I have an idea for start-up: sculpture style transfer, we learn from great piece of sculpture-art and we transfer it to new sculptures based on some simple sketch. Or we can design new buildings based on existing ones.
Holy moly Siraj, I think a year ago I commented on a video of yours asking whethere deep learning can be used to classify 3d objects. After being out of the loop of watching your videos for a short while, today I literally thought to myself 'I wonder what Siraj has released on youtube recently' and THIS VIDEO just came out 20 hours ago!! Thank you so much Siraj, im so excited to learn more about this field \^o^/
Perhaps I missunderstood, but Euklidian space is not just 1 or 2D but also includes higher dimensions R^n. You can describe a sphere in Euklidian space like r^2 = x^2 + y^2 + z^2. Euklidian space is based on the notion of uncurved space.
Siraj you have the only youtube channel I ALWAYS watch the outro of! &also I love how you sum up the most important info before the end & have it displayed in text form!
THIS is pretty descriptive of the principals I have been thinking of with the Aura Operating Zeitgeist I have been working on. The majority of the work has been setting up data collection, sorting, etc. Nearly time to implement and run some algorithms in VR. NICE CLIP SIRAJ.
Simply excellent. I really appreciate your videos. Just a tiny nitpick, unless you are near a black hole, objects in 3d space are also Euclidean. But I understand what you are saying!
@Hidan Gravity, especially strong gravity, warps in a non-Euclidean way. That is what makes General Relativity so hard. (Technically all gravity, but it is small enough you can mostly ignore it on Earth) Euclidean 3D, just means that the same rules (about angles, parallel lines, etc) apply in 3D as they do in 2D. Of course, this is somewhat missing the point of the video!
Idea to share: semantic topology decision networks. A key point to create a general artificial intelligence is an ability to create dynamic relations between abstract features(which are already achieved by DNNs). Lets say you have multiple interlinked neural network models. The first module is an input module which interprets the system state to a high level feature representation. The second module is a deep q table which evaluates the different abstract states. There can be a supervising model built in as well. In the initial training phase the idea is that a high level feature representation can also be understood by a human. (For example you have x,y coordinates on N maps, and the maps are linked)
They might benefit from other graph-cut algorithms applying Dijkstra's weighted or unweighted graph path simplification theory, as an alternative to Relu.
Have been looking forwards a incubating idea of mine where I abstract public opinions to form an "thought-heavy" neural network model. Something just clicked when you iterated the phrases "deep learning" for "non-euclidean" spaces. Damn, this is huge for me. Will spend some years exploring the thought. It's always exciting in the beginning. But when Siraj says: "Did this give you an idea for a startup?", I'm like "You're goddamn right boy." How did he know?!
So cool using non euclidean data for classification will make transfer learning much easier too as training a model to determine a non euclidian class say for audio data can also be fine-tuned to determine a plane I mean every data can be converted to a non euclidean format and a ML model can be made on it. I think this is a really cool concept and will help the AI community if researched further
I came here to hear all the different types of words that are used in these things, and I'll come back again when I have a proper level of understanding what those words mean. I'll be back
What are the inputs to a GNN? And also, assuming you input a feature vector per node and an adjacency matrix to describe the edges, aren't you hard limited by the size of the graph? How do you account for different graphs in size? I wonder if how much the curse of dimensionality is a problem here...
Very very interesting video! Keep up Siraj! I have one question regarding this. How do these methods differ from more traditional Probabilistic Graphical Modeling with missing data? Most of these problems seem manageable through some typical Belief/Markov metwork afaik. For example, given a graphical independence structure it is trivial (at least in theory maybe not so for large N) to predict MLE/MAP of missing verticies. Is this approach different due to efficiency or is there more theoretical poterntial?
This would be a great alternative to CNN, RNN and LSTM text classifiers that use word2vec. Word2vec is a three dimensions represents of words, so these new geometric models could be the next step in natural language processing.
If we use GCN on a regular image (i think there are ways to convert it to a graph), what accuracy can we achive for classification problems (lets say on CIFAR-10) ?
