Deep learning on graphs: successes, challenges, and next steps | Graph Neural Networks

ladies and gentlemen here we go to the professor at the imperial college in london ladies and gentlemen michael brownstein thank you very much so i hope that that somebody got five points to gryffindor so uh i would like to talk today about uh deep running on graphs and by saying graphs i mean uh networks basically that's the mathematical term for for networks and um i will discuss some of the foundations of these problems uh recent successes future challenges and overall next steps so allow me to start with actually a little bit taking a step back and from a little bit far away uh the concept of inductive bias uh this is a fundamental concept in learning it refers to the set of assumptions that the machine learning system has to make about the problem in order to apply it to previously not seen input and if we take a very simple machine learning system actually some of the earliest neural networks called multi-layer perceptrons we know that they can approximate any continuous function to any desired accuracy so this is a property called universal approximation that was proven uh in the end of the 80s for for these neural networks so it sounds like a good piece of news right because basically we can approximate we can represent anything we want with these neural networks but the moment you come to apply these these models to real data such as images uh we get into a problem because if we look at these images let's say the problem is to classify digits to say that this is digit three even if i move it by one pixel to the right as shown here basically the representation of the uh image that is passed into a neural network is just a vector and you see that it looks completely different so as a result the neural network will need to see a lot of data to be able to tell that all these instances are the same details so to learn environments from the data and because of the curve of dimensionality we don't really have uh anything close in the universe uh to that amount of data that is needed so the right inductive bias was really the crucial thing that made deep learning work in particular in computer vision problems where it really uh provided stellar performance and these are the classical convolutional neural networks where the inductive bias came in the form of what is called translation epi variants or shared local weights so the idea that you can recycle the same weights and apply them at different positions in the image was one of the keys to success of these architectures in computer vision and image analysis problems now let me show you a different problem so what you see here is a molecule so if you're interested this is a molecule of caffeine which others are very familiar with so this is a graph right so the nodes here represent atoms the edges represent uh chemical bonds and let's say that we want to predict some chemical property of this uh molecule let's say what a physicist or chemist called atomization energy the energy that takes to break it apart now how do we represent this molecule we can just take the properties the features of the nodes and put them into a vector but the problem here is that we have many more ways than before to do it actually any permutation of the nodes uh will result in a valid vector and this is really a big problem and we see that molecules are just one example of such graph structure data we see graphs everywhere probably the most prominent example are social networks where the nodes are users and the edges are social relations or interactions between them but we also encounter graphs or networks in biological sciences where we look at interactions between different biomolecules such as proteins in computer graphics in computer vision where we use graphs to represent three-dimensional objects such as meshes and many many more applications so the deep learning on graphs is essentially about finding the right inductive biases for these problems sometimes these are called relational inductive biases and uh there are multiple terms that are used synonymously in this field geometric deep learning is one of them it's because of a paper that we wrote a few years ago where we popularized this term but you will also find terms such as graph representation learning or relational inductive biases and graph neural networks are an implementation of such inductive biases so the history of graph representation learning is actually quite long and we can trace back some of the first papers to the mid-90s but i would say it is probably correct to say that that most of the recent works are from the past uh or most of the interesting or the the critical mass of the works are from the past couple of years and uh in the past years uh graph neural networks have really become one of the hottest topics in machine learning so if you look at the statistics of iclear which is one of the main conferences in this field you see that this is really one of the prominent keywords now what makes graph neural networks so interesting and so similar but at the same time different from the traditional uh uh deep running architectures so let's look at classical cnns so the input in this case is agreed so the image is a function defined on a grid and if we want to do uh operations on the grid like convolutions basically i need to look at each pixel and its neighbors and what conventional neural networks do they aggregate the features or the values in the pixels of of the neighbors by just simply multiplying them by some weights now i can do the same thing in the graph right so the neighbors will be the nodes that are attached by edges to some node so so far it looks exactly the same thing but one thing to know that in agreed when i move to a different position i have a constant number of neighbors on the graph i might have a very different number of labels so if you think of social networks some nodes are extremely popular they might have millions of neighbors some just a few right so the degree of the node might be very different another thing to note is that in a grid i have a fixed ordering of the neighbors i can always talk about a node to the left or a node to the right and this allows me actually to apply the same weight to the same order of the nodes so i can really share the weights in case of graphs the ordering of the nodes can be completely arbitrary so i don't