3 ms·

Separate, less tangential reply. I think what could really help computer vision is not as much emulating human eye tracking as it is understanding images as sh

Separate, less tangential reply. I think what could really help computer vision is not as much emulating human eye tracking as it is understanding images as short durations rather than instantaneous shots. The point is that during the eye's saccade reality is always changing. Think of how close-distance flirting works: you can't use your stereo vision to focus on both her eyes at once -- you have to alternate ever so slightly, occasionally meet her eyes focusing on yours and then lose them. All of this happens automatically -- our visual perception of the situation is intrinsically dynamic. Of course, here we're getting close to Henri Bergson's ideas about time and duration and Hubert Dreyfus' reading of Heidegger as a theoretician of situated existence. This is the same Hubert Dreyfus who told Simon, Minsky, etc. in the 50s they didn't have a chess computer yet and has recently passed away as a skeptic of automated cars. Did you notice self-driving cars are slowly disappearing as a near-term possibility? We still haven't figured it out -- we're trying to teach computers with frozen cases (MNIST, Imagenet, etc.) when the real problem is training them on actual situations. They don't need to emulate humans at all. They just need to exist in the same precise sense we exist. Computing power will keep increasing and I assume so will "deep learning" optimizers, but we need to figure out how to give computers a Dasein.

visarga 9y ago
> we're trying to teach computers with frozen cases (MNIST, Imagenet, etc.) when the real problem is training them on actual situations They are training AI agents on simulators, which are dynamic datasets as opposed to ImageNet which is a "frozen" dataset). > we need to figure out how to give computers a Dasein. Google and the whole AI community have been hard on the Reinforcement Learning bangdown. RL is based on situating an agent in the environment (real world or sim) where it learns to interact and achieve goals. That's situated AI. My intuition is that simulators are key to the next AI phase - where a neural net would be like a scientist creating new hypothesis and experiments, and the simulator would be like a lab, testing out those ideas (simulation + MC search + neural nets)
- thanatropism 9y ago
  > My intuition is that simulators are key to the next AI phase I hope so. Reinforcement learning and apprenticeship learning (inverse RL where the machine watches a human dealing with the problem) are probably the ground zero before which everything will look like toys and Kaggle problems.

visarga 9y ago

> we're trying to teach computers with frozen cases (MNIST, Imagenet, etc.) when the real problem is training them on actual situations They are training AI agents on simulators, which are dynamic datasets as opposed to ImageNet which is a "frozen" dataset). > we need to figure out how to give computers a Dasein. Google and the whole AI community have been hard on the Reinforcement Learning bangdown. RL is based on situating an agent in the environment (real world or sim) where it learns to interact and achieve goals. That's situated AI. My intuition is that simulators are key to the next AI phase - where a neural net would be like a scientist creating new hypothesis and experiments, and the simulator would be like a lab, testing out those ideas (simulation + MC search + neural nets)

thanatropism 9y ago
> My intuition is that simulators are key to the next AI phase I hope so. Reinforcement learning and apprenticeship learning (inverse RL where the machine watches a human dealing with the problem) are probably the ground zero before which everything will look like toys and Kaggle problems.