Hello, Alfredo) Great video! I see little response from students. When I was young, I didn't really want to study either. I want to study now, but I don't have time)))
@@doyourealise I enjoy watching them as well, honestly. Haha. And I *do* laugh at my own jokes, haha. There's some slight patronising tone that makes me crack up at times. And I don't think I even realise it when I teach. Although, to be honest, some do complain every year about my condescendence. But whatever, most of the students are having fun me included, so I'll keep it this way.
@@alfcnz yeah , like you know if the 8 out of 10 enjoy it then we should keep doing it :)...Specially its hard to find someone who teaches in a fun way or different way....
I checked the Residual Gated GCN paper but your notations seems little bit different. For the gate vector ( I mean the left vector at the hadamard product), they simply say it as sigm(Ah_i + Bh_l). I tried to simplify your notation to that but cannot made it. Do you have any other detailed resource for that notation?
Thanks for asking 🙂 If you check the notebook github.com/Atcold/pytorch-Deep-Learning/blob/master/16-gated_GCN.ipynb you can see it's an updated version of the author's drive.google.com/file/d/1WG5t6X12Z70JPtvA2-2PzdK3TMTQMsvm/ The formula came from reverse engineering the notebook itself.
I highly recommend this short introduction to GCNs ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-2KRAOZIULzw.html before watching Alfredo's lecture. It will give you a good overview and will make it easier to follow Alfredo.
Why have Graph Convolutional Networks been applied to image data based tasks like classifications? Is there any advantage to leverage domain sparsity as opposed to using CNNs for extracting features across the whole domain? I am very confused as to why GCNs are used for image data.
@@alfcnz I remember older SotA image segmentation methods used CRFs over the image's "superpixels" (irregular connected regions of correlated pixels) before U-Nets and the like took over. Maybe the (planar) graph of adjacent superpixels could be the bridge there (although I have no clue if it's a good idea or not 🤷♂️️)
