AIX Exchange wtih Max Welling, VP Technologies at Qualcomm Technologies Netherlands
My name is Max Welling. I am currently a Vice President of Technologies at Qualcomm and I am also full professor at the University of Amsterdam. So my job in at Qualcomm is to lead the AI strategy for the company and at the University of Amsterdam I lead a research lab, well actually there's a very broad spectrum of topics that we study mostly around deep learning these days but also causality is a big topic reinforcement learning is a big topic and sometimes with applications in for instance medical imaging.
Yeah so what excites me most is to work on new problems new directions in machine learning and connect them to to the real world applications within Qualcomm this is our bread and butter but I like to think of problems and abstract them into sort of mathematical problems or machine learning problems and then work on these abstractive problems so let's say to map a problem for instance in the design of chips back to maybe an optimization problem that can be tackled with reinforcement learning and so there's many more of these differences we are working there also not necessarily me but in Qualcomm, we're also working on source compression and we do that using variational autoencoders with which I've been involved, we work on compilation problems and you can map those for instance to to graph neural nets so I like to work on this interface and but also I like to work on completely novel directions within the field which do not necessarily have immediate applications. For instance the thing that I'm very excited about now has to do with quantum computing, so is there a possibility to apply sort of the mathematics of quantum mechanics and quantum computers to the benefit of deep learning and machine learning. This is of course a widely open field and quantum computers are not yet sort of fully developed and so that this field holds a lot of promise and then also connecting that for instance to the symmetries If we worked a lot on symmetries the world is full of symmetries and this can make this can put structure into your neural network that's very beneficial for learning and in some sense what's happening is that ideas from physics which are often mathematics very deep ideas from physics and mathematics are finding their way slowly into a sort of machine learning and it's kind of interesting in two ways first of all I think these ideas will help make our machine learning algorithms better and more principled in the future but also we might maybe learn something about sort of physics in some in some way at least for me it's interesting that I can learn more about physics by studying these ideas in in the context of deep learning. Yeah I think human-centric AI is actually very important but what sometimes confuses the discussion in my opinion is that AI and you know machine learning are extremely broad topics, so machine learning is a subset of AI so AI you know is a very large container term for all sorts of things and sometimes when some people talk about AI or machine learning, they talk about things where you know explainability and human centric are less applicable and then when other people talk about these concepts they are extremely important and I don't think necessarily that everybody understands that there is this very broad spectrum of machine learning and everybody talks about their own view of the world. So let me give an example to make this clear, so think a bank that needs to decide on who will receive a loan and so this is clearly a decision about people and there it is extremely important that these decisions are made in a responsible way, in a fair way and so there I feel that these you know regulation and you know and this human-centric design is extremely important. However there's also many
applications and actually many of those you find in Qualcomm where it is much less interesting and important to think about these things so for instance imagine I have a black box that will predict and I have no idea how it does it but it will predict for me the best design of a chip, right so I ask it to design the chip with these specs and it will provide for me the perfect layout of all the transistors and all the memory elements in a way that optimizes you know our our metrics, now I can check that it is a really good design and there basically I can just accept that it is a good thing and I can just you know build it and then done done with it. Now that to me is a you know and there's many of these example compilers are very similar, source compression is very similar but they're quite far away from the humans that actually use them and you can very easily check whether they are good designs or not so good designs right and I think so we need to really make this distinction when we talk about you know the importance of responsible AI and we have to sort of first try to classify you know what particular AI we are talking about and then we have to develop methods that will you know for that for every class that will be appropriate for that particular class and so I truly support the moves from the European Union to design this GDPR regulation that basically says that any anytime an AI system decides something you know about humans or for humans then you know this this needs to be explainable and it needs to be you know regulated so that it is for the benefit of those people. So yeah so that's kind of my high level view of this particular topic so myself I am mostly involved in the more technical aspects that are often but somewhat further away from the sort of human decision making about humans and other people who you know who talk about the sort of the importance of responsible and explainable AI, they are more talking about these applications of AI that are more that have a more direct societal impact I would say. Yeah so to me of course, so I think the biggest barriers are things like security, privacy and trust, so as we build as we put lots of these AI devices in our homes right now they start to interact with humans and you know they the more technology we use the more vulnerable it also becomes for hacking from you know from outside from you know from adversarial players let's say and so how do we keep everything secure and safe. How do we make sure that the data that is being collected or potentially being collected by these devices is not shared with parties that we don't want to share our data with, so that that all these things are private by design that we can truly trust that our data stays in you know in our own sort of safe and not moves to the cloud to somebody else's sort of system and sort of how do we certify things like this right, so