain circuitry and brain wiring. In addition, you know, people spend a lot of time and energy going to school to learn things, and that's basically just to put information into their brain. And if you have a full wiring diagram of the human connectome, then you could conceivably put information into someone's brain without having them go to school or pay for an expensive education or read Wikipedia or however they educate themselves. So in addition to the health and education aspects, the technology can be used to make other technologies, such as you could have like a personal assistant that has information that only you have access to. So you wouldn't have to train a helper or, you know, you can have your minions do your dirty work for you and you can do the things that you are uniquely suited to do. There are also entertainment aspects of the connectome, so with the connectome and the interface to the connectome, you can do interfacing with visual, auditory, and motor cortex, kind of like VR these days or VR headsets. And in the business realm, again, the neuromodem and connectome has many applications, like you could do advertising directly to the visual cortex, you can do neuro-advertising, you can see what people are really interested in by looking at their brain as they look at different products, and all sorts of things like that. So again, I get a lot of skepticism regarding this, so I'll just break it down really simply. So what is involved in connectome imaging and indeed what is involved with computer simulation of connectomes? So back in, I believe it was 2014, supercomputers were used to model 1 second of human brain activity. So if you break that down, that means that the supercomputers are about 2,000 times faster than would be needed to simulate the human brain in real time. And so that means something like we have 2 GHz computers now. If you had 2,000 of those 2 GHz computers or a terahertz computer, then you would have something approximating this kind of speedup factor that you need. And you can also do it from the dates in the '70s and the '80s and the 2000s, you know, going from hertz to megahertz to gigahertz and so forth. So the other, the other question that we get often is, so what organisms have we completed connectomes for? And the quick answer is we've completed partial connectomes of some very simple organisms and complete connectomes of the simplest organisms. So the C. elegans worm is the first connectome that was deciphered, and you can, you can grab the Excel spreadsheet of the connectome on GitHub. Today, and it looks kind of like you would expect. So there are an origin neuron, a target neuron, and the type of that junction, the number of connections, and what neurotransmitter. And so this, um, this table just goes on for another 300 lines, and it describes all the neurons in the simple organism. Next question is, what do you do with the connectome? So you could take this the simple connectome of the worm. And you can simulate it, you can modify it, you can study how the worm behaves, you can upload it to a robot, you could do all— just kind of whatever you want. Nobody really cares what you do with a simple 300-neuron worm. One of the most popular neuron simulation tools these days is this program called Nest, the Neural Simulation Toolkit. And again, if you're so inclined, you can download the latest release from 2015 right there. Can we play this video? So what this video shows is just walking through what happens when you do simulate the C. elegans connectome. So this is a connectome simulator. And what happens is all these neurons fire, or you can see that it interfaces with the outside world via the sensors that the worm has, like a sonar sensor, a touch sensor, a smell sensor. And it does simple worm stuff, like crawl around and find food and avoid walls and things like that. So here, this worm connectome was uploaded into a Lego. And it crawls around and does worm stuff. Uh, next slide please. We're stuck on the video repeating or something. Anyway, so here you can see the— Each of these little cells is an individual neuron, and you can see it firing when they turn green. Yes, please. So in addition to simple Lego experiments that, you know, a child could do or a child might be interested in, there's actually a lot of big science going into connective imaging these days, and big science requires big money. You may have heard of projects like the Obama Brain Initiative or the Trump Brain Initiative or whatever they're going to call it now. In America, the BRAIN Initiative is funding a lot of research into neuroscience, and in addition, there's been a huge uptake in all kinds of neuro-business. A really amazing new company called Kernel was just funded about a month ago to the tune of $100 million, and what Kernel is doing is it's investing in ways to implant a CPU into a human brain. They think that that's really interesting to sort of modulate memory in people who have memory deficits like Alzheimer's disorder. In addition to these really fantastic advancements in the business world, an actual whole human brain was recently uploaded, you could say. So all the— the human brain was uploaded to brainmap.org. And anyone can go through it and literally look at every single neuron in an adult woman's brain. And the way you can access that is by going to brainmap.org. You can also just click through their PDF. And so there are thousands of slides that look something like this. So different brain regions are displayed. And then on the right side is a stain of that particular brain region. So this