3h ago

Saining Xie, NYU Courant Assistant Professor, notes that weighting schemes for combining neural network block outputs and residual streams were already examined between 2016 and 2019 in feature pyramid research

Lucas Beyer agreed the historical record appears comprehensive.

——0——
Original post
OPSaining Xie#158Saining Xie@SAININGXIE

@giffmana @1jaskiratsingh great discussion. thanks! for better or worse, i feel like every weighting scheme, including soft/ learnable weighting, was explored back in 2016 to 2019 when people were studying feature pyramids for dense prediction tasks :)

6:28 AM · May 22, 2026 View on X
REPLYLucas Beyer (bl16)#55Lucas Beyer (bl16)@GIFFMANA

@sainingxie @1jaskiratsingh You're probably right, so you should be well placed to pick the good one to start with :D

Saining XieSaining Xie@sainingxie

@giffmana @1jaskiratsingh great discussion. thanks! for better or worse, i feel like every weighting scheme, including soft/ learnable weighting, was explored back in 2016 to 2019 when people were studying feature pyramids for dense prediction tasks :)

1:28 PM · May 22, 2026 · 101 Views
3:47 PM · May 22, 2026 · 38 Views
Saining Xie, NYU Courant Assistant Professor, notes that weighting schemes for combining neural network block outputs and residual streams were already examined between 2016 and 2019 in feature pyramid research · Digg