let's talk about the brain always a good idea um in more particular terms weather and how brain science as we think could still help us maybe with arriving at better machine learning it better artificial intelligence let's see so when we look at images like this we are pretty much aware that we are using neurons nerve cells to process this image detect that it's a face maybe even you know memorize or remember that this may be a composer and depending on your interests maybe even identify the name of that composer now the same is very well possible with the noisy image no problems and as you all know today is machine learning trying to solve these problems is also using neurons in some way or other in what we call in their own it but obviously neurons implemented in silicon and you know pretty in efficiently implemented usually and then network architectures like this as you all know are still at the basis of what is called today is deeply I think it is a large
fraction of it and definitely when it comes to image analysis now why is that called neuro and what is brain science to do with it as well as you know there are two things that came from biology as inspirations one nonlinearities of the neurons light of understood to be very important to be able to approximate almost any function and then secondly and this is after all of the feed-forward connectivity this came from as an inspiration from the visual system mostly the idea or the analysis of the first Anatomy if you want of the last century and that retina projects to then the later visual centers by the thalamus to cortex and then it's delivered further and further into cortex in the human and other mammalian brains now this is of course true but it's not the whole picture and this is what I want to talk about today now artificial intelligence today it is to a very large fractions I just said is deep learning meaning many layers meaning not so many layers of a
feed-forward perceptron well nothing really fundamentally different compared to what was known on the post and the sixties of the last and definitely you could say whatever we're doing right now in artificial intelligence corresponds to old neuroscience in the sense that I just explained now obviously I don't want to ignore there are more things to it learning rules backdrop as you know has you know some claim but I don't think any two plausible biological compliment my view is I don't think backprop has a has a natural correspondence others see differently and then obviously their attempts to go to more recurrent networks and you all know famous LCM learning rule that allows us to deal with our names at all it's the one learning rule that we know of again very difficult to see biological parallels there now why am I talking about this you could just say I come on you know you don't have to build a brain to build intelligent artificial systems right after all we don't have to build birds to fly and
that's an important thing here to see I'm not claiming that it's necessary to look at brains to improve artificial intelligence but what I am saying is maybe it is worth looking at it again and I'll tell you a little bit about what I mean again so luckily our planes are not as floppy as a bird's wings but anyway mostly at least these days now what are the key differences between biological and artificial intelligence today the first ones very obvious everybody talks about its energy efficiency and actually we haven't been talking about it so much because currently we can still fuel the the compute sensors that are needed for current AI but it's actually truly limiting if you think about much more powerful computations it's enormous amount of energy and it's nothing compared to the you know beautiful sandwiches you just had maybe already downstairs right I mean energy consumption is just know it's all of magnitudes different but I think there's another difference beyond learning rules and so on that is even more important and
that's also not so often talked about at least or if I said I think it's very important this is label efficiency so when we learn today's artificial intelligence we are essentially abusing the fact that we have too many labels and we're really not making use of labels very well what do I mean well if you think about it you know how does a today's a I system learn to call a tree a tree what we do is we present images of trees not just one that's another one that's another one there may be some with more than one tree and then we do that over and over again and usually actually with exactly the same images importantly always with a label attached to it and the worst of it is this billions of times as you all know I mean when you're trying to see an end that's what you do maybe a little bit of variation added to it but you're presenting this poetry and end with this kind of input and all of
you know that is not how we learn if you've ever you know thought about how you learn to call a tree a tree I think my estimate is you heard the word tree probably 50 to 100 times before you were able to tell a tree from from my car let's say at age 1.5 years the reason being you've seen millions of trees but you've used your labels much more efficiently and label efficiency is really I think the biggest problem we have or inefficiency in today's AI and this something where the brain is much better now I'm a neuroscientist so what I should be telling you about now and I will do this now is what we do in neuroscience and why I think what we do could help could help you here the contractors could help understanding better what's going on in brains and why in the end learning is as efficient as it is as one of the aspects now what we do is call connectomics so we care about the networks in the
