IDPOC: Class Incremental Learning Based on Identically Distributed Parallel One-Class Classifiers

Paper

Official repository of Class Incremental Learning Based on Identically Distributed Parallel One-Class Classifiers

Published at NeuroComputing

Setup

• Use ./utils/main.py to run experiments.
• Some training result can be found in folder ./result.

Datasets

Class-IL / Task-IL settings

• Sequential MNIST
• Sequential CIFAR-10
• Sequential CIFAR-100
• Sequential Tiny ImageNet

Performance

	MNIST	CIFAR10	CIFAR100	Tiny-ImageNet
DER++	85.61	64.88	24.75	10.96
LWF	21.62	19.59	9.20	9.36
OWM	96.30	52.83	27.63	15.30
ILCOC	86.05	38.40	24.39	16.97
DisCOIL	96.69	44.54	27.50	19.75
IDPOC	87.51	55.50	29.08	22.55

Visualizaiton of score distributions

Requirement

• numpy==1.16.4
• Pillow==6.1.0
• torch==1.3.1
• torchvision==0.4.2

Related repository

https://github.com/aimagelab/mammoth

Citation

@article{sun2023class,

title={Class Incremental Learning based on Identically Distributed Parallel One-Class Classifiers},

author={Sun, Wenju and Li, Qingyong and Zhang, Jing and Wang, Wen and Geng, YangLi-ao},

journal={Neurocomputing},

volume={556},

pages={126579},

year={2023},

publisher={Elsevier}

}