For my day job, I recently wrote an article on the use of artificial intelligence within a certain specialized field of investing.
I won't tell you what the investments in question were. Rather, I'd like to make a point about AI that I learned, or at least came to appreciate more fully, while working on this piece.
One of my interviewees for the story told me this,
“For many people, AI is still a black box. But the source of the innovation, and of the resulting errors, is in principle simple: a random number generator. This is attached to a machine-learning system that reins in that randomness according to a model of what is realistic and what is not.”
So: randomness is critical to AI, just as it is critical to, say, biological evolution. Random variations give the process of natural selection some clay to shape.
In the AI context, this notion of AI as a box containing a random number generator which has to be restrained reminds me of Goldilock's love of porridge if and only if the temperature of the porridge is JUUUUUST right. The porridge may have to be heated so that it can be allowed to cool down to perfection.
Too much randomness, and AI is concocting a dreamworld, and your risk management goes out the window. Or, to take a recent non-financial example: you can end up with a newspaper running articles about the hot summer books -- and offering an AI generated list of books that do not exist.
Too little randomness, though, and you have "overfitting," a slavish obedience to historic trends that leaves the algorithms blind to the possibility of the unexpected -- the unknown unknowns. Developers in the AI space will continue to try to get the amount of randomness that the system allows its number generator "juuuust right."
Just a thought. Not a porridge that I've allowed to cool yet.
More on AI tomorrow.
I will wait for the porridge. Or a good, hot bowl of French onion soup.
ReplyDelete