is a non-specific stain, which means you don't get much information about the subtype of the neuron, or really much information other than that there is a neuron there. So as I said, there are about 83 billion neurons in a human brain, but how big is the connectome, which is the network of connections between those neurons? A quick calculation shows that If you have 100 billion neurons and on average 10,000 synapses are between them, you get something like 909 terabytes. This is a gross simplification. And if you say that a terabyte hard drive is about $28, then the price of the storage for all human neurons is on the order of $20,000. What if we use a— so that's about the price of a Toyota Corolla. If you make an even more sophisticated model and you say that there are 500 inputs on neurons, the adjacency list is maybe 2 kilobytes per neuron, you say there are 1,000 neuron subtypes, and that each synapse has 1,000 states, then you have another 5,000 bits. And then you do the calculation again and you get 384 terabytes. So let's say it's in the range of a couple hundred terabytes, using a variety of models for how you calculate how to store this stuff. [NOISE] Our brain is on the larger scale of all of these brains. So neuroscientists study rat brains and mouse brains a lot. And a macaque brain has 6 billion. We have 86 billion. The mouse only has 61 million. And, you know, this is useful because it really does get more complicated as you go up in size from a couple million to a couple billion. So one of the topics that I want to talk about that isn't just connectomes is what we call a neural modem. So a neural modem is much like your typical modem where You transform analog signals that come across your phone lines into digital signals that your computer can interpret. The way that this works in a human or other types of brain-machine interface is you have an array of electrodes that sits on the surface of the brain. Usually you have two antennas, one antenna for power and one antenna for for data, so there's no battery in this device, and it transmits and receives from, in this case, 64 electrodes. And as I'll point out later, this is quite literally only the beginning, so there are several projects that are being funded by the DARPA agency to to bring this number up from 64 to 1 million, and stimulating 100,000 neurons within the next 3 to 4 years in humans. This is very different from the technology called EEG, which you may be familiar with, where you just put an electrode on the surface of the skull and you measure activity on the surface of the skull. So if you put an electrode on the surface of the brain, you get much higher frequencies, and if you have the inner electrode spacing small enough, you can actually measure individual neurons spiking This is an example of one such EEG headset. This guy— it works a lot better if you don't have any hair because the electrodes just go straight on your scalp. And so this guy is controlling a synthesizer using an EEG headset. This is a standard commercial EEG headset that transmits over USB and has a lot of neat applications. Like you can lift rocks and play games and stuff. These are the electrode signals coming off of the headset, and on the upper left you can see that if there's a strong connection, the electrode's green. So one might ask, well, what if we wanted to make a Wi-Fi connected brain? Then you could telepathically talk to all your friends who also have Wi-Fi connected brains. And given that there's projects to do a neural modem without a Wi-Fi chip, why not just add a Wi-Fi chip? Indeed, there have been a couple of telepathic communications from the US to India wherein someone communicated through the visual cortex, basically they saw a bright light and they were able to transmit information, binary information, over thousands of miles just using their brain. As I mentioned before, the motor cortex and this project called the Neural Information Engineering System Design from DARPA has this huge outlay of funds, approximately $60 million for the next couple of years, to record from a million neurons and stimulate 100,000 neurons and have it implemented in people within the next 3 to 4 years. And one might ask, well, what would you really want this neuromodem for apart from telepathy? Well, the government is really interested in re-enabling soldiers who have been injured. So if somebody has a spinal cord injury, if you put this device onto their motor cortex, then they could conceivably control an external skeleton and also interface with their sensory cortexes. The impetus to do non-destructive, non-invasive connectome imaging is that Currently, the technologies to do this essentially involve taking a brain out of an organism, slicing it very thinly, and putting it on a microscope, whether it's an electron microscope, an optical microscope, or any other type. And we figured that this was kind of sad because essentially you can only take one measurement of the brain. It's very destructive. And you can't image the process of diseases like Alzheimer's. You can't image the way people learn because if you can only take one image, it's kind of frustrating as a scientist. In addition, I just wanted to mention that a lot of the cryonics stuff is— it's difficult to reconstruct the connectome from dead brain tissue that has been sitting for more than 15 minutes. Basically, upon death, the cellular processes work very quickly to degrade tissue, And once that tissue has been degraded, there's nothing you can do about it. So a lot of people who are into brain preservation believe