brain and we measure them because they are beautiful they're also difficult to measure we have nerve cells these beautiful structures if you go at high resolution into your brain and this is just one of them with a cell body and the in white you see the dendrites as it's called the input side of the nerve cell and then in a second you will see in blue the axon the output cable of just one nerve cell it's not even all of it it's just the local fraction of it all this branching one nerve cell okay so this is really what we're dealing with and if you want to know how many of these you have the answer is roughly 86 billion okay so why are these structures so beautifully complicated the reason is that each nerve cell this is shown in the upper left corner each nerve cell in our brains connects to about 1,000 other nerve cells and it if people ask me what's the most number about brains it's not so much the number
of nerve cells I think it is the connectivity of single nerve cells so the main number of our brains is that we have an order of a connectivity degree degree on the order of 1,000 each cortical nerve cell connect synaptically and specifically to about 1,000 other nerve cells this is not paralleled as far as we know by any other system today especially because it's long range connectivity mostly it's not local when we talk about the nearest neighbor of a nerve cell in terms of communication this nearest neighbor can be at the other side of the brain and that is really a very very specific setup of a complexity that we haven't embraced at all I would say and we're trying to do the end by measuring this we don't know how these networks are structured at least not in high resolution and high detail now why are nerve cells so difficult to map the reason is they are morphologically very difficult they are very thin the cables I talked about the axons are easily 50
or 100 nanometers in diameter that is a thousand times thinner than our hairs and the problem is we don't have just one of these cables we have again the cables of 86 billion nerve cells in this piece of space that we call our brain pretty limited I mean it can't grow much more and the result of this I will show in a second now we are using techniques involving electron microscopes to analyze these structures again we essentially take a biopsy of brain tissue imbed a small piece of brain and plastic standards with heavy metals and the like and then we do experiments like this where we use scanning electron microscopes to image the surface of tissue blocks then have a diamond knife scrapes off the tissue we've just imaged and then take another image and do that over thousands tens of thousands hundreds of thousands ideally sections so we get to substantial distances in three dimensions because nerve cells are not in a plane like in the image there and 3d and 3d high resolution
imaging is seriously difficult since about 15 years based on techniques I should say developed in Germany and Max Planck society not in the u.s. not in China we are able to do this and here are examples of tissue data so this is a piece of retina actually so the first visual processing step in mammalian nervous systems and which contains already very interesting filter computations of the visual signal roughly 30 parallel filters of the visual world are transmitted to our brains so it's not a single pixelated version it's roughly 30 as far as we know today that is sent from the retina and this is computed in the piece of retina I just showed and this was our first target and this was a study where we for the first time were then able to get the mammalian connectome as we call it a smaller piece of enough to shoot from the mammalian brain analyze here there are cell bodies so the the somata the cell bodies of nerve cells involved in the visual system photoreceptors
and other the types of nerve cells and then we reconstructed them density meaning all of them about 1,000 URLs on this tissue block and all of their arborizations and all of their synaptic contact points you can already appreciate what it means to talk about dense tissue highly packed extremely densely packed nerve tissue and the cables within here's one of those nerve cells still intelligible still you know visible but then if you superimpose all of the bout again 1000 and nerve cells from this piece of tissue then you're starting to see what the problem is so this is the density of circuitry used in a small piece of in this case mouse retina to analyze the visual world just at the first stage so this is again the local and sometimes not so local processing used to compute these about 30 to 40 parallel theater representations of the visual scene that are sent to fairness and in higher order visual centers in the brain of a memo and more and more people realize that the same
or roughly the same may be actually happening in higher mammals like like parameters us which was for a few decades I think lost but I think this is coming back that the same kind of processing is happening now retina okay so this was a feat and we were then able to compute the connectivity matrices this is already a little bit abstracted here not as a single-cell resolution but using cell types that was a major step for us at least because now we were looking at communication maps of one tiny but at least finally mammalian circuit well I'll be seeing here this is an adjacency matrix you know just a neighborhood graft represented in terms of neighborhood entries in the neighborhood matrix and essentially indicates who talks to whom very much like we could create it you know connectivity matrix of this room here currently nobody is talking but but but if you were we would easily see whose friends and and who's not so this is what we got and then we you know started