when these devices become increasingly more complex and potentially even self-learning, how are you going to certify something that is changing and so that we have to I mean it's already hard to certify things which are highly complex but we do manage to do this somehow right, we certify airplanes and we you know and cars which are already extremely complex pieces of engineering but if these things become self-learning that becomes increasingly challenging to do that um and so overcoming that barrier of certifying these things so that we can sort of guarantee that the device is safe and privacy preserving for the people that that will use it and also fair and all these other dimensions which are important. That's a difficult question I am you know for every technology there is positive and negative applications it's always the same and you know it has always been the same, so if we design an axe you can use an axe to cut trees and build houses but you can also use it to you know to kill someone right and so this has always been. So personally
you know I feel a society evolves and becomes more technologically advanced there's no stopping to it there's no stopping to it and we should as people, we should make sure that we put in the right regulation and laws so that the we sort of mostly benefit on the positive side and avoid the negative side but I'm realistic enough to know that, that's not always true so first of all it's being done on purpose which means that you know people make military weapons from AI and it seems unstoppable, it seems very hard, it's like a weapons race and it seems very difficult to somehow stop this or you know I don't you know it's also not my field to so much to understand this but it doesn't seem that very soon here we will be living in utopia in that sense, so I'm realistic and not always all that optimistic about people and politicians and how they can handle these kinds of questions. I do think there's a lot of really good applications of AI in particular inferences in health care but also many other fields that are worth pursuing it but it's yeah so it's it in some sense we built the technology and we need to make sure and explain to the politicians that how this technology works and what is the potential upside and downside of these technologies and then we need to involve other people like you know legal experts, you know humanities experts, philosophers, politicians, regulators to make sure that this new technology that is emerging is used you know for the benefit of humanity rather than against it and sometimes I feel that you know too much of that is put at the at the research right so they often say you know the researcher itself should sort of make these decisions already when they develop the technology but I think that's somewhat unrealistic in the sense that you know it's like asking a mathematician not to work on a fundamental theory of mathematics because in the future you know it might be used for bad as well as for good but it might also be used for bad right, so I also feel I work on the fundamentals of machine learning closer to mathematics than to the actual application and it's when you actually start to work on applications that you're gonna make these may be tough choices, I mean I never work on military applications other people might but I don't and and so that's where you make these sort of choices I think. Yeah so what I find really important again and I'm just going to repeat that is that we this is a truly interdisciplinary effort so it's we should not leave these questions to the technologists and I find this very important because we have a very limited view of the world right, we develop these technologies but we have a very limited view of what technologies do with our society or what they do even you know with our own sense of security you know we will maybe adopt all sorts of technologies and you know they might change our psychology or they might change actually the way we function or the way we think and that's just very hard you know to predict and I think we should really make sure that experts who think about this get involved in order to decide whether certain developments are you know are something that we that we want to happen. So maybe one I can give you one example that sort of has worried me a little bit which is as we make, so we've seen for instance for social media which we use in our phones every day, that it does change people and in fact the things that we read, the things that we get served by social media does actually change the way we think and maybe even the way we vote and sort of I hadn't really thought about that very deeply, I've never worked on this problem but you know I could imagine that I could have worked on fundamental technology for this but you know maybe in you know improving a graph neural net somewhere but I never thought about you know what the impact would be along the lines. I think we need to listen and involve people who do know how these kinds of things will actually impact our lives and to try to help us a little bit figure out you know how to how to sort of make sure that it works in the direction that we, that is good for us and so here's one example that I worried about a little bit and I still worry about which is sort of virtual reality so it's quite clear that virtual reality you know, right now we're looking on our phones all day and you know I can see kids walking and you know they walk and they look at their phone at the same time all the time, so it seems very strange.
Now if we will develop these sort of augmented reality classes then we will be connected to internet into our virtual worlds all the time and so I think they will be adopted very quickly and what then you know and these systems might get so good in capturing our attention that we will sort of develop some kind of addiction to this, right so that we want to be in this virtual world all the time and it may be stronger than ourselves. I mean for some it might already be so some you know there's clearly the phenomenon of addiction to gaming and things like this but this this only gets worse and it doesn't feel that we as humans evolve to create better defenses against it right, so where does that end up at some point in the future. So this is one device that you know we might have around us and that we might just choose to pick up but we need to have people think very hard about how to make sure that doesn't go in a direction that's you know that will hurt a lot of people and might even hurt our society or might hurt our political system or democracy etc. But again I want to emphasize that I think it's really important that we involve other disciplines in thinking about that.
