An Active Learning Method Based on Variational Autoencoder and DBSCAN Clustering

Table 1

Mean accuracy (%) of different methods on MNIST from 20 runs.

Number of labeled examples	15	20	25	30	35	40	45	50
DBSCAN with VAE	49.76	58.51	59.67	60.90	64.38	69.02	69.68	71.28
Random	30.19	43.05	52.34	51.68	57.30	64.03	60.75	64.96
Uncertainty	37.74	48.29	44.15	59.35	61.41	67.69	61.70	68.59
Core Set	49.03	48.03	54.51	54.99	56.34	58.89	53.97	60.13
EGL	44.73	46.80	41.27	52.37	49.89	51.49	57.05	59.73
Bayesian	32.33	31.50	35.30	37.59	39.52	45.13	54.41	56.21