I don't have the numbers off the top of my head but regular CNNs tend to outperform some of the popular GCNs on these kinds of tasks. In general GCNs have not been as successful yet. There are no "final versions" of GCNs like we know for CNNs. There are different flavors of GCNs that are used for different kinds of problems. There's a lot more work to be done before they get to CNN levels of success.
Hey Siraj, Could you do a video explaining spiking neural networks and the use of them, as the search results for it on RUclips are not very satisfactory. Keep it up with your videos, awesome stuff!
*Thanks* for the awesome video, *you rock* man! The hype around the simple convolutional networks and image analysis in general is so high, that it's actually not that easy to find meaningful info materials about graphs and how to approach them. I have 2 questions/additions to the topic from the video: 1) Could you elaborate more on why exactly a 3D object is a non Euclidian object? I believe that it still can be considered just as a set of numerical values in 3 dimensions, just as like we process images as a certain amount of points in 2 dimensions, right? 2) I think it is also important to mention that in order to be able to analyze a graph you have to plot it with help of *force-based layout* algorithms. In this case after you visualize it properly (even in 2D I guess) - we can just analyze the picture to a usual well known 2D convolutional network, can't we? Actually the force layout stuff for the graphs is crucial, and maybe even deserves a separate video, since there is a lot of graph data to be analyzed out there - e.g. social networks, blockchain transactions, etc. Thanks in advance!
Im REALLY interested in 3d object recognition and pose estimation, preferably from 3d Mesh than point cloud but point cloud works as well..... Could you do a video of implementation of GCNs? That would be AWESOME! THANKS for always keeping us up to date with the latest in AI!!
Hi Siraj. Do you fancy doing a vid on the very latest tensorflowjs upgrade (0.14.1) ? Tensorflowjs is revolutionary. It doesn't need any client install or configuration, so anyone, anywhere with a browser and an internet connection can run it. It's also possible to cluster web services for a simple distributed learning system. And of course, it will run on the GPU. The only thing letting it down are the libraries and functionality : this is why the feature upgrades are so important.
Nice, I was just thinking of away to optimse djikistra algo alongside travelling salesman problem algorithm. Ps still trying to even gauge the practicality. thanks for this intro
Manifold learning is unsupervised and most of the algorithms involve no training. Oftentimes the term is synonymous with nonlinear dimension reduction and KPCA is usually considered an example. They get their name from the assumption that high dimensional data usually lies on or near a low dimensional manifold. Also, manifolds can have any dimension so this applies far beyond 3D objects.
Hi Siraj. Thanks. I thought 3d was old news, and that you just had mentioned or covered in other videos. Well, you know, vector calculus uses even more than 3 or 4 dimensions. of course the matrix and else gets a lot more complex. See you.
I was drunk and watching pbs then came here. I google'd for solutions to an n-dimensional problem in gauge theory, and ended up watching this somehow. From the comments, sounds like people working on similar types of problems...I guess, first one to the future wins. Hope it's a good one
Hi Siraj, I don´t know if it is a silly question but it came to my mind: Is it possible to combine Geometric Deep Learning with Hinton´s capsule nets? Because the vectors in each capsule can be three- or n-dimensional...
7:23 Does anyone know how these feature representation of graph nodes are defnined? What determines what the amount of feature values one graph node is represented as? Is this really only dependent on the neural network and initialized randomly?
Why are point clouds not considered Euclidean Data? Based on the wiki en.wikipedia.org/wiki/Euclidean_space, point cloud should just be 3D euclidean data.
Has there been any attempt to do pointcloud segmentation applying this kind of network layers in a similar manner to the U-Net for 2D? Anyone up for the challenge?
I'm so excited to watch this video oh man this is the final puzzle piece. THANK you Siraj. Thank you. I just couldn't figure out the best way to explain it but yeah basically that's how I've been doing things just ....manually. By hand. I've been manually doing exactly this...
