Taming Small-sample Bias in Low-budget Active Learning