Stupid anime reference from me but in the anime psychopass they averaged out the brains of the most societally outcast people to create a superintelligence to manage society, kinda feels similar to remembering the most memorable examples 😂 Great video 👍
I watched an episode or two of psychopass but didn't get that far! I gotta keep going... Popculture references aren't stupid, they can be great ways to communicate! One of my favorite animes is Gurren Laggan, folks have serious spiritual relationships to that one, haha
so ideally you have a data set of not just poor handwriting, but examples where 2s look like 1s and 1s look like 2s, so it gets good at solving the edge cases. you want there to be a fine line between concepts, and weird examples allows it to classify clearly. labeling ambiguous examples is worth more than labeling obvious examples, because they better define the boundary of the vector space holding these answers. extremely ambiguous examples are by definition, at the boundary of their classification, and if they cancel each other out, you don't really need normal examples, because those are just the average of the extreme examples. Weird examples don't just make it better at remembering, it makes it better at understanding the actual boundaries of the concepts its classifying.
Is it not possible to increase number of classes dynamically? like at start we only know there is 5 classes only(0-4) but we don't know how many more classes will come so made the model with 5 outputs, then 3 more classes came and we add 3 more neurons to last layer (let say 5, 8 and 9) also at this stage we don't know how many more classes will come and at end 2 more classes came so we add 2 new output neurons for 6 and 7.
@@Thinkstr yes, but what if we update fisher matrix before every new task. will it work? means is it possible to add new weights every time we add new classes in the model?
@@AbdulRazique-z2j Huh, I really don't know... You're reminding me of GANs which make progressively larger images, first learning small images and then expanding them.
I'm afraid I can't find the code for this video, but the next video did the same thing for reinforcement learning, and I've got that code here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-fxyttf6T5cA.html github.com/TedTinker/Tinker_FROMP_RL