Do not start with a book, about python or anything. Learn it like a language. We learn quickly from sentences that are usable but first step is vocabulary and grammar. But you actually start from a narrow understanding of grammar and as you use the language more and more you automatically develop hidden understanding of grammar, given you are quick to correct mistakes. You can't really speak if you have to first think in one language and then think about translation rules. Rarely, can a person state the rules of their mother tongue but every single person knows how to follow them. This is intuitive understanding. And, before starting to speak, you should first listen and have a motivating reason to speak that language. Now back to programming, if you already know at least one programing language, you can start by coding the same programs, simplest to most complex, in the target language. However, if this is you first language, even then there is a language which you already know, that is pseudo code. Start from simplest task, telling your computer to say hello. First go with something that already works, so you know your target. Ask someone else to setup the environment for programming and use smallest program as your starting point. Just try editing an already working code, a little at a time, to see how it behaves. Once you have an idea of the behavior, try rewriting the same code on your own and you will begin to realise the involved steps. Then, slowly move on, adding no more than one functionality at a time, so you keep learning as behaviour becomes complex. Remember, think of it like learning to speak to your computer. Once you get to the point where you can write a program that is a complex and big one, in your opinion, it is time for the next phase. By now you have some experience and this is the time to go deeper, because now everything will feel relevant. Now is the time to learn from a book, which goes into theory and standard and efficient practices and real scenarios behind development of a language. Even this time, its going to be a brief phase. Because, although a language can be understood better by learning about it but it can only be spoken better by trying to speak it everyday.
Maybe because it's late, I did not grasp the non-euclidean convolution stuff along with receptive field. Maybe you can point to some other refs, or make a new vid ? ;-)
Glad to see GCNs (Graph Convolutional Networks) from our 2016 paper (arxiv.org/abs/1609.02907) featured and explained in such a great way. Source for some of the material used in the video: tkipf.github.io/graph-convolutional-networks/ and the accompanying code release: github.com/tkipf/gcn PS: Most of us in the field don’t call it Geometric Deep Learning, but rather ‘Relational Representation Learning’ or ‘Deep Learning on Graphs’. See latest conference workshop: r2learning.github.io :)
Geometric deep learning is soooo powerful. Can't wait to start my latest project with it.
I have an idea for start-up: sculpture style transfer, we learn from great piece of sculpture-art and we transfer it to new sculptures based on some simple sketch. Or we can design new buildings based on existing ones.
Holy moly Siraj, I think a year ago I commented on a video of yours asking whethere deep learning can be used to classify 3d objects.
After being out of the loop of watching your videos for a short while, today I literally thought to myself 'I wonder what Siraj has released on youtube recently' and THIS VIDEO just came out 20 hours ago!!
Thank you so much Siraj, im so excited to learn more about this field \^o^/
Perhaps I missunderstood, but Euklidian space is not just 1 or 2D but also includes higher dimensions R^n. You can describe a sphere in Euklidian space like r^2 = x^2 + y^2 + z^2. Euklidian space is based on the notion of uncurved space.
Another good reason to be alive in the field of AI.
You have a lot to learn which means you never get bored.
Awesome Siraj 👍❤️
Siraj you have the only youtube channel I ALWAYS watch the outro of! &also I love how you sum up the most important info before the end & have it displayed in text form!
Bro!!! I just love the way you talk, as well as you body gestures. Thank you so much for the content.
THIS is pretty descriptive of the principals I have been thinking of with the Aura Operating Zeitgeist I have been working on. The majority of the work has been setting up data collection, sorting, etc. Nearly time to implement and run some algorithms in VR. NICE CLIP SIRAJ.
Simply excellent. I really appreciate your videos.
Just a tiny nitpick, unless you are near a black hole, objects in 3d space are also Euclidean. But I understand what you are saying!
@Hidan Gravity, especially strong gravity, warps in a non-Euclidean way. That is what makes General Relativity so hard. (Technically all gravity, but it is small enough you can mostly ignore it on Earth)
Euclidean 3D, just means that the same rules (about angles, parallel lines, etc) apply in 3D as they do in 2D.
Of course, this is somewhat missing the point of the video!
I watch and read a lot on AI and ML but this channel has to be one of the best right now!
Been working on a project using pointnet for 4 months. Really awesome video and a really fast changing field.
