@Natasha_Jay @quinn Holy crap, this is the first time I've heard of this.

I once worked as a developer with a team attempting to use ML for a non-invasive medical diagnostic device. The training data we had were... bad. Inconsistent, low quality, heavily class biased, and overall quite scarce.

I had to peace out when the CEO founder insisted that the answer was data synthesis—that we should just make random data with the same statistical distributions as existing data to even out class representation, and was completely bewildered at pushback that this was ludicrous.

I think I just heard him orgasm upon learning of silicon sampling.