A Study on the Effectiveness of the Class Imbalance of Training Data on Convolutional Networks(CNNs)
Experiment about class imbalance problems inside the CNNs
- pytorch, python : pytorch 1.0 ↑, python 3.7 ↑
- package : numpy, os
- Neuron Membership
2. Major class actvation - Minor class activation
3. Class Selectivity
- Data : Cifar10 with different imbalance ratio
Dataset |
Cifar10 |
Major / Minor |
Major class images per group |
Minor class images per group |
Accuracy |
Balanced |
[0~4] / [5~9] |
5000 |
5000 |
91.29 |
Imbalance 20 |
250 |
68.93 |
Imbalance 50 |
100 |
58.51 |
Imbalance 100 |
50 |
52.45 |
- Model : Vgg11 with batch normalization
- Neuron Membership
- Major class actvation - Minor class activation
- Class Selectivity