← Back to projects

Activation Function Ablation Study for 3D Point Cloud Classification

Completed

ReLU vs Leaky ReLU vs GELU in a PointNet baseline — accuracy, stability, and latency under a fair, single-switch setup

activations deeplearning ablationstudy
2025

About this project

  • Leaky ReLU and GELU tied for best accuracy (89.98%) on ModelNet10 in this setup.
  • ReLU delivered the fastest inference (2.4× faster than GELU) and is the most deployment/INT8-friendly.
  • GELU trained stably but was significantly slower on this backbone.

Gallery

speed_vs_accuracy

speed_vs_accuracy