I attended a nice Meetup last night, and one of the speakers quoted Turing’s description of a program able to modify its own instructions as a good description of artificial intelligence (AI) and another had a nifty typology relating all the standard tools in data science to their role in AI. This set me to thinking about the standards to which AI should be held.
I began a thought experiment:
Here, AI, is a dataset. I want to find out X. What do you suggest is the best way?
and AI would respond with one of three answers:
- This is a simple problem. Here are the standard solutions. You can read about the advantages and disadvantages through the help menu or you can select one to run.
- This is a complicated problem, composed of several simple problems. The following combinations have proven helpful. You can read about the advantages and disadvantages through the help menu or you can select one to run.
- This is a complex problem. Either more data is required or you will need further research in the domain for solution approaches. Standard tools can be expected to yield different results with the same input.
It doesn’t appear we’ve reached my heart’s desire yet. In a paper that I wished I had saved, I read a criticism of neural networks that they can operate as black boxes, producing perfectly useful results, but providing no traceable method as to how.
That’s not fair! After all, sauce for the goose …
Science is what we understand well enough to explain to a computer. Art is everything else we do.
Donald Knuth, in the forward to A=B (https://www.math.upenn.edu/~wilf/AeqB.pdf).
Why shouldn’t we hold AI to the same standard, and require it to explain to humans?
Finally, I am reminded of Croesus, a king in Asia Minor, not he of the Golden Touch. There were geopolitical issues with the Persian Emperor Cyrus, and Croesus made war plans. Like any good ruler of the time, he consulted the Oracles. A common report is
If you go to war with the Persians, you will a destroy a great empire.
Spoiler: Croesus lost big.
Herodotus’ Histories has a more likely version of the Delphic oracle
Wait till the time shall come when a mule is monarch of Media;
Then, thou delicate Lydian, away to the pebbles of Hermus;
Haste, oh! haste thee away, nor blush to behave like a coward.
(https://sourcebooks.fordham.edu/ancient/herodotus-creususandsolon.asp)
The king managed to torture this result into a favorable auspice.
As successful as AI is proving at pattern recognition and classification, for example, we should take care not to rely heavily on results given without explanation.