I have 63 rows and 17 columns in the cocomo81 dataset (see the information here). The first 16 columns are the inputs to the network and the 17th column is the estimated outputs.
I took 2/3rd of the data for training and validation (i.e 42 rows), 1/3rd for testing (i.e from 43rd to 63rd rows). I sent 1st row for training with random weights and then 2nd row with updated weights (got when 1st row sent) and so on up to 41 rows (1-1000 iterations) and averaged the error got in each row and took as training Error-1. To get error i used (estimated output-actual output). After this I sent 42nd row for validation and got error in the same way as before and updated weight so got validation Error-1(no iterations).
Next I sent rows 43 to 63 for training with the updated weights got in validation and got the error in the same way as before (no iterations here). I used back propagation for updating weights and sigmoid activation function.
What I did is the right thing or not? My sir said that the validation errors and the testing errors should be less compared to training errors but I'm getting large values. So please tell me, whats wrong?