For some reason, as a non-native English speaker, I've really struggled to pronounce "capsule". Hope you didn't mind that and you still managed to enjoy the video.
Thank you so much for the feedback! I was not aware that my English may be difficult to understand. I know that RU-vid generates subtitles automatically for videos in English. Have you tried those?
I've added manually edited subtitles to this video. I hope they help you to better understand the content and diminuate the confusion caused by my mispronunciation!
Most of my mathematical intuition about capsules was created by reading the following article: medium.com/@debasritac/capsule-networks-a-simplified-mathematical-understanding-and-why-they-fell-out-of-favor-458d150ed36a#:~:text=Mathematically%2C%20a%20capsule%20is%20represented,the%20presence%20of%20a%20color. Hope it helps you out! Pleae, let me know if helped and if you have any other questions. :)
@@datamlistic the mathematic is quite confusing but i think i get a liitle grasp in it. and can i ask what is the function that used in the digitcaps so it could know the spatial relationship between features. thank you sir!
5:17 Is there a way to visualize how these equations translate in practice? A visual animation as to how the code relates to the previous explanation. How and where it is used for less advanced readers. This paper is extraordinarily interesting, yet the final formula puzzles me a bit. Are we just to replace the pooling layer by a dense layer with softmax and a squash? Could you elaborate on that final part? Make it visual "we do that to get that result, because..."
Hmm, maybe I should have indeed spent a little bit more time explaining that formula better. In some way (maybe not the most correct one), I personally like to think of capsules as a vector generalization of a neuron (neuron: scalar in -> scalar out, capsule: vector in -> vector out). They both "route" information from one layer to the next one, but because capsules are N-dimensional, it's easier for them to encode pose information. But again that's my view of it and might be not the most right one. Maybe the image at the top of this repository helps you better understand what's happening by making a parallel with an artificial neuron: github.com/naturomics/CapsNet-Tensorflow. Let me know if it helps, and if it doesn't I may try to explain it in a different way.
@@datamlistic Oh makes sense now. When comparing the scalar vs vector it hit me. This is always the issue of knowing what goes where. understanding the language.
