News for hackers. Stuff that matters.
Re:Matryoshka Representation Learning by deepnet on Thursday February 01, 2024 @05:59AM (#39214670)
How do they sort the dimensions so the most salient are at the front and thus retained ?
Re:Matryoshka Representation Learning by yorwba on Thursday February 01, 2024 @07:39AM (#39215326)
They declare by fiat that retaining only the first few dimensions should lead to low classification error and construct the training loss accordingly. (Equation 1 on page 4 of the PDF.)