In table I we can see the mean accuracy (over the training
period) according to different budget values (ranging from
5% to 1%). We can see that as the budget decreases the
results tend to be worse, particularly when the budget is
only 1% or 2%. We can also observe that the strategies
VarUncertainty, RandVarUncertainty and SelSampling seem to dominate.
It is interesting to see that FixedUncertainty shows
the problems we mentioned before due to the dependence of
an adequate static threshold (0.9 by default). We investigated
the configuration (B = 0.3) and the high accuracy comes from
a very efficient selection of the most informative instances in
the early training (first 100 instances). However, this strategy
fails to generalise to other configurations as its outperformed
by other more flexible strategies.