abs() as activation function: a one-hidden-layer MLP just trains fine

burny_tech · x · 2026-09-27

A developer shared a counterintuitive experiment: using abs() instead of ReLU as the activation in a vanilla MLP with just one hidden layer (10 neurons) trains fine, with a video of the run.

The author claims abs() is faster than ReLU and more expressive, joking it could train in real time on an ESP32 — a low-cost exploration of an unconventional activation function that drew surprise from others in the AI community.

Original post →

More from Fun

Fun channel →