We use local input representation. A feedforward neural net cannot learn the problem, due to the arbitrary time lags that may occur. Instead, we need something like a recurrent net (e.g. [18] [37] [17] [40] [15] [39] [38] [20] [7] ).
A conventional fully recurrent net with
5 input units, 5 hidden units, 1 output unit, 1 ``bias unit'',
and a learning rate of 0.5,
needed less than 10000 training sequences to
come up with correct classifications in
100% of all test sequences.
(class 1: output activation 
,
class 2: output activation 
at the end of a sequence).
This sets the stage for the more difficult task described in the next subsection. It involves highly redundant sequences and will be harder to learn.