Can ML/AI understand incomplete constructs like humans?
Do humans have some inherent experiences in life which makes AI incapable of performing [some capacities of human intelligence]?
Comprehension of Literature and Film
Whether software exists today that is able to understand like humans is not something that the general public can know. No such system has been released to the general public by any military or commercial organization thus far. Yet such an achievement would not necessarily be something the government or commercial entity would want to disclose outside of the lab and its management.
Deeper comprehension of full or partial speech may have been achieved by anti-terrorist units, since funding has been available for that work for well over a decade, but it also unlikely that some software somewhere can read or scan parts of a book and do a book report that would produce a passing grade.
Determining whether a movie will return its investment at the box office may have been accomplished by the researchers for the big studios, but that doesn't require understanding like someone understands their favorite movie. The answer to the title question is, "Not yet."
Will it occur?
Most in the fields of computer science, robotics, and artificial intelligence say, "Yes." For religious reasons, some say, "No." I'm in neither camp, and have not seen indisputable mathematical rigor that proves either the inevitability of artificial brains or the impossibility of them. It is scientifically irresponsible to make a positive statement based on either recent technology trends or superstitious fears.
What are Humans Like?
The phrase, "Like humans," places a significant demand on researchers and engineers. Human beings can do much more than work with data sets (audiovisual in this case) with missing information, what statisticians call sparse matrices.
Whether software can realize higher achievements of human brains is unknown, and the predictability of such capabilities presents gross difficulties. Consider these human capabilities.
- Write a screenplay, cast it with artificial characters, direct it, and produce it.
- Initiate and partly develop a new branch of science, as did Isaac Newton or Lavoisier.
- Love beyond a superficial expression of love
Conversely, what humans generally do poorly is distinguish between reliable projection and baseless conjecture. Most humans are prone to musings of technical visionaries, propaganda, marketing, rumor, innuendo, and gossip. It would not be surprising for software to someday soon be wiser in this respect, but only because the bar has been set so low by human culture.
There are other capabilities of the human brain that are exceptional and present enormous difficulties in even considering an approach to realizing in software. These may be a result of a hundred thousand years of DNA refinement yielding significant complexity and precision. It is also not outside the realm of possibility that a form of causality exists in the human brain that defies scientific study.
The Possibility of Impossibility
No one has every proven that all things that exist can be measured. Heisenberg has actually proven the opposite to most theoretical physicist's satisfaction.
A phenomenon that cannot be measured at all cannot be studied scientifically. Whether the phenomenon of choice is a unique condition has yet to be understood even in question form, prohibiting the emergence of a proper answer thus far.
Imminent Change as Significant as Industrialization
Nonetheless, important capabilities of human intelligence have been simulated and others are emerging. These are now part of the world economy and will not likely disappear. The possibility of Asimov like scenarios of robots and humans coexisting and conversing in more human-like is very likely.
It is when that occurs that autonomous vehicles and walking robots will begin to have experiences that are like human experiences and we will be able to directly observe just how much like humans they can behave in terms of intelligence and also emotions.
About the Specific Capabilities Mentioned
These are the capabilities mentioned in the question.
- Predicting images
- Predicting objects in an image
- Understanding audio
- Understanding the meaning of spoken sentences
- Understanding a movie from the first half
- Understanding a movie from parts of it
These are the more canonical ways of stating the first four capabilities for which research has produced usable system approaches.
- Learning to distinguish image categories
- Learning to distinguish object categories from within images
- Parsing audio into notes, vocal tones, and transient sounds
- Extracting semantics from a vocalization sufficiently to respond intelligently some of the time
These are a more accurate cognitive science description of the last two.
- Guessing much better than random guessing would the story arch to the end of a movie from the sound and frame set of the its first half
- Filling in character and story arch details from portions of a movie's sound and frame set
There is no obvious reason why such couldn't be done and done well by software, given sufficient research time to develop such a system and sufficient data to train with. Also, significant computing resources would be needed, of course, and possibly some nontrivial period of time.