have an ordering of the nodes at least not a canonical one and this actually uh makes a graph neural networks quite different and if you think of a blueprint for how to do convolution like operations on graphs we only have two types of operations we can aggregate information from our neighbors and we can process it in some way and then update the node itself so these are the two operations aggregate and update so aggregate takes the most general form of a function that is applied to features of the node and the neighbor and importantly the function is permutation environment again because i don't have a canonical ordering of the neighbors there are some particular examples important architectures that can be applied for example if we do linear aggregation of the neighbors you can think of it as a node-wise transformation of the node features by some linear operation and a linear diffusion basically somehow mixing the the node features on the graph and these are poppy architectures such as gcn or graph convolutional networks we can also do something a little bit more complex so we can there's still a linear aggregation of the of the node features but the weights now will depend on the feature so this is for example our graph attention networks work and in general it can be some non-linear aggregation in the form of this function h that appears on this slide now let's look more into details of what is similar and what is different in aircraft neural networks if we compare them to traditional deep running pipelines and one thing to notice is that if we look at the historical development of let's say convolutional neural networks actually the early models such as alexnet were relatively shallow in alexandria we have eight layers and relatively large filters so 11 by 11 which is considered large filter but as these methods became more elaborate actually a few years afterwards networks became deeper and with smaller filters as small as three by three pixels and there are several reasons first of all obviously smaller filters are much more computationally efficient and second is that it it appears that in convolutional neural networks you can construct complex features from simple ones a property that is called compositionality so the few the first few layers will will construct simple geometric primitives such as edges corners and then as you go deeper into the convolutional network you will see more complex structures such as parts of the face for example eyes nose and so on emerge now this appears to be not the case in graph neural networks and this wishful thinking that we can uh compose complex structures from simple ones uh doesn't always work so for example it was shown uh recently that graph neural networks at least the message passing type of craft neurological that they showed before are equivalent to what is called the vice versa lemon graphite isomorphism test which is a classical uh graph coloring algorithm that is using graph theory and for example it is known that uh the vice versa element test cannot count triangles in graphs so this is this tells you that even such a simple structure cannot be constructed from uh from primitive ones another thing that you notice that unlike the euclidean the traditional case it is actually difficult to train deep craft neural networks and this is a typical result so this is actually a very recent paper that appeared in iclear this year uh so it requires quite heavy machinery uh special regularization to train a deep graph neural network but if you look at actually the results that you obtain with the deeper network you see that they're inferior to the results that you would obtain with a network with just a couple of layers and there are several reasons to it one of them is what is called feature smoothing it means that features on the nodes tend to collapse to a single point but probably more fundamental phenomenon that the game was described in a very recent paper is uh what is called information bottleneck and it has to do with the fact that uh in some types of crafts the number of neighbors as you go to uh to bigger and bigger neighborhoods tends to grow exponentially so an example of graphs where it happens are small world uh graphs and if you have this exponential growth of neighbors and you also need long range dependencies then you run into a bottleneck meaning that you have a lot of data that you need to squeeze into a single node vector and as a result message passing neural networks are inefficient in propagating information from distant nodes when uh the graph structure is like this and in some cases it might not matter in some other cases like molecular graphs for example long distance relations are really important so that's why depth is rather problematic now in a work that i did with colleagues at twitter we tried to take this idea to the extreme basically we wanted to see uh what can we do with really a very shallow graphical networks where we have just one graph convolutional layer but we allow for multi-hop diffusion approaches so rough analogy in classical cnn would be just one convolution but with big filters and the nice thing here that if we use linear uh diffusion linear filters we can actually pre-compute the diffused features and then it boils down to just multi-layer perceptron which is applied to these pre-diffused features and as a result the neural network is extremely efficient it can scale to extremely large graphs such as those that we see on twitter or facebook with hundreds of millions of nodes and apparently such simple architecture is very efficient so if you compare them to state-of-the-art models it is on par or in some cases actually performs better but it's significantly faster by more than an order of magnitude in uh training and in inference so we call this sign standing for scalable inception like craft neural networks because it reminds the classical google inception modules that were applied about five years ago for obtaining state-of-the-art in image classification uh with convolutional neural networks another thing that i mentioned basically if we already want to uh to enrich our filters uh and i mentioned that that we cannot count structures what we can do we can actually help the graph neural network uh to count substructures by just providing it as some pre-computed feature vectors or a kind of structural node encoding so we can uh pre-count some predefined uh