Graph Convolutional Networks - Oxford Geometric Deep Learning
ฝัง
- เผยแพร่เมื่อ 10 ม.ค. 2023
- In this video, I go over Graph Convolutional Networks!
Excellent blog post on GCNs (from one of the authors): tkipf.github.io/graph-convolu...
GCN paper (ICLR 2017): arxiv.org/abs/1609.02907
My website: federicobarbero.com - วิทยาศาสตร์และเทคโนโลยี
Came for the astronaut, stayed for the graphs.
Same bruh! :D
Mindblowing ! Thanks for the explanation!
Thanks for making this available on TH-cam.
Thanks for watching!
It will be great it you post all Geometric videos. Looking forward for more informative videos
Thank you!
This term will be fun!
Very nice, compact yet super clear explanation. Congrats! Idea for follow up video: downsampling graphs (e.g., max pooling or something smarter)?
Thank you! Definitely will consider that idea :)
Thanks a lot. Came from the demystifying GCN medium post. So, once we have "embedding" for each node ready, we may pass each node's embeddings into feature-vector matrix as row vectors. But I could not convince myself on the use of A.X.
A.X should be an nxd matrix, where, for each node in the row, there is a sum (average) of i-th dimensions of that node's neighbors, in the i-th column. I suppose that we are intending to represent the entire Graph network through A.X . But is it a good representation?
Nice course ! Thanks !
It’s good 😊
Geometric deep learning? Do you incorporate insights derived from Geometric Algebra and Clifford algebras?
👍🏻
the theory is based on spectral representation
👍
So, If I'm right, W must be d x 1. But is this all it is learning? Just d weights?
Hey Federico, I took a Linear Algebra for Big Data course this past semester and found that I understood most of the concepts that you presented in this video. I was wondering, as I have not began any Deep Learning courses, what does it mean for the weight matrix W to be “learnable?”
When he says it's 'learnable' he means that the values of that matrix will dynamically change with the learning process until it converges. As you said you didn't take any DL course (therefore, I infer, you also didn't take any machine learning course), I know this terms might be obscure to you. Briefly, the matrix W changes accordingly to your objective when using such a model.
Hi Gianangelo! As Caio said, when a matrix W is learnable, it means that its entries are optimized during the training process with respect to a specific objective function. In this sense you are able to "learn" the parameters of W during training. Once you learn W (and all the other learnable parameters), then you can make predictions with the model.
I am glad that you were still able to follow along without having taken a DL course! Unfortunately in this course I do assume prior knowledge in machine/deep learning so I would highly recommend to supplement this with a course/book on ML/DL.
@@dcaiothank you ciao, quite insightful! What type of objective W might you be looking for?
@@FedericoBarbero00Ahh, that makes sense. So the weight matrix finds principle components in a data set. If I understand correctly, you can make predictions because the weight matrix prioritizes the most prominent parameters in the dataset. Using the example in the video, if each node in A has a (high dimensional) vector of 10 variables, the variables could be waited | 0.93 0.05 0.01... | with the next seven summing up to 1% weight.
It means that the numbers part of the matrix W start being random. Then, through an optimization process, the numbers of W are changed until we find a set of numbers for W that can be multiplied by other variables and minimize a loss function. A loss function is a function that receives W and other variables and parameters and returns a number that measures how good a network is (less is better). The optimization process to find W (learn W) is tipically gradient descent.
Why Erlangen as rocket's name :)? (I am a student in FAU Erlangen, wondering if you studied there or you are from Erlangen)
u fucking god
P r o m o S M