We often hear that the AI is more accurate than humans in some image matching, but it also clearly makes mistakes that no human would make. Well here is a great example of such a mistake which a human wouldn't be making.
I can see a pattern. It helps if you think strictly in terms of what visual feature a ball and a bald head have in common: a crescent of glare against the sun.
* The ball is obscured by legs, shadow, or a player's body (this appears to be majority of the occurrences to me) - this one seems to trigger even if it's only out of view for a fraction of a second
* The high kick at 2:07 is similar to the above, except it's the glare from the sun instead of shadow that obscures the ball (ambient brightness matches the crescent instead of the umbra on the ball)
* Most of the rest I can't even find the ball myself, I think it might be too small in the far end of the field to register clearly
I think it is a testament to not just human complexity, but human efficiency.
Yes, the term has become accepted, but it's a horribly irresponsible label because it massively misleads laypeople about how these systems work and how trustworthy they are.
Technically AI is correct, since AI doesn't necessarily refer to human level AGI. Unfortunately it's highly misleading because that distinction seems to be lost on the average person.
Machine learning doesn't seem to carry the same science fiction implications for whatever reason.
Time: False. Time can be included as an input. Previous system state can also be communicated forward in various ways (ex the LSTM architecture).
Conceptual context: I'll agree, since that seems like it would require general abstract reasoning ability (ie strong AI). Similarly, robust treatment in the general case of physical continuity and cause and effect both seem like special cases of conceptual context to me.
That being said, bear in mind that we already have examples of not so robust treatment for special cases of physical continuity as well as cause and effect. (Example: https://news.developer.nvidia.com/transforming-standard-vide...) Who knows how close (or far) we might be from a general solution?
I mean that they don't understand time as something that's unidirectional. (To the best of my knowledge! Would be happy to be proven wrong.)
But I don't see any reason you couldn't train a model to incorporate the underlying assumption that the flow of time is unidirectional. I expect you would just need inputs (ie training data) reflecting that fact coupled with an appropriate loss function.
(Aren't networks that predict future physical states, such as motion, more or less doing this?)