Well I think it's not necessarily that they need to be involved at a particular point but I think you I mean they need to be involved now and all the time they need to be able to monitor what's happening what the technology can do and so they need to be also sort of technically sort of intelligent, you know they need to understand the technology and they need to basically, they need to to think about that and advise about that all the time but maybe you're thinking about a particular technology maybe like a self-driving car or something like that when would you involve people from different disciplines and thinking about the impact of a particular technology and well I would think as early as possible right as soon as you invest in a technology that you know that might have a societal impact, you can start to especially at the regulatory level you can start thinking about it and involve experts and the earlier you have regulation around certain technology the better it is also for the companies that are making that technology right so because they know you know in which direction to steer the ship, if you know if the technology is if the regulation is very clear at that point in time. Ok so a big theme throughout the last years that has been going on in my lab in the University Lab but now also at the Qualcomm is you know something that we call acquivariance or symmetries and so the idea is that the world has certain symmetries like it almost literally like the physical world has symmetries, so for instance if you take electromagnetism something looks like an electric field for one observer that's standing still but then if you move past that electric field it becomes a magnetic field just because you're an observer which has a different speed and so this transformation you know is typically encoded in this kind of what we call covariance or maybe equivariance of the mathematical objects that you're describing. So similarly for for neural net you could have that, you would transform something about the input it could be you rotate the input or translate the input or your shear or scale or whatever yet you think that the prediction of the neural net should be either invariant you know the object identity doesn't change as you rotate the object but also it could be you know equivalent which means if you do some kind of segmentation then the segmentation map should sort of transform or rotate let's say similarly as you rotate the input. Now building this kind of idea into neural nets in various ways has been a big theme, we now also do similar things for instance a reinforcement learning so you can think in reinforcement learning that you know something that if you know you have you're looking at a particular state of the system and there could be a symmetry that if you take this action in this state it should be the same as this other action in this other state and and thinking about these kinds of mappings and these symmetries that might exist in reinforcement learning is another theme. We've worked a lot on graph neural networks so this is if your input space is not a sort of an image or sound one dimensional signal two-dimensional signal but some kind of how would you apply the general ideas of deep learning to these graphs. This is important for instance for you know molecules studying molecules, so we recently looked at how could you take in a molecule, the structure of a molecule as an input and then map it onto some properties of that a particular molecule like you know will it will it cure, you know Covid or something like this right you know of course these properties are more subtle than that. So it becomes
like a prediction about you know what a molecule, how a molecule will act in the body and then you can combine these things you can also think about what if I would now add symmetries to this to this molecule clearly if I rotate the molecule maybe certain properties remain invariant maybe others change instead of building that into the you know into the structure of the graph and we've looked at that and so basically this whole field you know thinking about symmetries, thinking about deep learning on manifolds which I haven't really talked about like curved manifolds, thinking about graphs this called geometric deep learning. So that's been a big theme and what's interesting about that is that you see that the the mathematics that you're using when you think about that is very close to the mathematics that you also use in physics for instance gage field theory is a theory about symmetries in this case local symmetries that you would need if your if your variables live on a sort of curves manifold and that theory as it describes physics also describes this sort of these neural networks these deep neural networks and there's a striking you know resemblance between these two in some sense like a deep neural net is just an iterated map, like you you transform things again and again and again and if you think about an image this looks a lot like a space, the space can be curved and space can have certain symmetries and if you go through the neural net it's like evolve time forward and so in some sense we are building a little mini universe, in some sense in this neural net and you can apply the laws of physics to this kind of mini universe and so we recently extended these ideas to say well what if you know if we are describing sort of a mini universe in physical sense in a neural net could we also apply quantum mechanics, what about quantum mechanics and so this is on my mind right now I work together with people at Qualcomm on this but also people at University of Amsterdam and so the idea is there that quantum mechanics. First of all if you can implement things on a quantum computer you could get you know a lot more powerful neural networks because you know quantum computer is more powerful than a classical computer, that would be one way of looking at it which is already very interesting. The other way the flip side of it would be to say well is quantum mechanics itself perhaps a language that is different than ordinary probability theory because now instead of having just positive probabilities which can only add up we will now also have negative amplitudes and positive and negative amplitudes can actually cancel each other out and this is this interference is very different dynamics, could even classically on a classical computer sort of designing neural networks with the mathematics or the statistics if you want of quantum mechanics could that already give us a new sort of modeling paradigm and that is another sort of direction that we are looking into. Okay so the