structures of size k this could be for example triangles or clicks or cycles of or paths of different different lines and we provide this as a node descriptor and then we do standard message passing so in this case the precomputation might be expensive in the worst case it costs us n to the power k and here is the number of nodes and k is the size of the substructure but the message passing itself is linear and local and uh what we gain by in this way is that we are strictly more powerful than the message passing neural networks that are equivalent to vice fiber lemon graph isomorphism test and uh another way of looking at it you basically have here a problem-specific inductive bias because in some settings and some uh types of problems in some data sets you know a priori what structures uh what graphs of structures are important so in social networks for example cliques or triangles tend to be important they have certain sociological interpretation if you want in uh for example chemical datasets cycles are important so you see that if we introduce these inductive biases we get uh in some settings uh state of the art performance and the results are especially striking on chemical datasets so here we are predicting chemical properties of molecular graphs and if we uh count cycles of certain size in this case of size six these are very prominent features they are called aromatic greens so if you look at the gain of the molecule of caffeine you see that these cycles uh appear actually with cycle of size six and a cycle of size five and these are really abundant features in organic molecules so if we introduce these inductive bias we see a significant jump in performance of uh such craft neural networks so let me briefly go through what i believe to be the next steps in this field of craft neural networks and one thing that really made a breakthrough in traditional deep learning is this combination or conference of uh three pillars uh data compute and software so in a case of computer vision let's say convolutional neural networks the data was a large scale collections of images multiple millions of annotated images the compute was gpus basically used for general purpose computing and it seems that gpus are very well tailored for convolutional neural networks because they are they support very efficiently these type of single instruction multiple data operations and software was really open source software packages that democratized deep learning software packages such as pytorch or tensorflow so if we look at the situation in graph learning we don't really have standardized benchmarks both the data sets and the tasks that will be similar to imagenet that was the classical benchmark and computer vision on which uh the breakthrough happened uh in about eight years ago uh so the closest that we we have to it is uh open graph benchmark that was launched actually less than a year ago and it's a collection of data sets and uh and tasks uh and actually graphs are graph related problems are much more varied compared to computer vision problems and open benchmark uh open graph benchmark in particular supports three major types of problems these are graph classification node classification and link prediction so this is a large scale benchmark and it's uh always growing so i i hope that in the next few years it will become really the de facto standard for developing and testing graph learning algorithms the second pillar is software libraries and here again the situation is much more optimistic than it used to be a couple of years ago a couple of years ago if you were to implement the latest and greatest graph deep neural network architecture you had to rely on some buggy and unmaintainable code uh from the authors of the paper in the best case nowadays you have a professionally maintained library such as dgl that is supported by amazon or pytorch geometric which are widely used in the research community aspect is efficiency and scalability and um one of the uh issues that precludes or has so far precluded the uh the application of craft neural networks in industrial settings is the size of the graphs that we need to deal with so if you look at problems at theater or facebook sometimes the graphs have hundreds of billions or so or millions or sometimes even billions of nodes and tens of billions of edges and many of the craft neural network architectures that are described in the scientific literature are simply a no go for these settings so it is relatively recently that the community started looking at these problems and there are already several algorithms that try to look at how to scale graph neural networks and one of the issues for example is that when you want to apply uh stochastic optimization techniques that are uh often or i would say in most cases are used for for training uh deep neural networks is the assumption of statistical independence between the samples uh is usually not correct for graphs because if you take craft nodes they are related by ages so there are statistical dependencies so there are many fundamental questions of how to train efficiently large-scale craft neural networks another important question is that graphs that appear in many applications are actually not static they are dynamic so if you look at twitter twitter is evolving all the time so new users join the network uh some users are leaving the network uh people uh tweet something they they retweet reply they like uh so it's a network that is always living and always evolving so the right way of thinking of it is actually as a continuous time dynamic graph so you can think of it as a stream of asynchronous events such as node or age insertion or deletion and there are very few architectures that support these cases so at twitter with uh colleagues emmanuel rossi and and others we have developed an architecture that we call temporal graph networks which addresses these cases so this is generalization of message passing neural networks for settings where uh actually everything is time stamped so we have this asynchronous stream of graph events that form uh the graph in time and uh basically in this case the the the special part of the architecture is the encoder you can train it in unsupervised way basically to predict uh the edges in the future or some points of time and then the decoder can be task specific so you create node representation and then for example you can use it