14/ In the image domain, even detection tasks like interpreting radiological imaging, which humans train for years to do competently, can now be done by ML systems with superhuman performance. Not yet in general, but task-by-task, ML systems will become the new standard.
1
Replying to @celine_dionysus @default_friend
Good luck! I've got a similar feeling about game engines, which seem to be becoming straight up reality engines. Watching this video about producing a photorealistic cinematic sequence entirely in a game engine blew my mind. Makes me want to dive in deep. youtu.be/0iQJkSpOoOQ
The LargeVis paper says that the analogous term in the loss there scales from 1 to thousands, so they have a gradient norm problem, but in UMAP, these vary just between 0 and 1, and so it seems like this should work fine? Just like cross-entropy everywhere else?
Thinking about this more, is this really necessary in UMAP? Here's the original gradient as described in that post. Isn't v_ij just a probability? And can't you describe the whole loss as just cross-entropy between the probs implied by the source and target space distances?
2