to analyze directional circuits computation of Direction selectivity and so on and this is this is published now the update I want to give here and that is important to me in this context is also to realize that this is not how the brain looks this picture I have to say when I entered neuroscience was plausible to me I thought okay fine yes nerve cells you know a little bit of interstellar space or whatever ok the firing I was aware that firing was not as literally meant as it sounded but yes so there is activity and you know the put these cells kept you know like this but this is absolutely not what we have in our brains the actual data looks like this so this is the actual segmentation of dense neuronal Network again from the retina at this point but we're starting to get similar things from cortex highly densely packed neuro pull all these cables are on the scale of 100 nanometers or so in diameter we are flying through a tunnel which
is a black vessel that we took out ok so when you really dive into the neuro pool as we call it the nerve tissue this is what you get and I think this is really important to realize when we are talking about building computational systems again I'm not saying we have to do this but I'm well I'm trying to remind you offer put out here is this is how evolution solved it again there could be better solutions again we don't need to be birds to fly but but we should be you know maybe appreciating the level of complexity that's used by nature to systems that can compute the inefficient energy efficient and able efficient way that we talked about okay just to give it a little bit of an update also a little bit of technological excitement this is basic research obviously but this is happening again not somewhere in the world this is happening in the middle of Europe the fastest edicue microscopes are those scanning with 6090 in the future hopefully hundreds of
beams built by Zeiss a German company three of these microscopes are in the Max Planck Society in Germany they're five worldwide so we have three or five this is good we get very large data sets not just of mouse also of human I'll talk about that happily some next time so but this is happening data analysis is a big challenge we use a little bit unconventional tools to combine again AI and human interaction to analyze our data that's a separate story but I want to come back a little bit to just give it a small glimpse on what we mean by analyzing these networks and these are early days so this is just happening right now so when you think about brains and so natural computers if you want and our man-made computers then architect on the architecture side it's very clear what the differences are we talked about you know the cable stuff but even more abstractly speaking the key thing about our current computers is that memory and processing are separated as you
know and this is absolutely not how the brain does it the nervous system combines storage and processing we have absolutely no evidence for separation of the two so that means it's really a different computational architecture and that also means by the way that looking at the structure can be much more meaningful than looking at the structure of a computer chip obviously if you don't slice up a computer chip all you learn about is the microprocessor architecture that's not going to tell you what program is running that's not going to tell you what software algorithms are implemented and being you know used on this on this processor so but but when we look at brains looking at the networks means reading out potentially both the algorithms and the storage and this is why we think this is so exciting there are many more differences actually brains as slow as you know if you want the clock as Mondamin micro to millisecond scale but highly recurrent our networks and operate in 3d but for the last couple
of minutes I would like to focus on another aspect of brain architecture and that is the existence of inhibitory nerve cells so that's another difference right I mean currently at least standard AI is not particularly looking at any concept of inhibition that is not explicitly so we have no activity that can be high or low but negative activity is not really something that that is implemented I mean not saying exclusively but in most things that we know of so but brains have that so we all have roughly 20% twenty to twenty-five percent of our nerve cells of fourth to a fifth are inhibitory nerve cells what do we mean by it when they are active they dampen the activity of the postsynaptic neuron they you know transmit gaba as it's called it's a particular transmitter it's kind of a substance that it leads to chloride in facts and so on and so the point is when you activate an inhibitory neuron one of these you know 20 25 percent cells that we own the next
