A person who hasn't considered very deeply the complexities of comprehension would be unimpressed if a computer vision system categorized a girl as a car because she is wearing a tee-shirt with a Mercedes-Benz depicted on it. Perhaps the photographer deliberately featured the shirt in the photo rather than taking a head shot, inadvertently fooling the computer vision system. The creator of the system might defend its output by pointing out that the image legitimately fits into four categories.
- A photo
- A girl
- A shirt
- A car
We can dismiss the shortcomings of the computer vision system with appreciation of what has already been achieved and optimism about the bright future we see for computer vision systems and AI in general. Bright outlook is characteristic of all scientific and technological communities and beneficial to the continuation of progress. The hope of making an important contribution to that progress is why people get involved.
Another ingredient to excellence in that progress is self-evaluation. The creator of the computer vision system may realize that the end user of the computer system expects an output of
girl rather than
car. It would be a credit to the system's creator to transcended the myopia of the specialist outlined by Thomas S. Kuhn in his classic book Scientific Revolutions (1962) and enter a deeper avenue of research.
In this case, this deeper avenue is into comprehension. What is the semantic relationship and meaning of the objects found in the images? The car is printed on the shirt worn by the girl photographed.
If a small child was shown the same image, it would not be surprising if the child drew the same conclusion, pointed, and said, "Car." With this example and others like it we can see that the artificial nature of the system is not the cause of its limitation.
The ability to distinguish a car from a shirt depicting a car does not have to do with the carbon or silicon nature of the learning system. It has to do with the depth of abstraction to which he, she, or it is able to learn.
Dysfunctional Bias in Thinking
Many of those in the AI field have become predisposed to thinking about the degree of intelligence as a single bit. It is either narrow or general. Others have proposed middle ground, but these midpoints are merely an amalgamation of characteristics from narrow and general categories. These extensions of the narrow-general paradigm are still biased by that paradigm.
It is important to understand this when discussing, "Human Intelligence [in contrast with] Artificial Intelligence." Attachment to the ideal of general intelligence probably began with the pioneering work of statistical psychologist Charles Edward Spearman and most often referenced in his General Intelligence, Objectively Determined and Measured (Am. J. of Psych. V 15, no 2, April 1904).
The idea of narrow intelligence did NOT come from AI books in the 1990s or news articles from this century. The concept was widely known in the 1970s and 1980s as a result of work at the labs at MIT and Princeton, mostly funded by the U.S. military. Rome Air Development Center at Griffiss Air Force Base in New York wrote in a now declassified December 1983 report the following.
In the Intelligence Analytical Methodology subthrust developments
will enrich the intelligence analysis process by integrating advance data
base techniques, knowledge based expert systems, and analytical aids. In
the area of artificial intelligence, S-O-A techniques such as knowledge
acquisition, knowledge representation, and inference aids will be applied
to a narrow intelligence domain. Analytical methodologies will address near term operational issues by exploiting trend analysis, hypothesis formulation, and forecast generation.
There is no strictly general intelligence on earth. Human beings can imagine their own intelligence to be general simply because the constraints of human intelligence are shared. Although some have imagined super-intelligence, we cannot easily conceptualize it because it is outside our experience. If it were to exist, it might possibly fall outside our capacity to detect, and, if detectable, appreciated.
All intelligence is narrow in that it works to effectively achieve objectives with in a FINITE SET of scenarios within a FINITE SET of domains. There are problems the objectives of which are purely intellectual and well understood but the solutions of which have eluded the brightest humans in history.
The emergence of synonyms to narrow AI (like weak AI) are attempts to develop an approach to general AI, which is an objective that rides on the misconception that it exists. Not only does general AI definitely not yet exist, it has not yet been proven mathematically that unlimited intelligence can exist. It is likely that there will always exist at least one problem for which the current maximum intelligence is below capability.
Face of the Problem
The Latin term prima facie does not have a short literal translation into English but is pertinent when comparing the intelligence of a human with that of the artificial. If we strip all the preconceptions about intelligence we have acquired from the literature and look at the face of what people associate with intelligence, we find that there are three main categorically different levels of intelligence with regard to a topic or scenario.
- Can't say a thing
- Poor in judgment
- Usually gets it right
These are not formal definitions, but every adult can identify with these three overarching levels of effective intelligence. Most would agree that no one or thing ever gets everything right all the time, so there is no fourth category.
All intelligence tests such as these are attempts to quantify.
- Aptitude testing used by selective corporations
- Credit histories used by investors
- College boards
However, the above three (silent, wrong, usually right) are categorical measurements of intelligence. These are on a per scenario basis. A person or artificial system may have nothing to say in one situation but be trusted to frequently get it right in another.
These prima facie categories are more important than numbers when considering the question, toward what answer do we design the computer vision system in the photo-girl-shirt-car example and in what direction should research be directed.
Human Intelligence Vs Artificial Intelligence
To approach this distinction, the term, "uncertainty problem," mentioned in the question we are answering is an excellent entry point. Uncertainty is not a problem. Uncertainty is actually an indication of intelligence.
The small child that calls the image of the Mercedes on the shirt of the girl photographed in the example above is not yet trained enough in cognitive skills to raise their tone at the end the word, "Car?" indicating an interrogative.
The parent may say, "Yes that's a car," or, "No, that's a shirt with a car on the front." Even if the parent chooses the first option and none of the child's teachers point out the distinction, the child will arrive at the deeper abstraction naturally, during socialization.
