Atomic Brain? - Computerphile

Key Takeaways

hi sean really great to be back on computer file we are in the lab again i think a few or maybe about a year ago you were down here i can't quite recall if the instrument was running at the time but what we're actually going to talk about today is based very much on this paper i wish i could say it's or a paper it's not our paper it's by a group led by this guy alex could turian sorry alex if i've got your name for not pronounce your name incorrectly um at least you're not stuck with moriarty for a surname interconnected single atoms could make a quantum brain and we're going to talk about this there's an awful lot of physics we need to get through and sean hangs his head back uh in pain the camera's not drooping oh sorry we'll try and condense our physics down last video we did shows on neuromorphic computing which is about trying to take the type of computing that happens in here rather than the sort of on neumann architecture which is what happens in there and use that to do processing because there are certain things our brain does incredibly well that silicon doesn't do so well and vice versa when this particular title popped up in my physics world institute of physics feed it obviously repeat my interest directly they've basically made neurons out of single atoms which is really really quite remarkable not only have they done that they've got single atom neurons and then they've also controlled the connections between those atoms and the way those atoms talk to each other so they've also built artificial synapses again effectively using single atoms so this is really quite remarkable we've got neuromorphic computing pushed all the way down to the atomic level and that work is is based around an instrument called the scanning tunneling microscope now we've talked a little bit about this before i bored brady harren's ears off over on 60 symbols about scanning tunneling microscopes this is a scanning tunneling microscope all the stainless steel surrounding it is because it's in an ultra high vacuum we're manipulating individual atoms we want to have this controlled environment as possible we don't want lots of contamination we're actually working with silicon and if you were to prepare silicon in air it will oxidize in a fraction of a fraction of a fraction of a second so we have everything in ultra high vacuum and a pressure comparable to that you have on the moon and actually in real time what you're seeing here these are atomic rows of silicon these little holes are where atoms are are missing and we along with a number of other groups have been embedding machine learning approaches like convolutional neural nets and there have been lots and lots of videos on computer file about convolutional neural nets we've also actually had a number of very helpful conversations with mike pound who i know many of you will be familiar with on taking machine learning ideas embedding them into the scanning probe microscope and instead of us sitting there laboriously preparing that probe we're training the instrument using reinforcement learning and strategies along those lines to actually get the tip in a state where it gets the best possible image without us having to sit there and babysit it for hours days weeks months on end in a nutshell sharp probe atomically sharp you bring it in close to a surface and when i say close i mean within an atomic diameter or so there is a means of doing that 60 symbols you move it back and forth and you measure some type of interaction be that a current and the tunneling in scanning tunneling microscope comes from the fact that it's a quantum mechanical effect called tunneling that happens on the atomic scale [Applause] or a force and we're now at the point where you can measure the force and not just measure but exploit the force between single atoms and that's what they've done here they've imaged and controlled right down to the single atom level after that long introduction we're going to talk about this particular paper and what i'm going to do is try and summarize it down to what the key ideas are so the paper itself is here nature nanotechnology an atomic boltzmann machine capable of self adaptation so what's a boltzmann machine well boltzmann is an incredibly important physicist mostly because he thought about the statistics of physics he thought about matter in terms of its component atoms and the dizzying bewildering array of gazillions upon gazillions of atoms you you know try to count up the molecules in this room so that makes it a statistical problem in terms of how those atoms interact because you know trying to follow the trajectory of an individual atom or an individual molecule if you have to do that for 10 to the 23 some of you might know where that 10 to the 23 number comes from in terms of a mole of material if you have to do it for that number of molle it's an impossible task so you think about it in terms of statistics and that was boltzmann's genius was to think about the atomic draw comparisons between i'm not sure comparisons raw parallels between statistics the atomic nature of matter and also how energy is distributed and that's absolutely key his ideas were soundly rejected by some of the key figures in physics at the time particularly boltzmann was one of the first to really accept the reality as would wear of atoms we've been vindicated long before those images appeared in terms of the evidence for atoms but you know he would be blown away we're seeing individual atoms on the screen we're seeing rows of atoms um there so where does the boltzmann machine's idea come from the boltzmann machine idea is really clever and physicists really like it because it involves taking the ideas from physics in terms of the distribution of energy and feeding that into a machine learning process in physics you have what we'd call an energy landscape the simplest possible energy landscape is think of two hills with a valley in between we have a ball it's at the top of the hill we let go of that ball what happens it falls down the hill to get to its lowest energy state and actually there are very very few physics questions that cannot be answered with well the system wants to reach its minimum energy state if you want to use that at dinner party at some point or in the pub please feel free to do that pretty well any question can be answered with that of course the devil is always in the details but it's the system wants to get to its lowest energy or what we call the ground state you can think of much more complicated landscapes energy landscapes with lots and lots of hills and valleys and it doesn't just have to be gravitational potential energy if we're talking about atoms and molecules for example it'll be a type