Idea to share: semantic topology decision networks. A key point to create a general artificial intelligence is an ability to create dynamic relations between abstract features(which are already achieved by DNNs). Lets say you have multiple interlinked neural network models. The first module is an input module which interprets the system state to a high level feature representation. The second module is a deep q table which evaluates the different abstract states. There can be a supervising model built in as well. In the initial training phase the idea is that a high level feature representation can also be understood by a human. (For example you have x,y coordinates on N maps, and the maps are linked)
Hello this is Siraj ,👍🏻Thumb up
They might benefit from other graph-cut algorithms applying Dijkstra's weighted or unweighted graph path simplification theory, as an alternative to Relu.
Have been looking forwards a incubating idea of mine where I abstract public opinions to form an "thought-heavy" neural network model. Something just clicked when you iterated the phrases "deep learning" for "non-euclidean" spaces. Damn, this is huge for me. Will spend some years exploring the thought. It's always exciting in the beginning. But when Siraj says: "Did this give you an idea for a startup?", I'm like "You're goddamn right boy." How did he know?!
Yes!! You gave me a great idea for a startup using point cloud data!! omg thank you!
Have u created it..?
@@ardan7779 is coming together!!
I'm so glad you decided to do a video GCNs. I asked you on LinkedIn about this and you delivered. Really Cool!!!
Great, great, great video, and the image from Thomas Kipf of GCN! Simply love it!
Damn, Finally!! I have been waiting for this since so long. Thank you :)
you are the best.... and you help me understand many concept in AI and ML.. thanks for that..
So cool using non euclidean data for classification will make transfer learning much easier too as training a model to determine a non euclidian class say for audio data can also be fine-tuned to determine a plane I mean every data can be converted to a non euclidean format and a ML model can be made on it. I think this is a really cool concept and will help the AI community if researched further
I came here to hear all the different types of words that are used in these things, and I'll come back again when I have a proper level of understanding what those words mean. I'll be back
This is too awesome. Please elaborate on GCNs in the future when you feel that there's more to be discussed
Wow. Love the level of detail here. Thanks again!
-Ev
What are the inputs to a GNN? And also, assuming you input a feature vector per node and an adjacency matrix to describe the edges, aren't you hard limited by the size of the graph? How do you account for different graphs in size? I wonder if how much the curse of dimensionality is a problem here...
Very very interesting video! Keep up Siraj! I have one question regarding this. How do these methods differ from more traditional Probabilistic Graphical Modeling with missing data? Most of these problems seem manageable through some typical Belief/Markov metwork afaik. For example, given a graphical independence structure it is trivial (at least in theory maybe not so for large N) to predict MLE/MAP of missing verticies. Is this approach different due to efficiency or is there more theoretical poterntial?
This would be a great alternative to CNN, RNN and LSTM text classifiers that use word2vec. Word2vec is a three dimensions represents of words, so these new geometric models could be the next step in natural language processing.
If we use GCN on a regular image (i think there are ways to convert it to a graph), what accuracy can we achive for classification problems (lets say on CIFAR-10) ?
I don't have the numbers off the top of my head but regular CNNs tend to outperform some of the popular GCNs on these kinds of tasks. In general GCNs have not been as successful yet. There are no "final versions" of GCNs like we know for CNNs. There are different flavors of GCNs that are used for different kinds of problems. There's a lot more work to be done before they get to CNN levels of success.
Hey Siraj,
Could you do a video explaining spiking neural networks and the use of them, as the search results for it on RUclips are not very satisfactory.
Keep it up with your videos, awesome stuff!
Finally... 3D deep learning
*Thanks* for the awesome video, *you rock* man! The hype around the simple convolutional networks and image analysis in general is so high, that it's actually not that easy to find meaningful info materials about graphs and how to approach them. I have 2 questions/additions to the topic from the video:
1) Could you elaborate more on why exactly a 3D object is a non Euclidian object? I believe that it still can be considered just as a set of numerical values in 3 dimensions, just as like we process images as a certain amount of points in 2 dimensions, right?
2) I think it is also important to mention that in order to be able to analyze a graph you have to plot it with help of *force-based layout* algorithms. In this case after you visualize it properly (even in 2D I guess) - we can just analyze the picture to a usual well known 2D convolutional network, can't we? Actually the force layout stuff for the graphs is crucial, and maybe even deserves a separate video, since there is a lot of graph data to be analyzed out there - e.g. social networks, blockchain transactions, etc.
