BCE averaging convention was secretly tuning your learning rate

BlackHC · x · 2026-09-16

A subtle gotcha from the author's continual learning experiments: mean BCE averages over a pair's two outputs; switching to summed loss (×2) drops accuracy from 58.1% to 51.2% because at a fixed SGD learning rate it doubles the update. The averaging convention was doing implicit LR tuning — treat loss scaling as part of your optimizer setup.

Related event: BlackHC's toy continual learning experiment: PCA replay hits 91.1%(6 posts)→

Original post →

More from Research

Research channel →