kalomaze: STE works fine under backprop even for binary weights and latents

kalomaze · x · 2026-09-30

Responding to skepticism about straight-through estimators, kalomaze argues STE works unironically fine under backprop even in the binary-weights case, noting it also powers autoencoders that learn binary latent spaces at the bottleneck — pointing to LFQ and the BitNet papers as evidence.

Original post →

More from Research

Research channel →