That is what is remarkable about human intelligence.
It is not general intelligence but rather the ability to acquire knowledge about a somewhat general set of domains, which is an ENTIRELY DIFFERENT CAPABILITY than machine learning or executing rules in a production system. The child, without being trained, acquires the ability to comprehend images in increasing depth, and this depth is a quality related to abstraction, not reliability or accuracy.
The science, technology, and linguistics are independent of the object recognition sophistication in distinguishing what is depicted in a visual field. Here are some potential expressions of comprehension from the novice to the more sphisticated.
- Car? — Uncertainty is indicated by the interrogative as the child points.
- A picture of a car
- A picture of a girl wearing a shirt with a car on the shirt's front
- The photographer is featuring the car on the girl's shirt
- Did the photographer like the car, the graphic on the shirt, or the girl's figure? — This increased uncertainty is an attribute of the higher intelligence in this list, placing the system that conceptualizes it in the Usually gets it right category for this kind of comprehension task, not the Can't say a thing, or Poor in judgment categories.
These seven are levels of cognitive ability specific to the photo-girl-shirt-car scinario.
Here we can now see the disparity between artificial intelligence and human intelligence clearly. Humans, if there is no organic damage to the brain and optical acquisition system will grow from #1 to #7 through exposure to normal life. Artificial intelligence, at the current point of development, must be trained and may only reach the level of #4 or #5.
The emergence of software systems that can move through these layers of semantic meaning with ease and apply them to decision making, will, if possible at all within the span of human intelligence on planet earth, take time.
Note also that these semantic and largely statistical conclusions about meaning are not completely rational. A child doesn't write a logical equation to understand that a graphic is on a shirt. Cognition is an ability UNDERNEATH rationality, not above. It is because we can cognate that we can learn the rules of logical inference.
It is also possible that artificial power may precede by centuries artificial depth of thought.
An intelligent system that could not cognitively detect the uncertainty around photographer's possible interest in either the car, the graphic, or the girl may yet be able to overpower humanity in a struggle for dominance. Books, teleplays, and screenplays discuss exactly this. The more humanistic writers always place the depth of human understanding above that of the machines.
Other writers have doubt about the depth of human intelligence. When the following statement is fully considered to the depth of human understanding and without rationalization, it is not difficult to side with the doubters.
I'm going to take the car to the gym to use the stationary bike for a cardio workout. I'll be back in an hour.
The traditional response when this is brought to attention of the gym member is,
"Ride my bike? The tire is flat, ... and ... the traffic is too dangerous round here. Plus I don't want to waste my gym membership."
An anthropologist with an interest in psychology and the environment might comment about this reasoning.
"That is a rationale for essentially neurotic behavior. Such trends in the emphasis on convenient personal transportation is the cause of the unsustainable combustion of an important natural resource that took tens of millions of years to form in the earth on only a few centuries.""
If the anthropoligist knew thermodynamics too, they may say ...
What is more pertinent is that the inevitable depletion is thermodynamically irretrievable, and future generations will probably look back at what we now called developed economy and say that humans in the twenty and twenty-first centuries were either stupid, selfish, or insane.
Are humans intelligent or stupid? It depends on the scenario. Same as AI systems. All intelligence thus far is narrow.
Regarding a Other Concepts Presented
Causality is not a new thing in science. In fact, it necessarily preceded science. Faraday would not have set of an experiment to test a particular electromagnetic effect if causality escaped his mind.
Volition is not proven. If Minsky's statement that the brain is simply a meat machine, it is possible that what appears to be free will may be as deterministic as Faraday's electromagnetism. That is actually the perspective of much research: That everything is deterministic and in time will be predictable.
Is this optimism scientific? No. It is unprovable, and this faith in determinism has a huge subconscious force: The emotional pain of uncertainty.
The more tempered understanding throws four practical and evidence-based wrenches into optimistic futurism.
- Heisenburg's uncertainty principle
- Gödel's incompleteness theorems
- The butterfly effect (causal sensitivity to initial state that was exposed by the mathematics of the chaotic behavior of complex systems)
- The Sixth Extinction (Richard Leakey, Roger Lewin 1995)
A Computer Guiding Politics?
Can AI, with finality answer the question, "Is climate change real?"
Computers already determined that. Climate change was occurring long before industrialization, and the current rate of change seems to be exacerbated by it. We can demonstrate using basic statistical modelling that the increase in northern hemisphere terrestrial and ocean surface temperatures track about thirty years behind increases in artificial carbon emissions.
That thirty year lag was the output of a basic Levenberg–Marquardt curve fit. So a grandparent of AI technique already answered the question of statistical correlation.
The question of causality and the auspice of climate based dangers is something that humans would never entrust to a computer. There would be a FaceBook argument about whether the AI programmers were in one political party or another. You'd see in the news allegations that someone from the pool of usual suspects hacked the system before it produced the result. Some of it might even be true.
Intelligence may be, like beauty, in the eye of the beholder.
What does what better is only meaningful in the realm of science when one defines the two systems and the comparison operation clearly.
- A particular human with a particular level of caffeination
- A particular computer with a particular set of computing resources)
- The meaning of better, define in such a way that the results from a series of tests can be formally evaluated
Then one would have a scientific and possibly factual determination. Most of what you read and hear about AI in the years surrounding this writing is best characterized by one of two words.