that, you know, within, you know, a couple of minutes of death, you should be injected with some kind of protectant or cryopreservant or something that essentially fixes the tissue so that this degradation process doesn't happen. Because if it does happen, there's no going back. Whether you're into cryonics or not is of course your personal opinion. I wanted to mention that apart from these types of optical imaging methods, there are ways of using genome sequencing to actually do connectome imaging. So the way that this works is you have a virus that infects a bunch of neurons, and then each virus has a barcode, and then you take the cells and you do the genome sequencing on them, and you use the barcodes to to decipher where the virus infected them. So the— with this, we'll call it a one-step process. By grinding up the cells, you could actually get a connectome in one step for, let's say, about $1,000. The trouble is, of course, in order to get it onto the genome sequencer, you have to destroy the brain once again. One approach to this is using contrast agents. So using CT or MR— CT is essentially bouncing X-rays off of your sample— you can image down to a micron or a submicron. Most of the interneuron distances— about a micron is good enough to get most interneuron distances. And gold, iron, potassium, iodine, gadolinium can be used for MR and CT contrast. So once you have this image, what can you do with it? And there's been a lot of progress recently to decode visual cortex and decode auditory cortex. So the idea here is, can you take a direct measurement of the brain and see what the person is seeing? What they're seeing, or can you take a direct measurement of the brain and hear from that measurement what the person is hearing, or hear what they're articulating, or hear what they're articulating in their mind? For instance, you can have a piano player playing piano, and take away the piano, and you can still hear what they're hearing in their mind from a measurement on their auditory cortex. I'll explain this experiment briefly. So this was an MRI where a person was presented with some images, and they took the MRI scan. And then they use the model that they built from when they knew what the picture was to infer what this new model is. And so the reconstructed image here is on the right. As you can see, it's not very accurate. MRI has many inherent flaws. One of the biggest in terms of doing these experiments is that the minimum size of the voxel you can measure is about a millimeter. So a millimeter or 2 millimeters is really not a lot in terms of measuring the activity. You can measure regions of activity, but measuring very accurate activity just doesn't work. Next one. So in this experiment, direct— they're called ECoG arrays, they're basically electrodes that are placed on the surface of the brain. Were used to infer what the person was hearing. And if we can get some audio on this one, you can hear. "Pencil." So this is just with 64 electrodes, and you can kind of make out the word. Of course, it sounds garbled, but just using this, somebody who has injury to their ear or somebody who wants to have some kind of interface could, using this method, at least have the option of getting some kind of really garbled sound out. Directly from an eCochG array using technology from, let's say, 10 years ago. So what do these neurons look like if you zoom them up? Here is an image of some quantum dots on different dendritic surfaces. So these are quantum dots with, like I mentioned, the ligand for different receptors on it. So these are quantum dots that have a glycine receptor ligand on them. And so where these dots are, are where the glycine receptors are. And the groups of them show where these are more densely packed. This is about the resolution you want, so the scale bar is 500 nanometers, or half a micron. And this— the whole process just kind of gets more technical from here. So the brain is very complex. And there's tons of data in the brain and tons of subtypes of neurons and tons of proteins that are on the neuron surface and lots of metadata that you can store about the neuron. But I will briefly go through a couple of the features that are required to measure this. So firstly, you need adrenergic receptors, dopaminergic receptors, and GABAergic receptors. And these all have a number of different subtypes. So measuring the subtypes of dopamine and GABA receptors is very important, as they have different functions. One of the kind of trade-offs that you get into as you start measuring through the skull and through a couple of centimeters of tissue is what type of imaging do you really want to do? You could do X-ray and MR, as I mentioned, but MR only goes down to a couple of millimeters. So with, with CT, you can get to 200 to 400 nanometers, but the radiation is pretty intense. So what that means is that if you put someone in a— if you put an animal in a CT machine of this type, it will cause irreversible damage and they will die. The current X-ray CT at this level is a postmortem scan. However, you could do near-infrared imaging. So near-infrared imaging can image up to 10 centimeters through tissue, but it has other disadvantages. So there's the near-IR1 window and the near-IR2 window, and the trouble is that it's not a very clear image. It looks like you're looking through very dense fog. And then you can go on to fluorescence microscopy and electron microscopy, and there again, there are disadvantages and advantages of each of these methods. I'll just give one example of technique that we use, which is a nano-CT