[grugagag]

Yeah but you missed the point op was making

[markburns]

That seems an uncharitable view of the reply.

The search space is huge, we sometimes find needles in haystacks by accident, isn’t it exciting that we have tools now that can systematically check every piece of hay?

[richrichie]

ML search is more about ‘averages’ based on samples.

Innovations like these are more about ‘shocks’ that surface fitting cannot capture.

Note universal approximation theorem applies only to smooth surfaces.

[tomrod]

Not always. Quantile regression exists. And you can develop "no match" categories.

[richrichie]

Quantile regression is also about averages.

[tomrod]

Averages are formulated as measures of centrality in the L2 norm ("straight line" distance), sum(values) / count(values). Quantile regression uses modifications the L1 norm ("city block" distance); if median (50%) then it is a measure of centrality. Not everything is an average. If you're interested, this is a good (but math heavy) treatment: https://en.wikipedia.org/wiki/Quantile_regression#Computatio...

[radarsat1]

But the better the mean surface is fitted (in a generizable way), the easier it is to spot outliers.

[kylehotchkiss]

Well said.

[dojomouse]

Perhaps. I was thinking along the lines of MarkBurns response - ML will allow us to efficiently look in those places we might otherwise only have searched by accident.

If ops point was rather that “accident”/“luck” are uniquely human… I don’t agree. Luck is when probability works out in your favour - and that can happen all the time with any sort of probabilistic search, which is rife in ML.