for different tasks such as classifying the nodes predicting for example if a user is doing something wrong predicting edges so you can predict engagements and this is basically bread and butter for recommender systems another thing that is important i already mentioned it briefly is the use of high order structures so so far graph neural networks have really been about nodes and edges and we know that in many complex networks especially biological networks or social networks we we have complex high order structures such as triangles and maybe more complex graph graphic graphics and graph motifs so uh one may be a simple and to some extent naive way would be to use them by just counting them but you can think of maybe more complex uh message passing architectures where the aggregation uh happens at these uh higher order or higher dimensional structures so you can think of uh topological constructions such as simplicial complexes and this has been largely unexplored or unknown in this field of deep learning of graphs another important topic is actually the very assumption that we are given an input graph in many cases is incorrect so this is a typical situation for biological problems where we have let's say interactions between molecules such as proteins so this graph we have about 20 000 proteins in in our human body we don't really know how they interact with each other so some proteins are considered important so they're probably researched to death and we know how they interact but some others are maybe less prominent so we don't know many of the edges in this graph so this graph is given only maybe partially or maybe it is noisy so the bottom line that that in the best case we have maybe just an approximate version of the graph so it is possible to design uh graph neural networks that build the graph uh as part of the learning process and uh together with my collaborators from mit we actually did some of the first works in this domain we call these dynamic graph cnns where the graph is constructed on the fly using let's say k nearest neighbors and uh it can also be updated between layers because the graph is really dependent so depending on the downstream task you can build the best graph that that suits your task and this really brings to this important question is whether the computational graph that is used for message passing in graphical networks has to be the same as the input graph if it is provided at all and in many cases we see that it doesn't need to be the case so we really want to separate or decouple these uh these two constructions so in dynamic graph cnns the graph is actually constructed to represent some local structure geometric structure of of the the input which is a point cloud we applied uh these methods first so the problems in computer vision and graphics where we were dealing with 3d comp 3d point clouds but they can really apply it to any kind of data so in recent work with collaborators from munich we applied it to healthcare electronic records where we have uh data on patients and the data can come in the form of phenotype features such as age sex and so on and also maybe imaging features that you can get from uh basically from medical devices so we want to build the graph that represents some kind of relations or similarity of the patients but we don't know a priori how to build it so we learn it for the given task and we see that uh using this construction we can significantly improve uh the performance of important problems such as disease classification we can predict for example whether a patient has alzheimer's based on the brain imaging features now i should say that in retrospective this kind of methods are related to what was called manifold learning or non-linear dimensionality reduction which was a proper class of machine learning algorithms about 20 years ago which looked at high dimensional data with the premise that the data might live in a very high dimensional space but it actually uh it has low uh intrinsic dimension and the convenient metaphor for it is that the data comes from low dimensional manifold so the the usual metaphor a way of visualizing it in few dimensions is the swiss roll surface so the the typical way that these algorithms worked were to start with first building a representation of these low low-dimensional structure usually in the form of a nearest neighbor graph and once you construct the graph you create a low-dimensional representation of the data for example in popular algorithms such as isomap this is done by using multi-dimensional scaling to preserve the geodesic distances on this graph and once you get this low dimensional flattened representation of the data you apply some machine learning algorithm that in many cases was simple clustering so now with modern graph neural network pipelines you can bring all these uh steps into a single pipeline so you can build the graph and perform machine learning on this graph in the same pipeline in an end-to-end way so maybe with some stretch i i could call this manifold learning 2.0 so another question that is important and i briefly mentioned it uh is actually the theoretical understanding of rough neural networks and one of the big questions is actually the expressive power how powerful are graph neural networks so i mentioned already some results the equivalence between message passing and graphics over physical tests but these are actually very recent results from last year so this is still an open question that is uh interesting and important and last uh but obviously not least the killer apps so when i started working already more than five years ago on deep learning on graph somehow i was hoping that in a matter of a couple of years we'll see a breakthrough revolution similar to what happened in computer vision with graph neural networks and i cannot say that it really happened uh but i think that it might happen and uh we really see graph neural networks being applied to a lot of different problems and really crafts are very abstract and universal models for systems of relations and interactions that can be applied to modeling in many different fields of science you can see graph neural networks applied to recommender systems so particle physics i had a work with uh you know with high energy physicists doing uh neutrino detection in the ice cube observatory uh we had a startup that was acquired by twitter last year where we applied