Example: When cutting a scene with a lot of people eyes, you can really push the limits because of the high likelyhood that the audience will have a certain amount of focus on some eyes.
I tried to major in film (it was a program halfway between film studies and technical work), and remember a big technical thing being not producing an "axis break" (I won't know the terms in english) by changing angles too suddenly or even changing lenses too abruptly from wide to long. Unless you were trying to break the rules on purpose. The big exception was "plane/counter-plane" shot sequences where we alternated between seeing two characters who were talking face-to-face directly. This was 15 years ago and not in a country with an actual film industry.
Then the reason why this worked was said to be insert long French theoretical evocation you identified with the characters. The lacanian gaze, that kind of thing.
So in light of what you just said, the whole "axis break" makes perfect sense: film sequences are trying to show images I might have plausibly apprehended in real life via eye movements. Because I'm always seeing it from someone's point of view. The lacanian gaze.
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.
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)
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.
Eye movements while taking to someone, posture during conversation, and voice characteristics could all be measured.
The point is feeding the network slightly different images as they're captured by a small field of vision. The real perceived object is not static.
You're going to need a temporal, recurrent type of model. Also: a bigger boat.
For example on two different occasions when I was hanging with two different friends of mine we looked at Tinder profiles of women on my phone and both of them had almost the exact same reaction when I started swiping; "wow, you and me have really different taste in women!"
In both cases I was swiping on the women that I found to look the most attractive. Previously I might have thought that everyone would agree that these women were attractive.
Absolutely agree. I'm just saying that even if they were I think it'd be hard to measure attractiveness objectively because different people have different likes and dislikes.
However you probably are right that there are features that are considered attractive by almost everybody and that we just overlook those.