koshian2/pseudo-label-keras

Pseudo-Label: Semi-Supervised Learning on CIFAR-10 in Keras

Pseudo-Label-Keras

A semi-supervised learning approach using pseudo labels. Implemented on Keras, evaluated on CIFAR-10.

Result

VGG-like 10-layer CNN

# True Labels	Supervised	Pseudo Labels	Pseudo Gain	# Pseudo Labels
500	39.81%	39.02%	-0.79%	49500
1000	47.95%	50.89%	2.94%	49000
5000	69.41%	74.58%	5.17%	45000
10000	77.96%	80.46%	2.50%	40000

vs MobileNet Transfer Leaning

Pseudo Labels	ImageNet Weights	N=500	N=1000	N=5000	N=10000
No	Yes	51.77%	60.13%	68.23%	74.92%
No	No	22.70%	30.81%	55.27%	65.77%
Yes	No	34.26%	44.95%	63.39%	72.94%
Yes	Yes	46.14%	51.37%	65.00%	73.86%

Reference

Dong-Hyun, Lee. Pseudo-Label : The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks. 2013 http://deeplearning.net/wp-content/uploads/2013/03/pseudo_label_final.pdf

Details(Japanese)

CIFAR-10を疑似ラベル（Pseudo-Label）を使った半教師あり学習で分類する https://qiita.com/koshian2/items/f4a458466b15bb91c7cb