"Tell me about these Brain Backups, Russell" - NeuroTechBOS

professor at Mount Sinai, and he does genomics and multi-scale biology. He's going to talk to us about brain backups. So I'll hand it over to you. Thanks, Alex. It's great to be here, and thanks to Leaf Labs for hosting. And I'll get right into it. So I'm a professor at Mount Sinai, and I have a startup company that does connectome imaging. It's called Brain Backups for obvious reasons that I'll get into. A little bit about me. I used to be really into computer hacking, like, con-style computer hacking, and gave talks at places like the Chaos Computer Congress in Germany. Vice Magazine wrote about me doing brain imaging. Over the last little bit, I've been giving talks about the same topic in Finland and Istanbul and the Canary Islands and all kinds of fun places. And I also work in New York City at Mount Sinai. Everybody loves a good demo, but the fact is, is that in neuroscience, the machines to do good demos are really hard to carry around in your back pocket or bring on an airplane or in my— the trunk of my car. So I have a lot of videos instead. Please forgive my lack of good demos. You know, stuff like— and in addition, many of the experiments require animals, and if you've ever worked with animals, they live in a very tightly controlled area of the facility called the vivarium and you have to go through many months of training to get access to the vivarium and I just can't bring animals around with me. And when I have it, it was a really bad idea. But you may have seen little experiments like this where instead of putting a mouse in a maze, you put a mouse in a virtual environment and he runs on a little ball. Oftentimes in talks I'll talk a little bit about brain hacking, which we can talk about as ways of interfacing with the brain in new and innovative ways. And for me telepathy was one of the first things that I thought was really cool in neuroscience. So you can actually at present communicate brain to brain with somebody. It's extremely crude and primitive and it's like ones and zeros, but it really does work. In this case they communicated Hello world from India to somewhere in Europe. In addition, people are often skeptical of uploading, which in this case would be like uploading your brain to a supercomputer or a server or something like that. And the fact is, is it's already happened. So if you go to brainmap.org, a little while ago a large group of scientists put together the first complete upload of a whole human brain. And it was actually at a full cellular resolution, so every single cell in the brain was uploaded in this atlas, and it's freely available to anyone. You can scroll through the whole human brain. It was a lady who passed away in her mid-30s, so it's only one brain, but it's a really amazing dataset. In addition, it doesn't have any specific contrast, so you just see general contrast of the cell bodies. So you see blacks and whites and outlines of cells. In our view, it lacks this information about densities of receptors, density of transporters, and more local chemistry. I often get asked, what is a brain backup and why should I care? Well, a brain backup, famously people have said that the connectome encodes who you are. So the brain backup is the connectome. It is who you are. Your genome encodes the things that allow you to be who you are, but your connectome at any point in time really is who you are. And you know, there are people who work on this from health aspects, whether it's Alzheimer's or Parkinson's or dementia. There are people who obviously work in the educational field trying to get information into people's brains and then back out for God knows what reason. People do technology of how you could You could have a virtual assistant or an AI or a robot or— Yeah. Like the bouncing paperclip in Microsoft could be encoded by a brain image. Entertainment, VR currently interfaces through your eyes and other sensory organs. If you could interface directly with those sensory organs, it might work as well. And business, you know, you could communicate from brain to brain. Which would alleviate the need to have like a little piece of glass in your pocket. Again, you know, I just— I'll put the criticisms up front. I get a lot of, you know, "I don't believe you. You don't know what you're talking about," so on and so forth. That's fine. But there's tons of information that I will provide to you that I hope will perhaps inform you or at least steer this audience in that direction. So in 2014, The Telegraph wrote a paper about how a supercomputer modeled 1 second of human brain activity. If you run through the math, that makes it maybe 2,000 times slower than real time. So if you look at how much time it took for current CPUs or our current type of CPUs to