craft neural networks for problems of fake news detection and there are many many other examples so there are really already first success stories of this of these methods if you ask me what would be one field on which i am willing to bet where these methods would probably make a breakthrough i would say these are applications in medicine and biology and you can really apply graphs there from all scales from nano to macro from modeling molecules to modeling interactions between molecules and basically interactions between patients what they mentioned already with patient networks and some of the results are really extremely promising and i would say almost dramatic so for molecules modeling molecules and predicting their properties is really the holy grail of drug design and drug development because if we look at the space the search space the the space of possible synthesizable molecules it is extremely large we probably have more synthesizable molecules than the number of atoms in the universe i think the estimates are above 10 to the power 60. on the other hand what we can test in in experimentally in the wet lab or even in the clinic is way way smaller we can probably test maybe a few hundreds or a few thousands of substances so we somehow need to breach this gap and this can be done computationally what is called computational funnel so basically you need to break through this gap and uh you can do of course uh super computing simulations uh you can do some some cheap and dirty approximations such as density functional theory that is uh used in computational chemistry graph neural networks allow you to do uh more or less the same accuracy but significantly faster orders of magnitude faster and they're already interesting because uh there was a paper from a group at mit earlier this year where graph neural networks were used in a pipeline that allowed to discover a new class of antibiotics it was published in cell in february i had a collaboration with uh colleagues from epfl uh that are experts in protein design so we developed graph neural networks or geometric depending methods for designing proteins from scratch and proteins are potentially interesting classes of drugs biological drugs that for example can be used for cancer immunotherapy so here we're designing proteins with certain functionality such as binding to a cancer target so that was also a paper that appeared on the cover of nature methods in february uh taking uh uh uh taking one step back basically looking at a more abstract level of interactions between molecules these can also be modeled as drugs as as graphs and here we can describe uh drug to target interactions as graphs and the interactions between the targets which usually are proteins as as graphs so uh we can talk about problems such as drug repositioning or predicting uh side effects of multiple drugs or predicting synergies of multiple drugs uh using graph neural networks and uh just a few days ago there was an announcement uh i'm taking part of a collaboration with mila in canada and a pharmaceutical startup called relation therapeutic where we are trying to use these methods to develop uh drug repositioning for covet 19 which is obviously a very important and noble cause i will finish with a last example uh basically talking about drug repositioning we don't really need to look at drugs themselves we can look at drug-like molecules and many such molecules are contained in food so you might know that foods especially from the plant kingdom are rich in molecules that come from the same classes as many oncological drugs so we can use graphene law networks to predict this dry drug likeness and we can find what are the the foods that are most rich in these molecules and then we can use them maybe as if not the therapy maybe to prevent uh cancer and other diseases and we have a collaboration with the molecular chef who is actually using these uh discoveries the ingredients that we are finding with craft machine learning to design very appealing and very tasty dishes so i guess this yummy note i would like to finish and probably a good question here is uh let's meet again in probably three years and see whether these promises have materialized so thank you very much thank you so much michael braunstein for your uh for your talk here um while i'll see if there is any questions coming into the chat uh if not ladies and gentlemen then send your questions to the chat and i will pass them on to michael uh waiting for those first questions to come in michael um what is keeping you awake at night scientifically speaking please um if you look at your uh at your research well i i should say that i am involved in multiple projects so uh for for graph uh neural networks uh these are obviously problems that we are working on at twitter related to social sciences things like platform health or uh detection of misinformation especially in light of the coming u.s elections uh biological problems so kovid 19 is really an important thing and i would say this will be really the the call to arms and maybe the the the test of how uh uh of the promises that uh artificial intelligence and machine learning uh uh was maybe a little bit overhyped whether it really can deliver and i believe that that if we are successful that will be really uh really the case and finally uh i'm involved in a completely crazy moonshot project which is uh using uh machine learning uh techniques that are developed for natural language processing to try to study and understand the communication of sprung whales wow this is cool okay um there are some questions coming in so i will start with one of our regular bringer of questions in dots and he says will it be beneficial and if so how much if the initial features number of triangles for instance of graphs are only approximated by sampling before inputs to a network the complexity will be reduced from n hyphen k to s hyphen k brackets s is oh brackets one close brackets i hope i make sense to you obviously it makes a lot of sense so i think uh this is an excellent question so uh i should say that the the end uh end to the k is really the the worst case complexity so for many structures that that appear in important problems such as clicks or cycles there are better complexity algorithms there are stochastic approximate algorithms with actually bounds on performance that are significantly cheaper so the computation of structures