cells will decrease their activity now obviously that's kind of plausible that you need something like that to prevent your brain from going crazy from becoming epileptic fine that's definitely one part of the story but maybe there's much more and I think it's fair to say we don't know yet fully what this means and I just want to show you a little bit how we start thinking about that so this feature of the brain which is called inhibition means their population of cells as I already said have a negative impact on the postsynaptic activity and they're obviously 80 percent of the cells are doing what's shown here with us a positive sign so exciting each other that's what we know from artificial neural networks that's the same story there but this inhibitor in ourselves are special and when we look at their activity obviously one reason you could do that in addition to preventing epilepsy is that you simply overall reduce the energy load on your system right so if you are able to encode negative
activity by activity of other nerve cells and essentially you're reducing your overall activity by certain factor depending on how active these are so this could contribute to energy efficiency I don't think it's the key deal but as one aspect but circuit wise what what is it so in a recent study where we analyzed the circuitry and mediate and ride the cortex a part of the brain occupied with spatial navigation and so on so what I'm trying to do here it's run by me entering the cortex partly at least and they're unrelated to the spatial activity we think we found a very fundamental aspect of these circuits that made us wonder if we found that synapse along the cables in our brains in this case mouse brain I should say I extremely precisely placed along the cables much more than we thought we knew for some insects and other species that this could happen but for a memo this was extremely exciting and the key thing was that these synapses were used to always create what
we call a triadic circuit mode and this was much more preserved than we thought in the sense that whenever an excitatory nerve cell contacted another excitatory nerve cell always there was an inhibitory neuron in the loop not sometimes but always an always the exact same connection what's happening so this cell is excited the inhibitory cell which would make sure to inhibit the very same cell also contacted by the single prison optic cell so what we call it called cellular feed-forward inhibition because it operates at the level of each and every nerve cell extremely specific and now we found that this is actually a closed gate we found that the biophysical properties of sign up sorting thick cables fast cables and a few other tricks make the circuit in a way that when you have a very active population there is no transmission of signal to them to this excitatory population because the inter neurons kick in now that made us wonder why would you do that I mean obviously you can say I find I
can build a big gate easily but but this is this is very interesting because now you can think about how how can you unblock such a gate right do you have a switch signal there you have expectations that come in and this is really when we're thinking about whether this is essentially a very basic hypothesis circuit where you encode by wiring all the things that could happen but you're not activating it you would go crazy if you're thinking about everything that could happen all the time and only when information comes in like I'm now heading in that direction or more abstract context like whatever it's not so sunny day or something like that that makes me want to my umbrella that then this contextual information unlocks the most plausible hypothesis laid down in this highly parallel circuit now that's going very far well I'm just talking about a hypotheses I'm not talking about you know something that we've proven to be there but what we've proven to be there are the circuits for it and
this is really a first step very very early at first step I think in that journey to use circuits in the brain to understand how its operating and this could lead to novel computing right again this is a proposal I laid down a few years ago we could maybe learn from these recurrent networks in the brain and maybe help build more exciting technology or also in terms of artificial intelligence and again we don't even know currently on what scale of recurrency our my main cortices really operate this is something we're right now trying to find out in local circuit reconstructions from Mouse cortex is it you know is there anything like LCM probably not exactly but on what level of recurrency are we operating what are other specific circuits that we find in each and every animal and obviously also in human brains and with this I want to conclude with think that you know neuroscience inspired machine learning 70 years ago that's clear and then for a long time the two fields a little
bit you know separated partly because computer science wasn't really able to compute with these ideas and partly because we were occupied with other things and then you know that's two decades or so there was an inverse thing where machine learning became so powerful that it helped us a lot with data analysis we're using a lot of machine learning for our image analysis it's nice twist here but now we think is the time where we at least want to try by mapping these circuits which we'll do anyway but we'll ask questions about them like the ones I've shown you again early steps and that may lead to ideas about how the brain is able to be so energy efficient and so label efficient and this is what's happening right now and again is happening in Europe it's happening in Germany it's happening here so it's something that obviously it could inspire totally new technologies and I hope may contribute to some transformative developments in this field this is my summary thank you very much for your