get 2,000 times faster, that might only require an acceleration or an amount of time equal to 10 to 12 years. The second thing is, you know, I want to put up front what is a connectome, what does it look like in terms of like an Excel spreadsheet or a database file. And this is a real connectome, this is the connectome of the C. elegans flatworm. And it's an Excel spreadsheet. It's 5 columns, 303 rows. And what it shows is the origin of the connectivity of all the neurons. So the origin, the target, the type, how many connections there are, and what is the neurotransmitter along that connection. The human brain has about 86— yeah, there's a question. [Speaker] Do you think the strength of the connection and how many connections? [Speaker] Yes, that's a great point. So the strength of the connection or the synaptic weight or just generally the strength is also a very key component. So as you learn something, You get better and better at doing that thing, and it's encoded in the strength of the synapse or the strength of the connection. So yes, I would love to have the strength of the connection. I'm not sure if in C. elegans it really matters. I don't think it does, actually. But it does certainly in people. Is this the minimum number of dimensions you need to encode a connectome, or can you continually add more dimensions? Of course, yeah. [INAUDIBLE] You can label it with a chemical dye or antibody or measure it using a mass spectrometer or a bunch of other ways. ELISA. What can I do with the connectome? One of the first things that we like to do is to simulate it. So if you have a circuit with the weights and the connectivity, you can simulate that circuit and you can get— you can attempt to get the same behavior out of that same circuit. This particular simulation tool is called Nest, the Neural Simulation Tool. It's an open source tool that's been around for a long time. You can download it, play around with it if you like. There are many other simulation tools. Brian is one. I know we're a little bit short on time, but I— a guy in my lab made an animation that I want to show you. It's about Brian the Brain. It was just finished yesterday. So hold on a second, I'll bring up the animation. Sorry. I promised I might show him. He just finished this like— Hello everyone.

My name is Ryan and welcome to RyanHackers. I'm sure you're all wondering, what is brain backups and what do we do? Well, that's why I'm here. Follow me. As you look around you, you're inside me. Inside the brain, those tree-like structures are called neurons. There are around 100 billion neurons in the human brain. Each one of those neurons responds and transmits signals to one another. The neuron is made up of several parts that help transmit and receive these signals. A neuron first receives a signal from something called a dendrite. As you can see from the animation behind me, the dendrites are the branches that are extruding out of the center of the neuron. After the dendrite receives those signals, it is sent to the center of the neuron called the cell body. The cell body is the main processing center of a neuron and decides whether the signal that is received is strong enough to transmit to another neuron. If the signal is strong enough, then it causes a chemical-electrical reaction inside the cell body which is sent down the neuron's axon. The axon acts as a pipe to where the signal is sent. Once the signal has been sent down the axon, it is transmitted to another neuron through the synapse. The synapse is where all neurons send and receive their signals. Whether you are walking down the street, feeling angry, or even breathing, the neurons are all responsible for this. While other brain imaging techniques require destroying the brain, Brain Backups uses non-invasive technology to scan so they don't have to destroy me. We hope to eventually compile enough imaging of all the neurons so that we can create something that neuroscientists have coined the connectome. The connectome is the map of the neural network in your brain, kind of like the blueprint for the neurons in your brain. The connectome can help us better understand diseases like Alzheimer's, Parkinson's, and dementia, and help us better understand how to cure them. It can also be used to image of the brain after someone has learned something, and then maybe be able to transfer that knowledge into another brain. The connectome can be used for endless possibilities, and hopefully in the future turn things we think as science fiction into reality. Brain backups is just one part of the scientific community that is aspiring to make this technology a reality. And in the future, we hope to implement the Connectome into our daily lives to make it safer, healthier, and easier to live. That's all for today. Thank you for watching.

