3 ms·

Preventing undesirable behavior of intelligent machines

• salawat 7y ago
  Now this is fascinating! This is a start toward what I'd call a sane form of ML Engineering. That is, assuming you can manage to mathematically express the fairness constraints ahead of time. It'll go a long way toward proofing against known hazardous functions. One thing I'm curious of though is whether or not this approach severely increased training time. I'd imagine it would, since for every constraint you're stripping out a higher and higher chunk of possible functions to emulate. Although it makes me wonder whether the time viable candidate function could be found may be usable as a sort of walking stick to tease out relationships not immediately apparent in the data. I'd expect convergence/divergence of constraints and primary function optimizations to play a big role in determining whether your dataset will converge to a solution, or whether you just sent you're supercomputer loose on a wild goose chase.