Thanks in advance!
Really great video, Thanks man for all the effort you did to simplify this for us.
Im REALLY interested in 3d object recognition and pose estimation, preferably from 3d Mesh than point cloud but point cloud works as well..... Could you do a video of implementation of GCNs? That would be AWESOME!
THANKS for always keeping us up to date with the latest in AI!!
Thanks Siraj 😎
Hi Siraj. Do you fancy doing a vid on the very latest tensorflowjs upgrade (0.14.1) ?
Tensorflowjs is revolutionary.
It doesn't need any client install or configuration, so anyone, anywhere with a browser and an internet connection can run it. It's also possible to cluster web services for a simple distributed learning system. And of course, it will run on the GPU.
The only thing letting it down are the libraries and functionality : this is why the feature upgrades are so important.
This does bring ideas for a start up, do you have any links or books that relate to ethics and tech?
Nice, I was just thinking of away to optimse djikistra algo alongside travelling salesman problem algorithm. Ps still trying to even gauge the practicality. thanks for this intro
*Madrid fans disliking the video*
Chiringuito de jugones dataset :)
Ha visto tu cuenta y eres de Perú man ta bien que estés aquí dame tu whatsapp 51947559008 whatsapp
Only 27
thank you... very useful video
great video! where can i find the gcn example code in the video
Wonder if geometric deep learning could be used to track an object, like a hand, in 3d space, producing it's xy and z coordinates?
Any suggestions for a master thesis topic?
Finally 🤗
What’s the difference with manifold learning?
manifold == 3d object
Manifold learning is unsupervised and most of the algorithms involve no training. Oftentimes the term is synonymous with nonlinear dimension reduction and KPCA is usually considered an example.
They get their name from the assumption that high dimensional data usually lies on or near a low dimensional manifold.
Also, manifolds can have any dimension so this applies far beyond 3D objects.
Hi Siraj. Thanks. I thought 3d was old news, and that you just had mentioned or covered in other videos.
Well, you know, vector calculus uses even more than 3 or 4 dimensions.
of course the matrix and else gets a lot more complex. See you.
man this guy is awesome
First one to like and view and comment 😀😀
Sir plz one video on ROS, how to start it and what are the prerequisites for ros
Good info!
Thought of using this in space time object identification and manipulation.😎
Hello Siraj, can we apply Graph CNN to time series data? specifically for traffic flow ?
I was drunk and watching pbs then came here. I google'd for solutions to an n-dimensional problem in gauge theory, and ended up watching this somehow. From the comments, sounds like people working on similar types of problems...I guess, first one to the future wins. Hope it's a good one
great video!
You do know that DLis about to go meta, when: NN are graphs and causality graphs are a 'thing'
Does the geometric learning help us in the long way to replicate animal (humans eventually) brains?
Hi Siraj,
I don´t know if it is a silly question but it came to my mind: Is it possible to combine Geometric Deep Learning with Hinton´s capsule nets? Because the vectors in each capsule can be three- or n-dimensional...
I might have missed something, but didn’t you only list 4 Axioms for Euclidian Geometry, while you were talking about 5?
One question
I want to develop a textile pattern designer bot
So will it fall in this geometric deep learning ?
at 5:22 "Do a singe tree and get Paid" not a thing right?
7:23 Does anyone know how these feature representation of graph nodes are defnined? What determines what the amount of feature values one graph node is represented as? Is this really only dependent on the neural network and initialized randomly?
How did the persons from ted exposition generative design " this achieve such chassis desing
Hi Siraj,
do you could make a video over Tube Neuronal Networks?
best regards.
Translate landscape data to a form of 3D Morse code to aid the blind!
Now that is a revolutionary thought!
Please make a video on pointnet. Thanks siraj
Generate pickup lines?? Snd ful prjkt lnk plz!
Why are point clouds not considered Euclidean Data? Based on the wiki en.wikipedia.org/wiki/Euclidean_space, point cloud should just be 3D euclidean data.
So what's the difference between geometric deep learning and 3D CNN?