attention we have a little boy actually I have a little bit of time left would like to ask a question in SS first yes thank you so much um being a neuroscientist myself I have a I have a question about well I think usually we only incorporate the electrical signals but do you see a chance of incorporating the chemical signals as well well I mean as you know the transmission from one neuron to another is always electrochemical right so you have the electrical part which is fast and goes to the synapse but then the ten nanometers or 20 nanometers from the axonal terminal to the postsynaptic neuron are chemicals diffusion as you know so that says every signal transmissions electrochemical because then postsynaptically it's again electrical so in that sense it's all electro chemical electrical if you're talking about things like modulation so state dependent as we think you know transmission off of substances that change the entire state of the brain is very simplified and that's obviously also something that is very relevant for
instance for reinforcement learning again another topic inspired by for in science where a dopamine is the key player so there that's obviously also very highly researched area reward encoding okay so no other question oh I'm not sure he raised first give it to him not throwing say 50 fish hi yeah just the clock speed hundred milliseconds for the brain biology does it in many seconds and then our seconds for silicon is it I mean can we if we sped the biological up to nanoseconds would we lose something I'm not sure that's true I'm not sure it's possible at least with a where the current device I mean even at the level of proteins and and the need to transmit an iron through a channel which is at the core of our electrical activity if the eight nanometer membrane across which you want to conduct a charged ion so I don't think that even that speed would allow for nanosecond so I'm not it's more practical thing but what I'm definitely interested in it's a solution
which is not just about making it faster it's a solution where slow kind of slow is okay because maybe I mean people talk about then obviously the parallelity of the system the sparse coding right it's also not highly active didn't mention it but only ten percent roughly of our cells are active at any you know second so we opened one will it help again I mean could you do it I don't think you could do it if you did would it help I mean okay there's another interesting thing to think about if I may what sets actually the speed that we need to operate it and I think gravity is key right if you think about I mean the object we care about fall from trees usually or maybe mountains but essentially that determines a time scale on which we need to it to be able to react and if you take that time scale then milliseconds is fine I mean we react within the fastest is 100 milliseconds of brain if cortex is involved
it's usually fine for the most falling apples okay is there another question yes you have to stand up and raise your voice because we are running out of battery yeah I mean yeah so I mean the first key difference is that in in the brain and our brains not all of animals but in mammals there are specific neurons specific nerve cells only doing inhibition so that was as if you were separating those with a negative weight from those was positive right so that's already different from what you usually have when you just get a negative way interestingly some species don't do that flies don't do that the worms don't do that so it seems to be something it's called Dale's Dale's law it seems to be an achievement or development of mammals and some other species again I don't think we understand why you would do that I think that's the fair answer thank you very much just a another question a last question from the audience then I'm oh they're sorry what car and
like because it seems that you really focus on the hardware not so much only like software challenges with and and what current like hardware initiatives are you most excited about because there are these like neuro inspired chips by Intel and all of these initiatives so first I should say I would like to emphasize again that I try to point out that the software hardware distinction is true only for our computers not for our brains so in that sense when you say I'm focused on hardware I want to hope you don't mean it dismissively in the sense like who cares about hardware fine just make that point again so yes and the brain hardware is what we call hard and software in computers and it's party I mean there's a little bit more but party now to your to your question about current row morphic if you want approaches I think they're you know they are interesting and they are trying to represent what we currently know about the brain what I try to show to
you is that we are only now measuring these circuits and all the current amorphic initiatives have been based on knowledge that was available let's say 20 years ago good learning rules SCDP and so on plasma learning rules but the network architecture totally not matching what we see so I think that is an update needed and it might be that we need ago need to go to 3d implementations you know if you want to build a network of other 1000 you make it by cables but non by non wireless transmission when wire base transmission you may need to go 3d you know some people are thinking about that obviously