can be much better talking about the the sub sampling it is actually very interesting so sub-assembling methods for graphs are used for uh mainly for scalability purposes and uh one of the the early scalable graph architectures called sage it's called sampling and aggregate uh the acronym sage uh used uh uh use some graph sampling for scalability but i believe i have the feeling that it actually has to do with the this bottleneck phenomenon that i mentioned that graph sampling allows to reduce or break the bottleneck and as a result it's not only it improves the the scalability but also improves the performance of crafty knowledge okay thank you um before moving to i think the last question for now just a quick remark by kingsley he says okay three years awesome presentation so that one is in the pocket um then i go to to neil and he says in 3d vision of the competing types of state-of-the-art approaches which include cnn based point based and graph based what are the main advantages of graph based methods over the others do you consider the dependency of having to pre-compute a mesh rather than operating on raw point clouds to be a limiting factor yeah good question so for let's start from images right so if we are talking about two-dimensional images uh probably convolutional neural networks because they are so efficient it is probably hard to compete with them unless you have a very good reason to use a graph you can of course represent an image as a graph maybe as graph of super pixels you probably don't want to do it now i should say that graphs are being used in computer vision as a side information so at least in two ways one is as a kind of a scene graph where you can represent relations between between things so architecture such as capsule networks can be reinterpreted in the in the optics of of craft neural networks second thing is uh fu shot learning because you can use graphs uh to represent the structure of your data so your uh basically you you apply you would apply a normal convolutional neural network to your image but you will also be uh be exploiting the structure of the data space we actually used uh this hybrid architecture of cnns and graph neural networks to for example increase the robustness of of conversational neural networks go to adversarial perturbations now talking about uh 3d data measures versus point clouds so of course meshes are much nicer for uh for uh for a geometer uh because they they keep the the geometric structures you can basically you have the underlying the underlying uh discretized surface whether they are necessary probably not if your point cloud is sufficiently dense and sufficiently nice uh it probably has all the information that uh that is needed for the problem now uh the graphs uh are a kind of uh intermediate way between measures which have a lot of uh structure and basically you can think of them as local uh euclidean representations of discrete manifolds versus graphs that are just uh sets of points sorry the the point clouds which are sets of points so the graphs have uh this local representation of structure which is maybe a little bit coarse and primitive but uh it is better than just one clouds okay thank you um meanwhile the compliments keep coming in in that says a great presentation um and there was one more question uh before we go um and i know it's hard for a science scientist to uh say yes to me asking you can you come up with a short answer but i would like to uh to try and have this one answer the question by hassan it says graph notes can represent multiple information sources that have different features with having nodes in the graph that have different features influence the dynamic graph cnns i'm not sure that they completely understood the question but of course you can you can represent uh so uh in in practice these uh the graphs that what we call graphs uh what i call graphs are in fact uh what is more correctly called multi graphs so in multi graphs you might have multiple edges between nodes so this is probably this is probably what was asked in this question so uh for example when you have two users and uh on social network they interact between each other multiple times so you have multiple edges so uh because they also have time stamps so they are somehow spaced in time so these are uh in reality multi-groups yeah so the the short answer the temporal graph neural network framework can address these settings thank you um so we settle for the yes thanks um okay um ladies and gentlemen we are going to uh another question in the fastest finger first competition before we say goodbye to michael and i would like you to answer your answer to this question as quickly as possible and the question is michael showed us somewhere in his talk a mess of an animal can you remember which animal it was please enter the animal in the chat on the right side of your screen the first one gets five points with the right answer number two gets four number three gets three four gets still two points and number five gets one point a mesh of what animal was part of the presentation of michael and i am waiting for the right answers to come in and while i do that i will tell you michael that another few compliments came in that was a really interesting talk thank you said someone uh great talk says but here then thank you impressive talk says menwa thank you great talk says so all compliments here and now waiting for the right answer to the question yes there are answers in and i will scroll down and see who was number one in that says a mouse that was i'm sorry that's not correct in that the first one i will count this as right was ulrich who says a bunny i will count that as correct so five points to ulrich four points to nigel who says rabbits three points to burn at also rabbits two points to juke rabbits and the final point with also rabbit goes to menoir so those were the points for this round in the five fastest finger first competition ladies and gentlemen um can i have a great thank you and i know you can't applaud and so i will do it for you thank you so much for your talk michael bronstein thank you

Channel: SAIConference

Views: 19,140

Rating: 5 out of 5

Keywords: Deep learning, Graph neural networks, GNN, machine learning, Michael Bronstein, Graphs, applications of deep learning, Deep learning on graphs

Id: PLGcx65MhCc

Length: 43min 25sec (2605 seconds)

Published: Fri Sep 04 2020