machine. So the nano-CT machine has the advantage of that you can go through any depth of tissue and also achieve a submicron image. So these nano-CT machines are rated to— it's a 200-nanometer voxel to 500-nanometer voxel, and these are commercial machines that you can get from GE. This is one of those scans This we got off of a synchrotron CT machine. Again, this is a submicron voxel, so it's 0.6 micron. And it was a— this is a slice of mouse brain. And you can see clearly individual neuron bodies, and you can see axons going through the sheet. If we zoom it up more, you can see more features. And so one of the advantages of this is that it's really fast. It only took 6 minutes to do the scan. You don't have to take slices of tissue and carry it across the room and image it on an electron microscope, which is both very expensive and very time-consuming. As we progress into, you know, what's left of technologies, this type of nanobot animation is a little bit fanciful, but it provides a goal. Instead of having a nanobot that has arms that extend, you can have a nanobot that has strand replacement ligands or strand displacement ligands. So you can have a nanoparticle that attaches to a single site and then moves on to a second site and a third site and so forth. Next slide. So this is another volume from an electron microscope, and I like this animation because it shows, you know, what are these synapses that we keep talking about. And what this image does is it goes through and it reconstructs all the synapses. You'll see them as red pancakes in a few moments. So it's constructing the red pancakes, and then it'll take them away again, and so you'll be able to see How many red pancakes or synapses are there in this volume of tissue? It's quite dense. Next slide. One of the questions that people often ask is, well, if you do have this connectome, is it really possible to compute the, you know, consciousness or the functions of a human being? Is it not just too complex? Aren't there too many unknowns? And one of the theorems from computability theory essentially states that if there's a finite amount of information, then any finite state machine can model that finite amount of information. So if you were to claim that there's infinite information in the brain, I think that that would be antithetical to the facts. So most people who work in this field don't believe that brain functions cannot be effectively modeled using computational models. In addition, I mentioned a couple of uses for these brain-machine interfaces that address medical applications or injured soldiers or spinal cord injuries from car accidents. And I think that people who have endured these types of injuries, it's not ethical to not provide them with this kind of interface to allow them to interface with their world or to have a somewhat normal semblance of life. So I think that the argument that, you know, we shouldn't be doing this because it's dangerous, you know, how are you going to tell that to the person who who got in a car accident. A little while ago when I gave a similar talk, I asked, you know, what's next? And the first thing on my list was an implanted CPU with a database of neural codes. Guess what? Just a few weeks ago, this company Kernel was funded to do exactly that. The neural codes are the way that the— it's kind of like a binary code. And the way that the brain talks with other brain regions is via these little neural codes. It's kind of like binary or ASCII or what have you. Secondly, the process of using deep learning to improve these interfaces, for instance, like deciphering all the synapses or the red pancakes in that image is a current area. So you need sophisticated machine learning models to be able to process all these data and make them usable and interpretable. As I mentioned, this neuromodem thing, what's happening is that there's become increasing business cases for building these types of things. So before, you know, you might have a prototype and you'd sell 100 of them, but if you start having to build 100,000 of them, then the companies start caring because they can make money off of building these things. So right now it's kind of a process where big electronics companies are seeing the business case in building devices that address these issues. And lastly, you know, maybe sometime in the near future there will be a $1,000 connectome and we'll have it just like we have a $1,000 genome these days. It's been great to be here and I'd really like to thank the Biohacker Summit people and all the volunteers out there who are donating their time, and you for listening. If you have any questions, find a microphone. Can I get a microphone over there? Let's go on the left.

Hi, thank you for that talk. Thank you for sharing these results with us. Maybe this is a silly question, but what other neurological imaging techniques do you see up and coming? Kind of the takeaway was that we still have to improve the way we take these images. Have you heard of any unproven still, but new neurological imaging techniques?

Honestly, I think I covered most of the methods.

They're, you know, genetically encoded, validated, the scientific validated methods. For example, I know of an auditory device that kind of flipped around. That's some research— researchers are trying to prove that they're collecting some data. Do you know any kind of futuristic imaging techniques maybe? Thank you.

For auditory devices, the ones I know are direct sort of cochlear implant type things. Maybe you know something I don't. Any other questions? Comments? Well, thanks a lot.