iVoro

Official code for Progressive Voronoi Diagram Subdivision Enables Accurate Data-free Class-Incremental Learning, ICLR'23.

Introduction

The iVoro method is based on the idea of Voronoi Diagram subdivision from Computational Geometry.

A. Establish Voronoi Diagram based on base model.
B. Insertion of a new class as a new Voronoi cell enables the minimal intervention to the overall structure.
C. Divide-and-conquer (a classical algorithm for Voronoi construction) efficiently introduce a batch of new classes into the system.

Results

The results of MNIST in 2D space below clearly showed different space subdivision results from conventional fine-tuning, PASS, and different variants of iVoro.

Reproducing the results

Step 1. Training of the base model, please follow PASS (github).

Step 2. Download the feature files. Google Drive

Go to the directory:

cd MNIST

Then run analysis/CIFAR_voro.py in following order:

A. iVoro

B. iVoro-D

C. iVoro-AC/AI

D. iVoro-L

Reference

ICLR'23

Acknowledgments

Contact

If you have any problem please contact me.