Description: "To test whether our trained networks are following either the cumulative or recursive strategy, we
train diagnostic classifiers to predict the sequences of intermediate results of both these strategies, as
well as the variable mode used by the cumulative strategy to determine whether the next number
should be added or subtracted. As the diagnostic model should merely read out whether certain
information is present in the hidden representations rather than perform complex computations itself,
we use a simple linear model as diagnostic classifier."