Do you like the video? It's the work of Chris Seidel, a very talented animator. [LAUGHTER] This is a quick— another video of how it looks when different neurons are firing in the brain. This is a simulation of the C. elegans connectome, again about 302 neurons. And so very simple computer programs can simulate this very, very small connectome. And you can download it into a Lego robot and you can have the connectome of the Lego robot cruising around on your kitchen floor. The C. elegans is a very simple organism, so it just looks for food and avoids obstacles and avoids noises and, you know, very simple sensory responses. [Speaker] How many connections? [Speaker] I don't know how many connections, but it's about 300 neurons. And a lot of people are involved in the business of doing neuroscience, AJ and Leaf Labs and us in New York. And in the past 6 months, there's been a huge explosion really in companies who are trying to do neuroscience in new ways to improve human health and do all kinds of things. So Bryan Johnson at Kernel and Elon Musk at Neuralink both invested hundreds of millions of dollars into new neurotech applications. And it's really a very expensive endeavor. It's very expensive to run a neuroscience lab and do meaningful work. What do you think of their efforts? I don't want to get into it right now. We can talk later. You know, I'm a very concrete person. I want to, you know, lay out what— how I think this stuff works in simple terms and this particular calculation is perhaps too simplistic for some people in the audience but I think it gives you the gist. So I was curious to know. How much computer storage would be required to save the connectome to a bunch of hard drives. So if you have 100 billion neurons or cell— neuron cells, and on average each neuron is connected to other neurons through 10,000 synapses, and a terabyte is so many bytes and you encode every neuron or connection in one byte, then you get about 909 terabytes of storage to encode just the connections and the and all the neurons in this fashion. And if you— whenever I did this calculation, a terabyte was $28. And if you do some math, it's about $25,000 to save, you know, would you rather have a car or a copy of your brain? So, it's about the price of a Toyota Corolla. You know, There are really countless ways to do this calculation, but a more complex one is if you include— assume 500 inputs per neuron in adjacency list, you maybe need about 2 kilobytes per neuron. Assume there are maybe 1,000 cell types, that's another 10 bits. Assume that every synapse has a state, so what we're adding is we're adding 1,000 different cell types and 1,000 different synaptic states. Then we only get to 300 terabytes and, you know, if you compress it, you know, the brain is obviously highly compressible, you know, the number goes down with compression to 300 terabytes.

What does the upper end look like?

The more complicated models? It's immeasurable. You can keep adding more and more features. You know, we as humans probably think mostly about human brains, but in reality all these different types of brains, as you go from the smoky shrew to the mouse to the hamster, we work a lot with mice. Mice have only 71 million neurons and have many, many different things happening than in human brains. But you know, every type of brain has different things that it's designed for. You know, human brains do human stuff, mouse brains do do mouse stuff. And there are just so much— there are many other things to do in neuroscience other than human connectomics. In addition, there's been a lot of work in doing brain-computer interfaces. This is a project from a group in France. And what they've done is they've taken what's called an ECoG array and they've moved it from a type of system where you could only do it when under surgery to an implantable type of device where you can have somebody have the wound closed and have it stay in someone without having an open wound. These eCochG arrays typically only have about 64 electrodes, so an 8x8 array. But even with these many electrodes, as I'll show in a bit, you can get some really outstanding features. It's currently going through trials in France right now. I might optimistically say it might be approved in France in a year, year and a half, something like that. It's hard to say. And many people get into this through these commercial EEG headsets. EEG headsets are put on the surface of the brain or the scalp and only measure field effects at the surface of the skull, and the effects are really difficult to analyze. However, I was able to control a synthesizer with one. So this is a guy controlling a synthesizer with a commercial EEG headset. It works better if you don't have any hair. That's why this guy is my test subject. So that noise in the background is him controlling the synthesizer with his headset. He's just learning. You know, it would probably be pretty awesome to make a Wi-Fi connected brain. So, you could do ifconfig brain0 up and just bring up your— [Speaker] China could give you mental illness just for