the brain basically has two systems one is a somewhat unconscious system that you know you can think of bottom up where you know information hits our senses and sort of gets automatically processed through layers of sort of biological neural nets that you know in our brain to basically see things in a completely unconscious way or hear things right so if I hear you talk I'm not consciously processing you know necessary, you know the translation from the sound that hits my ears to the words that are down here and similarly if I look around me I see objects but it goes instantaneously and completely unconscious. So that is one type of process you can think of fast processing, you can think of it as a shortcut in a way we have learned to do these things effortlessly. Now there's another process which is much more conscious and I think that Daniel Kahneman uses this type 1 and type 2 thinking or something fast and slow thinking, so the slow thinking is much more iterative where you sort of build an argument slowly over time it's reasoning and it takes much more effort for humans to do that and it's also conscious you have to use your consciousness to basically do this reasoning. It's the kind of process if you try to solve you know a math puzzle right you're, you're really thinking you know about you're building steps and you know you're joining sort of reasonings together, you're chaining things together and that's the part that hasn't become conscious and I think that you know when we learn we you know we're putting you know things like if you practice something a lot right if you practice playing the violin a lot then some of this conscious effort turns into more and more unconscious effort and you're starting instead of you know the unconscious things now you know maybe playing some small little you know piece you know becomes much less conscious but then you can start to focus on the bigger sort of holistic things about a musical piece and you know how to how to put you know your interpretation on a musical piece like like a director would do. So okay so that's sort of these two processes,
but so of course when we build neural networks all the way back they are inspired by how the brain works in particular this fast processing sort of these bottom up neural nets but I think more recently we have come to understand that the slow thinking the reasoning. I mean there was actually sort of a separate line of research maybe symbolic AI or logic AI where this is more this reasoning type of methods and people have started to think well maybe you know reasoning and causality and sort of symbols and all those kinds are actually you know, can we merge those with the sort of fast processing, so do we have a fast shortcut sort of recognizer, classifier can we join that with a sort of more slow reasoning machine and that's I think one of the big questions that the field is facing now and how do you how do you sort of emerge these two things together. Now that's one part of the one part of the equation, the one where I believe neuroscience is becoming increasingly relevant is because in one aspect the human brain is a lot more efficient than the hardware that we're building. In fact it's estimated that the human brain is about 100 000 times more power efficient than hardware for the same task and so clearly as our you know AI tasks you know the models become become more and more complex like gpt3 huge you know is a huge model and so it will also cost a significant amount of energy to execute it and so you cannot really execute it for every task because at some point the revenue that you'll get maybe out of your ads is not enough to pay the energy bill or if you would run it on a phone your phone would get way too hot in your pocket right, so there is ceilings where we basically cannot sort of use the latest and the greatest technology because it's basically due to you know energy hungry. So I think there might become an increasing pressure to get you know to make these models and make the hardware more energy efficient and so at Qualcomm we do things like model compression and quantization so that we run these models on much less precise hardware like 8 bit integer rather than 32 point floating point in sort of precision, that would save a huge amount of energy if you do it that way right. Now there's still a lot to win because the brain
has a fairly different compute paradigm and it's far more efficient and people nowadays think that that's because we're in this vonneumann type of compute paradigm where the the memory is separated largely from you know the registers, where the compute happens right where you do the big matrix multiplications and stuff so you have to move the weights of your neural net, you know off chip to the places where you do the computation and you have to do the computation and have you move the results all the way back off chip and this movement is extremely power hungry. So people are starting to think well can we design hardware more like a brain, can we make the memory for instance sit very close and distribute it but close to where the computation happens so we don't have to move this sort of this information back and down all the time and so that kind of developments I think are extremely promising and they can bring down the power sort of consumption of these networks on the set of new type of hardware and then of course we haven't even talked about spiking neural nets but it's yeah it's unclear yet whether spiking buys you something in the hardware the silicon hardware that we build or whether that's something that's mostly useful in the sort of wet wear that's in our brain but these ideas at least are very interesting and we will be looking to the brain and in order to make our AI more energy efficient in the future. Yeah so first I wasn't necessarily talking about edge computing because sort of edge computing is sort of a different dimension which is that some of the computing happens nor on the cloud nor on your device but sort of more on sort of the service which are on the edge of the compute network and then sort of your device you know I guess your device could also thought of as the edge, some of the computation happens at your device and at the sort of smaller sort of edge clouds and the 5g network would sort of move data back and forth between these things. So that's more like distributed learning or federated learning if you want, it's a whole different topic so here I was talking about you know any any chip that needs to do a computation, whether it's in a phone or it's in a cloud or whatever. You know how can we