of energy that describes how electrons behave which is very much not gravitational energy and so what we do with a boltzmann machine is we take that idea of minimizing the energy and we couple it across so we draw a parallel with trying to find the optimal solution to a problem so we have a particular energy cost or an energy gain associated with following some trajectory or other and what we're searching for in both cases is the minimum energy state or the optimal solution and that's that's where the boltzmann machine idea comes from we're borrowing from ideas from physics and importing those into machine learning what we want to do is we want to model the brain what do physicists do physicists are very reductionist and for example this is a big wet squishy thing with lots of very strange weird stuff happening so what we want to do is reduce down so the question to ask yourself is what do we have in the brain we have neurons and we have synapses could we have a neuron which is basically one atom so we want something whose state we can change so how do you change the state of an atom though is an atom's an atom an atom is an atom but what's inside that atom inside are the electrons so you can change for example a really simple ways you could have a potassium atom or you could have a potassium ion where you've added or subtracted a charge or you can rearrange how the electrons are organized in that atom you can change the electronic configuration leave it neutral but change the electronic configuration that's a brilliant question shown because that's exactly what they've done in this paper they've changed the state of the atom by changing the electronic configuration now so what we have is the possibility of having a one or a zero but also we can flicker back and forth between those two we can go once it runs stay in one state go to the zero state for a little while go to the one states do the zero state these are individual cobalt atoms we don't need to go into the gory detail of this what we have here are individual cobalt atoms what we're imaging here with our scanning probe microscope is actually effectively how the electrons are distributed so although it's an atom and you might think of an atom as a little billiard ball actually what the stm is sensitive to is the configuration of the electrons can you give us a two sentence what's in an atom for people who haven't like me have forgotten so the the traditional model of the atom two sentences is showing you asking me for two sentences on the atom physicists like to reduce yeah um the typical model of the atom which is called the bohr model which is unbelievably wrong but it's helpfully wrong is that you've got the nucleus at the center of the neutrons on the protons and you've got the electrons spinning around outside it's not how it works quantum mechanically we don't have these well-defined orbits but we certainly have that nucleus at the center and then the best way of thinking about it is you've got some distribution of that electronic charge right which is effectively like a cloud around that nucleus what they're doing here by is by using the tip and changing the voltage on the tip they're changing that that cloud with that electron cloud and effectively what you're imaging there is that electron cloud so they basically have a two-state system they've left to its own devices it flips back and forth between zero and one right so now with about equal probability our energy landscape in this case is like that right so we've basically got the zero state and the one state in terms of what the atom is and the configuration of the atom is bouncing back and forth between those two states which is why you're getting high current low current high current low counter high current logan and to use the physicist term it's stochastic or it's random it's bouncing back and forth between those two states and so in one state let's see call this the zero state it looks like this in the other state it looks like this and that's because the configuration the electrons in the atom is changing so this is our basic model of a neuron right but that's one so when you bring two of those atoms together like this what they then find is that you have a larger number of different current values or basic unit our basic neuron is a single cobalt atom and that cobalt atom can be in either a zero or a one state we're measuring and monitoring what state that is in by looking at how current flows through it from the tip that's above it the zero state we get a low current in the one state we get a high count it doesn't matter which way around we label those now when you bring those you bring two of those atoms together they start to to affect each other they affect the probabilities and now you've got instead of having just two choices you've got two atoms you've got four choices they can both be in the zero state this one can be in the zero state this can one can be in the one state this can be in the one state this can be in the zeros here or it can both be in the one state you've got two bits right so effectively you've got two bits and indeed that's what they see they now see that the current has four distinct levels or a number of distinct levels yeah four distinct levels which is what they've plotted out then you can add more and you can i build up the complexity of your system but physicists love this idea because it links to boltzmann and b it's as reductionist as you get it's we've broken it down to single atoms instead all that wet floppy stuff with gazillions of bloody atoms widely flopping around in molecules we've got an atom on a surface in ultra high vacuum that we can probe right down to you know the electronic configuration that's the type of thing that physicists love and when that connects to computing it's really neat they've even gone a step further so that's our neurons what they've then done is they've basically built in artificial synapses which is nothing more than an additional atom and they bring that additional atom in and that controls the linkages the weighting of the link between those those two neurons and then they've even gone on further and they've actually got this thing to learn to some extent in that it can detect or remember the voltage that's been applied to it by the tip because that changes the weightings of the connections between the neurons it changes the synapses so that the system can learn about its previous experiences and so on we can't do a lot of interesting things with this image after just one set of convolutions but we're getting there so this one is starting to be transformed some of them are noisier than others purity atoms in that sample what happens in the real sample is you have a voltage and that generates an electric