Thank you so much
Amazing
I'm confused about why you can't apply a normal convolution to a point cloud.
Because you need to define a neighbourhood for a convolution. This is difficult for an unordered and unstructured pointcloud.
What sorces do u use for finding these latest topics
Reddit, twitter, friends, facebook(?),Arxiv/Arxiv sanity , Slack, RUclips comments
Every time I watch one of your videos I just realize how much I suck at math
Guess it's humanities for me
Sir how you edit your video ???
mean which software u use for it??
Video starts at 5:25
Can you make video on competative programming
Has there been any attempt to do pointcloud segmentation applying this kind of network layers in a similar manner to the U-Net for 2D? Anyone up for the challenge?
Can you make a video on startup
He already has made one.
Yes he's made one
Oh damn you guys finally caught up lmaoooooo this was basically what I've been doing.
I'm so excited to watch this video oh man this is the final puzzle piece. THANK you Siraj. Thank you. I just couldn't figure out the best way to explain it but yeah basically that's how I've been doing things just ....manually. By hand. I've been manually doing exactly this...
you are gaining weight siraj ... stop eating multi-dimensional non-euclidean doughnuts ...
Will do thanks for keeping me in check
He's gaining momentum! Get your facts straight
chicken salads, siraj
don't you mean n-dimensional tourus's? No need to over complicate it man :-p
Wow 😍
When are you coming to India 😀
കൊള്ളാം മച്ചാനെ
What does it mean
@@subschallenge-nh4xp Nice video man
Good
Yes white hair with black background is like 3d
💡🚀 🙏
Nice video but Karate is cooler than football ;)
I am a student of btech cse 1st year which book is best for python
Do not start with a book, about python or anything. Learn it like a language. We learn quickly from sentences that are usable but first step is vocabulary and grammar. But you actually start from a narrow understanding of grammar and as you use the language more and more you automatically develop hidden understanding of grammar, given you are quick to correct mistakes. You can't really speak if you have to first think in one language and then think about translation rules. Rarely, can a person state the rules of their mother tongue but every single person knows how to follow them. This is intuitive understanding. And, before starting to speak, you should first listen and have a motivating reason to speak that language.
Now back to programming, if you already know at least one programing language, you can start by coding the same programs, simplest to most complex, in the target language. However, if this is you first language, even then there is a language which you already know, that is pseudo code. Start from simplest task, telling your computer to say hello. First go with something that already works, so you know your target. Ask someone else to setup the environment for programming and use smallest program as your starting point. Just try editing an already working code, a little at a time, to see how it behaves. Once you have an idea of the behavior, try rewriting the same code on your own and you will begin to realise the involved steps. Then, slowly move on, adding no more than one functionality at a time, so you keep learning as behaviour becomes complex. Remember, think of it like learning to speak to your computer.
Once you get to the point where you can write a program that is a complex and big one, in your opinion, it is time for the next phase. By now you have some experience and this is the time to go deeper, because now everything will feel relevant. Now is the time to learn from a book, which goes into theory and standard and efficient practices and real scenarios behind development of a language. Even this time, its going to be a brief phase. Because, although a language can be understood better by learning about it but it can only be spoken better by trying to speak it everyday.
Maybe because it's late, I did not grasp the non-euclidean convolution stuff along with receptive field. Maybe you can point to some other refs, or make a new vid ? ;-)
Mother... Real Madrid FTW!!
Just a little correction, Real Madrid is the best team.
Awesome
GENERATING PICKUP LINES?
Cool....
the newoton method is very old and its deep learning in a sens
First ♥️😛
Glad to see GCNs (Graph Convolutional Networks) from our 2016 paper (arxiv.org/abs/1609.02907) featured and explained in such a great way.
Source for some of the material used in the video: tkipf.github.io/graph-convolutional-networks/ and the accompanying code release: github.com/tkipf/gcn
PS: Most of us in the field don’t call it Geometric Deep Learning, but rather ‘Relational Representation Learning’ or ‘Deep Learning on Graphs’. See latest conference workshop: r2learning.github.io :)
Cheers
smart dude.
Go Barcelona!
“Pun intended lol” -robot voice.