896  33152 365568 nic23a1.dat   In 29  Out  8
    896  33152 365568 nic23a3.dat   In 29  Out  8
    896  24192 267008 nic23b1.dat   In 19  Out  8
    896  24192 267008 nic23b3.dat   In 19  Out  8
    896  33152 365568 nic23c1.dat   In 29  Out  8
    896  33152 365568 nic23c3.dat   In 29  Out  8
    840  31080 342720 nic8a1.dat   In 29  Out  8
    840  31080 342720 nic8a3.dat    In 29  Out  8
    840  22680 250320 nic8b1.dat   In 19  Out  8
    840  22680 250320 nic8b3.dat   In 19  Out  8
    840  31080 342720 nic8c1.dat   In 29  Out  8
    840  31080 342720 nic8c3.dat   In 29  Out  8

nic23a1 and nic23a3 are the same feature, different target due to
different experts.

there are 3 data set labelled by a, b, c, respectively,
there are 2 experts giving the teaching output labelled by 1 and 3 resp.
data set a has 29 features dim but 3 or 4 of them are all 0's
data set b has 19 features and
data set c has 29 features. 

It seems normalize each feature individually should give better result.