Conversation

KFist (kfist@gs.smuglo.li)'s status on Sunday, 19-Nov-2017 22:57:45 UTC

KFist

in reply to

@augustus This is basically a good case study for the dangers of using rudimentary machine learning and markov chains to generate content or single-mindedly meet a goal. It's why there should be a high level of being careful about defining the goals and a large amount of thinking about just what a given learning algorithm will do.

It might absolutely annihilate the goals you set for it and become extremely efficient at what you created it for, but the resulting codebase and thought processes will be incomprehensible to humans. Sometimes the answers, even if you do understand them, might be insanely uncomfortable to realize. This machine learning algorithm had one goal: Achieve highest amount of ad revenue from Youtube Kids videos, using keyword-oriented video creation and tagging mechanisms. That's what it got.

You also see it in other fields. AIs which are optimized to find the objectively best candidates for a job negatively select against demographics such as recidivists or niggers. AIs that are optimized to identify objects based on visual characteristics and a shitton of keyword-linked criteria end up identifying a group of black faces as gorillas.

This is apparent to pretty much anyone who's done any amount of work on anything as basic as sequential solution-finding in response to a numerically complex ill-defined set of problems knows this. One of my friends from uni went on to do his Master's and Ph.D at Purdue on aerospace engineering. As proof-of-concept for one of his solution-finding techniques, he set up one of these machine-learning algorithms and told it to find the most energy-efficient way to replicate a SpaceX first-stage landing: start from a high altitude with a given speed, and arrive at a spot with a certain vertical and horizontal velocity, at a certain orientation, using a certain amount of energy deliverable by a single-vector source without "restarting the engine" or exceeding a given speed. There were a certain amount of aerodynamics simulations too.

One thing he forgot to add in was the ground. He thought a solution which involved going below the ground would be energy-inefficient due to gravitational potential energy considerations, but forgot about the aerodynamics fiddliness he programmed in. The program spent 7 hours crunching everything, only to give him a path which overshot the landing pad by a few hundred meters, went through the ground, and did a spiral burn down and around the pad, before popping up out of the ground and doing a perfect landing. The parameters as set up made the algorithm maximize control surface lift and take advantage of the elevation/pressure-dependent characteristics of the engine. Having no ground limitation (just a 'finish at at x,0' condition) made that possible. Technically it was the most energy-efficient solution, but it wasn't a useful one.

I have a feeling that machine learning and AIs will keep finding maximally efficient solutions to the problems they tackle, but they won't be useful to humans.