kicks. [Speaker] Yeah, you'd get the Chinese version because it's 10 times cheaper and it would send all your brainwaves back to mainland. [Speaker:JAMES WHITTAKER] And in addition, funding agencies like DARPA are putting big money into projects like neural modems. In this case, they want to go from the 64 electrodes in the ECoG that I mentioned before up to reading from a million neurons and stimulating 100,000 neurons in a cubic centimeter. And a bunch of us went down to DARPA and put together proposals and talked with them about the feasibility of doing this type of thing. And like many DARPA projects, it was, I think, in most people's opinion, way beyond what is possible when they proposed it. But having this type of thing could enable direct interfaces with the motor cortex and sensory cortices. How do people do connectome mapping today? Really, it's very slow, very expensive, very destructive. It's all ex vivo and usually it consists of slicing pieces of tissue, putting them on a microscope and imaging them. And— Is 1 million a special number for some reason? It's a round number. Or something? No, it's arbitrary. It's a round number. Yup. And companies like 3Scan have seen how histology, which is the analysis of pieces of tissue, has— is ripe for disrupt— disruption and technologies like the knife-edge scanning microscope are pushing sort of faster slicing methods to do imaging of tissue. In addition, people are really interested in ways to use DNA sequencing to do connectomes in one step. And one way to do this would be to have a pseudorabies virus, which is a rabies-type virus that you would inject into a brain and then the rabies virus would infect the neurons and it would deposit barcodes in between all the neurons. Then you'd crush up the brain and you'd sequence it and you— by reading out where the barcodes were transmitted, you'd actually like in one step get the complete connectome. There are many, many difficulties with this system, but the obvious one is— the obvious advantage is that it's very, very cheap given how— the price of current DNA sequencing. One of our approaches is to use contrast agents to do in vivo imaging. We do this using ligands and contrast-providing agents. And a couple of cool clips. In this case, fMRI was used to decode what people were looking at. So people were presented with images inside an fMRI machine and a model was built and then they reconstructed what someone was looking at based on that model. Weren't they YouTube videos? Yes. This was a group at Berkeley a couple of years ago. The difficulty with fMRI or MRI is that the resolution is a couple of millimeters, and so it's really, really reduced resolution from, you know, all the details that are happening. So that last one was electrodes implanted in the brain? No, it was fMRI. Yes. And this one was using the eCochG array that I mentioned, so a 64-electrode array that's placed on the auditory cortex. And again, a model reconstructs what the auditory images are saying here. I won't talk over it. Did anybody— did everybody hear that, or should I replay it? Okay, sorry. Is that just the reconstruction sound or is that also the source sound? Yeah, it has the source sound and the reconstruction sound. I'll be quiet. Waldo. Waldo. Waldo. Structure. Structure. Town. Town. Waldo. Waldo. They have a piece of matlab code and it just. Yeah, Magic happens. You know, so what the goal is, is to take things like these optical images and electron microscope images and figure out ways to get enough information using in vivo or non-destructive methods. And so, we've been hitting our head against the wall on this problem for quite some time. But this is sort of the gold standard of what we would like. So what this is, is an electron microscope image with quantum dots. Or just gold nanoparticles that are labeled with glycine receptors. And so what you see are all these little black dots are where the glycine receptors are. So if you have a map where all the glycine receptors are, where all the dopamine receptors are, where all the serotonin receptors are, you have the metadata about the neurons that you need to make a meaningful connectome. And the way that we do this is by making ligands for all of the different subtypes of these receptors. So it turns out that there are many different subtypes of glycine receptors and dopamine receptors, you know. Here, there are 5 different types of dopaminergic receptors: D1, D2, D3, D4, D5. There are a bunch of different types of adrenergic receptors, many types of GABAergic receptors, and so forth. But we've had some success. So here is a CT image of some mouse neurons that have been treated with contrast and it's at a submicron resolution. And you can see individual neurons and dendrites and axons as well. So, CT is computed tomography or it's like a 3D X-ray. What you do is you bounce photons in the the X-ray spectrum off a sample. The shortcoming of X-ray CT, as you may well know, is that it causes ionizing radiation. So CT scans are kind of like— what we want to do is we want to get the radiation dose similar to a