|
@ilyasut | |||||
|
The reason most (not all) methods don't add value (over baseline) when scaled is because they're "extra training data in disguise", so their benefit vanishes in the high data regime twitter.com/ilyasut/status…
|
||||||
|
||||||
|
Jan van Gemert
@jan_gemert
|
7. tra |
|
Well.. *All* current deep learning and machine learning is just "data in disguise": the 1-nearest neighbor algorithm converges to the best possible classifier just by adding more data. (converges to twice the Bayes error with infinite samples; en.m.wikipedia.org/wiki/K-nearest… )
|
||
|
||
|
Vixtor Namaste
@_rdm_8
|
7. tra |
|
"extra training data in disguise" ? can you plz elaborate.?
|
||
|
|
||
|
Dzmitry Pletnikau
@spring_stream
|
8. tra |
|
"New" algorithms introduce hardcoded prior "structure" to the solution, which baseline is capable of learning on its own, given enough data.
Therefore, on large data sets, "new" loses its advantage over baseline.
|
||
|
|
||
|
Hassaan Hashmi
@hassaanhash_mi
|
7. tra |
|
Should we, then, introduce a "Scalability score" metric for the algorithms; kind of like what the DxOmark is for cameras.
|
||
|
|
||