make that particular chip more power efficient um and so is there an environmental angle, so yes there is but I want to say that I am not too confident that that if we make our chips more energy efficient we will be using less power because I think it's basically we will use as much power as is possible you know and as is economically sort of viable. So in other words if you make things I mean it is what I call the refrigerator effect so when I when i moved from the Netherlands to the U.S to live there for 15 years my refrigerator was a lot bigger but like four times bigger or something and it took a couple of weeks to fill it right and then that one was full again all the time. So this is the same with computing if I give you the same intelligence but you know for four times less power you will use all of that you know you will use four times more intelligence in some sense because your power is basically the budget. So it still is very useful but I think we will have to think about this in a different, way we will have different mechanisms to use less power in a way, so for instance I would believe that in order we would make sure that the polluter pays right in other words if you use power you got to pay for your carbon print your carbon imprint that you're making by doing that computation and I think we should really tax pollution and the use of energy a lot more than we're currently doing because you know now our future generations are paying for the pollution that we cause where we should really be paying for it right now. So to come back to your question of do I think there is an issue with AI using a lot of energy, my answer is yes as AI is increasing you know becoming increasingly important it will use an increasing fraction of the available power and you know as there is a very strong economic driver behind, it won't just stop by telling companies or people hey you're polluting, you should stop that right so that's that's just not gonna happen you should basically tax uh and you know put a sort of a climate tax on the use of energy and I'm a very strong you know proponent of doing just that but it but it is a big issue the amount of energy that's being used you know by by AI certainly in the future.
So definitely to me you know it's easiest to answer the question in terms of an application I've been working on myself that I'm very excited about, so this has to do with the fact that there are nowadays machines that can radiate tumors at the same time as image the body right, so with at the same time as image the body right so with an MRI machine, so this basically means that you get a real you know a a very fast sort of imaging of the body as it moves, let's say as you as you breathe and at the same time there is some kind of radiation beam that sort of destroys the tumor but of course as you're breathing this tumor is moving around and you want your beam to sort of adapt to that movement and nowadays this is not possible, you first make an an image and then you sort of immobilize the patient hoping that the patient doesn't move at all but of course he or she does and then when you radiate you often hit healthy tissue which is very damaging and you don't hit the cancerous tissue which means that it's it's not dead and so I was very excited about you know seeing that application but what it means is that you will have to do far faster MRI reconstruction. So what we need is basically so MRI works as follows so you basically observe sort of you know measurements in the frequency so the fourier domain is what it's called basically certain frequencies and from that you need to reconstruct the image and typically you know you need let's say end of these three a moments that you need to measure but that would take you too long to reconstruct. So now we're going to do with 10 times less of these measurements so we can get 10 times faster imaging yet you still need to be able to reconstruct that image and so one way to do it is to basically learn from many images that have been reconstructed in the past basically look at okay with these types of measurements this is what the image should look like and then this type of measurement this is what should look like. So you basically learn to reconstruct a high resolution image and so there was this particular competition a fast MRI competition which was organized by Facebook and NYU and and sort of we participated in that with the technique which are called neural augmentation where you know you use a neural net but you also use a more sort of a classic signal processing technique and you sort of combine these two together to get the best of both worlds. So we applied you know this kind of technology and and we won in one of the tracks together with the collaboration with Phillips and so you know that I thought was a very neat sort of I would say application of of AI to healthcare and we are still advancing this further to now also think about which of the measurements do you actually want to do not just the number and doing well with a small number of these measurements but actually thinking okay which sequence of measurements should I do. So do the first measurement reconstruct look at the result then
okay what is the next measurement that we need to do right and instead of you sequentially reason your way through you know which measurements you should do, which could give another boost in terms of efficiency of these kind of methods. So in healthcare I think there's enormous opportunity but of course there's many other you know also in mobility there's clearly huge opportunity. You know to to build safer cars and you know and safer ways to move from a to b right now there's about you know what is it one point point you know two million deaths per year and so we should be able to get that down a lot with with applying AI technology. Yeah that's a great question so I think it's important to think about machine learning as a horizontal technology and with this I mean it is not so much you know don't think of it as kind of an another application like autonomous driving or something like or wireless even but think of it as a technology that will permeate everything that we do. So all other technologies and this is what we are currently seeing slowly every field is being basically transformed by AI and machine learning and so it started from you know speech recognition and then image recognition and video and all these kinds of things and then entered the medical field, medical image analysis and so it's much more like the internet or a computer or something like that. So there's no way you can avoid it right, it's like it is a basic technique that will be part of every other application you know that you will build in the future and so where will it make big impact, well it's hard question because it will make an impact everywhere but I can give you a few examples of where I think it's going to make a very interesting impact.
