mae is the mean of the absolute values of the errors.
If you use mae for learning, then tabulate mae.
mse is the mean of the squares of the errors.
If you use mse for learning, the tabulate one or more of
1. Root-mean-square-error rmse =sqrt(mse)
2. Normalized mean-square-error nmse = mse/mse00
3. Coefficient of determination R^2 = 1 - nmse
where mse00 is the mse of the naive constant model that
always outputs the mean of the target, regardless of the input.
When the NNet model is better than the naive constant model,
0 < 100*R^2 <= 1 is often viewed as the % of target variance
that is "explained" by the net and R is considered the nonlinear
correlation coefficient between input and output.
Note that mae and rmse have the dimensions of the target and
are dependent on how the output is scaled; whereas nmse
and R are normalized and scale independent.
I favor using R^2. Search wikipedia for "coefficient of determination".
Also search comp.ai.neural-nets and comp.soft-sys.matlabin in Google Groups for
greg heath MSE00 R2
Hope this helps.
Greg