An Active Learning Method Based on Variational Autoencoder and DBSCAN Clustering

Table 5

Accuracy improvement percentage over random sampling on MNIST.

Number of labeled examples	K-means (%)	Mean shift (%)	DBSCAN with VAE (%)
15	3.72	3.70	5.12
20	10.56	2.76	11.30
25	4.21	2.43	5.11
30	6.42	1.96	5.48
35	3.82	3.07	4.11
40	0.33	−0.22	3.85
45	−0.96	1.30	2.15
50	1.20	1.77	3.24