42 thoughts on “Geometric Deep Learning”

  1. 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

  2. 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!

  3. Why are point clouds not considered Euclidean Data? Based on the wiki https://en.wikipedia.org/wiki/Euclidean_space, point cloud should just be 3D euclidean data.

  4. 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!!

  5. 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 ? 😉

  6. 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. 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.

  8. How did the persons from ted exposition generative design " this achieve such chassis desing

  9. 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?

  10. Hey Siraj,

    Could you do a video explaining spiking neural networks and the use of them, as the search results for it on YouTube are not very satisfactory.

    Keep it up with your videos, awesome stuff!

  11. 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?

  12. 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.

  13. Hololense scans the room it wouldn't be great if we are able to segment out the scanned object and run object detection ? Basically using hololense as a camera what could go wrong ?

  14. I dont understand why i couldnt use 2d Conv for 3d objects? If i move in a 3d environment every frame is a plane. And the sum of an objects 2d plane views is the object. Maybe it is computational to expensive doing it like that. But you are also talking of graphs and not image recognition of 3d objects. Otherwise informative video.

  15. Man,can you please make a video maybe explaining the different functions in this pytorch library. I'm kinda new in the field.


  16. 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!

  17. This does bring ideas for a start up, do you have any links or books that relate to ethics and tech?

  18. I'm so glad you decided to do a video GCNs. I asked you on LinkedIn about this and you delivered. Really Cool!!!

  19. Glad to see GCNs (Graph Convolutional Networks) from our 2016 paper (https://arxiv.org/abs/1609.02907) featured and explained in such a great way.

    Source for some of the material used in the video: http://tkipf.github.io/graph-convolutional-networks/ and the accompanying code release: https://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: https://r2learning.github.io 🙂

  20. 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

Leave a Reply

Your email address will not be published. Required fields are marked *