Another great question! Interpolative generalization works as long as the manifold hypothesis applies. This works independently of whether your latent data space is entirely connected, locally connected, or fully discrete.https://twitter.com/etsiJnaJ/status/1450558699714535428 …
-
-
1. The handheld camera: just keep the camera rolling as you walk from your kitchen to your backward. Boom, a smooth transition where *every single frame* is extracted from the real world.
Show this thread -
-
2. Fade to black: this trick exploits the fact that changing the brightness of an image still gives you a "valid" image. Just fade one frame to black then fade in to the new frame. A bit trivial, I know
Show this thread -
3. The shared focus: zoom in on one object shared between the two scenes (or background color/texture, e.g. the wood of the table) then zoom out into the new frame.
Show this thread -
4. A variant of this would be to, say, turn the camera towards the sky you can see from the window, then move it down, and boom, it's your backyard
Show this thread -
So anyways, the visual world is in fact an entirely connected space (you can go from any valid image to any other valid image via a continuous path where every frame is a valid image). This means that there exists a natural, intrinsic manifold of the visual world.
Show this thread -
This is true for every perception problem, remarkably. The *physical world* lies on a manifold. And that's what makes deep learning so effective -- its assumptions are a good match to reality.
Show this thread
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.