transatlantic flight, so, you know, something that is safe. But here's some new data. I mentioned this to a couple of people earlier. We took this image last week. This is a whole rat brain vasculature imaged at once in 30 minutes at a 20-micron resolution. To our knowledge, this is one of the only whole rat brain vasculatures that have been imaged at this resolution to date. No, it's ex vivo. I don't know, there's something wrong with that image. But there's room for improvement. And this video shows an imaginary nanorobot and this nanorobot cruises around inside the brain and it spreads out its arms and it says, "Oh, you're neuron 1, you're neuron 2." I'm it's sitting on neuron 3 and it saves this data to DNA ticker tape and it has a power source or whatever and, you know, in some imaginary world this is a make— this is— it's possible to make one of these. If you know a little bit about chemistry or nanoscience, you know, you start breaking this down and you're like, "Whoa, this is really, really not feasible right now." So the way we do it is we have like 20 to 45 nanometer gold nanoparticles and we have a surface functionalization on the nanoparticle and then we can attach ligands to that nanoparticle. But going much beyond that is really difficult. And this is a— this is a volumetric EM stack and in this case what is traced out are all of the synapses. So the synapses in this image are the— these red pancakes which you'll see reconstructed shortly. Why do you get squash and then un-squash? Does that represent— It's just slice by slice. It's just slice by slice. Oh, it's going through a slice. Okay. Okay. Yeah. So this is admittedly very impressive. So the size of this cube is 9 microns by 7 microns by 5 microns. So it's really, really dense and it's really, really high resolution to get, you know, this many synapses in 10 microns, 5 microns. You know, maybe it's 10 to— 12 synapses across in 10 microns. Philosophers and others are curious about, you know, why do you think that you can simulate brains anyway? There are many approaches to, you know, this. One approach is to look at mathematical logic and computability theory, and there's a theorem in computability theory that says that if there's a finite set of bits, it can be simulated. Simulated, period. So obviously the brain is a finite set of bits, therefore it can be simulated. Other people say that, you know, there are quantum effects and quantum effects can be modeled as well. Other people are really anti-brain imaging. They feel like it's your— I don't know. I'm not a philosopher, but maybe there's some part of philosophy where, you know, you should separate the brain from— I don't know— knowledge of the brain from use of the brain or something like that. But I work in a hospital and, you know, if we could image the brain better, we could make people better. Like, there are so many ways. Just figuring out which drugs work for Alzheimer's patients or dementia patients or Parkinson's is really important. Now in the future, in 2017, there are big concerns about AIs. Brain imaging can save us from the AIs, or at least it's worth trying. So I hypothesize that the Russell— so if you go over time and computer performance, computer performance is— maybe in some cases human performance exceeds computer performance, but there are many cases where computer performance exceeds human performance. But with the perfect connectome in the future, human performance will again exceed machine performance. [Speaker:AUDIENCE] Why do you say that? Because I believe it's true. [Speaker:AUDIENCE] You said that on faith. Do you have any logical scientific reason for making that conclusion? [Speaker:MICHAEL STOLER] Can you give me some more background about what you're— [Speaker:AUDIENCE] You're saying that in the future, human performance will always outperform machine performance? Is that what you said? Really what I'm saying is that without knowledge of the connectome, it's hopeless. Oh, but with connectome— Yes. —then it would be possible. Yes. Okay. I promise you, James. Yeah. What's next? I think that there is going to be a lot of work in implanted CPUs with databases of neural codes. Neural codes are ways that different regions of the brain communicate with other brain regions. So neural codes go from your ears to the rest of your brain or your eyes to the rest of your brain. Many other pieces of the brain. Using deep learning to improve interfaces, using electrophysiology arrays to parts of the sensory motor cortex, artificial intelligence and machine learning to trace neurons and axons in the image stack. I think the neuromodem thing is— it's fantastic that DARPA has put this much resources behind it. And one of the key things is microelectronics manufacturers, they need a certain volume in order to make good sense for their business. And I think that they may actually get to that volume if DARPA supports it at this level. And lastly, brain backups, $4,000, which will save us from the AIs. I'd love to thank Neurotech Boston and our sponsors and whoever brought the beer and pizza as well.