So the first one is perhaps in sort of materials design and drug design. So there I think with machine learning techniques we will be able to model molecules in the way that molecules interact a lot better with machine learning tools and we will be able to not only predict which molecules will have you know the right type of properties that we're looking for either as a drug or some other material but we can even sort of specify the properties and then let the system generate the molecules directly, so I can envision a tool where I could say I want to I want a particular material that does this and this and this is this and then the system will turn on and it will just spit out possible molecules that you can test in the lab and then you know that that have you know potentially these desired properties. So I can see that will that's on a very interesting sort of trajectory that kind of application Similarly I could imagine in an agro culture where you know I can see a self you know I would say a self-regulating greenhouse or something like this right, so in the future we may not want to use you know large swaths of sort of field in order to grow our vegetables and other stuff. We want to maybe like we do in the Netherlands at a very large scale build greenhouses or maybe you know we build huge factories multi-level factories or something like that where exactly the right amount of carbon dioxide and exactly the right climate exactly the right moisture and direct exactly the right light type that is being used every phase of a plant growing is being optimized and self-regulated, so that we will grow you know stuff in these in these factories and you know at the unprecedented efficiency. I think that might
help us sort of you know feed the world and it also it also might be much more independent of the actual environment in which you place these things right it could you know it could be in the Netherlands but it could also be in Africa it's some arbitrary place you could put these things and you could sort of you know you run them almost automatically as like a self-driving car but you know you know the software would sort of self-learn and self-adapt. And then within Qualcomm I think there you can also see a couple of directions where AI is making a profound impact I would say machine learning that's for instance in wireless communications of course Qualcomm is a company that is all about wireless communication. So up to 5g I would say many things were hand designed engineers basically building on decades of knowledge they designed these systems, these modems and all these kinds of things by hand. What you see now is that machine learning is going to optimize these kind of systems, you know much further than humans ever could right so you generate a huge amount of data so think about you want to send data over a noisy channel right now, you don't know precisely what the channel looks like but you can sort of create huge data sets of you know data in you know and noisy data out and then you can learn a system that will learn to decode or what's called demodulate you know the signal on the other end of the channel much better than sort of the hand design systems that we currently have. So
what you see is the machining again asset of this holistic you know technology will help and replace the traditional signal processing tools with these sort of self-learning tools and improve these systems dramatically and so in the next generations of wireless communication systems I think machining will play a very important role and then maybe one other one is perhaps chip design, so in chip design we're also seeing that techniques such as reinforcement learning can sort of where usually teams of hundreds of engineers would sit together and build these chips together over at you know months and months time. It's a very very complex process to you know the engineering process so what's happening now is that bits and pieces of that process are being taken over by machine learning tools and eventually it's not hard to predict that machine learning tools will take over the entire process and since this is such a complex problem they will do a much better job than these human teams do. Now here it's also interesting that humans will have to work together of course, you know with the machine learning algorithm certainly in the beginning you know there's there's you know they become sort of tools that humans will use to do a better design but at some point I think it will be mostly automated and it will it will be clear that the automated procedure is actually better than the procedure where the humans do the work because it's just too complex. So these are just four examples but there's many many many more basically everything you can think of will probably be prone to at some point in time to be sort of revolutionized by machine learning in one way or the other that's my prediction.